10 Grok 4.5 Deep Research Prompts for 2026

The short answer: the best Grok prompts for deep research in 2026 don’t ask for a long answer. They define the question, set a source policy, force competing explanations, and make uncertainty visible.

Grok is now a moving target. The current xAI API pages list Grok 4.3, Grok 4.20, Grok 4.20 Multi-Agent, and Grok 4.5. The exact model, tool access, and interface can differ between the Grok app and the API. That distinction matters if you’re using Grok for serious research rather than a quick answer.

I wrote these AI deep research prompts around the features xAI documents as of July 15, 2026: Web Search, X Search, code execution, file search, citations, structured outputs, reasoning effort, and the beta Multi-Agent model. I’ve also added a plain-English check for labels such as DeepSearch, Think, and Big Brain, because product names change faster than most prompt guides do.

TL;DR: the best Grok prompts for deep research

Use Grok 4.5 for a strong single-agent research workflow, and use Grok 4.20 Multi-Agent when you need several research perspectives working in parallel. Give either model a narrow question, a date range, allowed sources, a claim ledger, and an explicit rule for saying “not verified.”

Here’s the fast version:

• Use Grok 4.5 for source-heavy analysis, reasoning, code-assisted checks, and structured reports.
• Use Grok 4.20 when you want current agentic tool calling with a large context window and a fast research loop.
• Use Grok 4.20 Multi-Agent for complex research that benefits from parallel source gathering and a final synthesis pass. xAI labels this capability beta in its current documentation.
• Turn on Web Search for current public information. Without a search tool, Grok’s answer is limited by its model knowledge and prompt context.
• Use X Search for posts, threads, user profiles, and public reaction. Treat posts as leads until a primary source confirms the claim.
• Ask for citations beside each material claim, not just a source dump at the end.
• Ask Grok to separate fact, interpretation, estimate, and unresolved question.
• Use code execution for calculations, data cleaning, and reproducible comparisons. Do not ask the model to “do the math” when a small script can show it.
• Use files or Collections Search when your research depends on PDFs, filings, reports, or internal documents.
• I don’t treat “DeepSearch” or “Big Brain” as stable API instructions unless the current interface exposes those labels. The xAI API documentation I checked uses Web Search, X Search, reasoning effort, and Multi-Agent Research instead.

A good research prompt does not make Grok infallible. It makes wrong answers easier to spot.

What Grok 4.5 and Grok 4.20 DeepResearch actually are

Grok DeepResearch is best understood as a research workflow built from a model, live-search tools, reasoning, and a citation policy-not as a magic answer mode that removes the need for checking. In the current xAI API docs, the most concrete building blocks are Web Search, X Search, code execution, document search, structured outputs, and multi-agent orchestration.

Deep research means a multi-step process that gathers evidence, compares it, and produces a documented answer. It’s different from a normal chatbot reply because the system has to decide what to search, what to read, what to discard, and how to reconcile disagreement.

Grok 4.5 is the default choice for one careful researcher

Grok 4.5 is the current xAI model I’d pick for a single research thread that needs reasoning, tools, and a clean final report. xAI’s July 2026 model page describes it as a model for coding, agentic tasks, and knowledge work, with Web Search, X Search, code execution, function calling, structured outputs, and configurable reasoning effort.

The current model card lists a 500,000-token context window and an API price of $2 per million input tokens and $6 per million output tokens below the long-context threshold. The same pricing is shown on xAI’s pricing page, and TechCrunch reported the same $2/$6 figures when covering the July 8, 2026 release. I would still check the live xAI pricing page before budgeting a production workflow because prices and availability can change.

Grok 4.5 also has a knowledge cutoff of February 1, 2026, according to xAI’s model page. That does not make the model current by itself. For events after that cutoff, I use Web Search and a date-bounded prompt.

Grok 4.20 is built for fast tool use

Grok 4.20 is the faster agentic option in the current model family, while Grok 4.20 Multi-Agent adds parallel research roles. xAI’s model card lists Grok 4.20 reasoning and non-reasoning variants with a 1-million-token context window, tool calling, structured outputs, and reasoning support.

For a short market scan, a quick news check, or a first pass over a long source set, Grok 4.20 can be a practical choice. I would not pick a model only because a product page says “lowest hallucination rate” or “most truthful.” Those are provider claims, not independent evidence. Ask the model to show its sources, then audit the sources yourself.

Grok 4.20 Multi-Agent is the closest verified match for deep research

Grok 4.20 Multi-Agent is xAI’s documented beta model for multi-agent research. The current xAI guide says that multiple agents can search, analyze, cross-reference, and contribute to a leader agent that writes the final answer. It supports built-in tools such as Web Search, X Search, code execution, and Collections Search.

The documentation describes two agent-count settings: a smaller four-agent setup for focused research and a larger sixteen-agent setup for complex, multi-perspective work. xAI says the larger setup uses more tokens and can add latency. That tradeoff is real: a wider search is only useful if the final synthesis can distinguish independent evidence from repeated copies of the same claim.

I use Multi-Agent when the question has separable subproblems. For example:

• one research pass maps the market and competitors;
• one checks regulations and primary documents;
• one looks for counterexamples and failed implementations;
• one audits the source quality and dates;
• the leader agent merges the findings into a claim-by-claim brief.

I don’t use it for a one-sentence fact. More agents can mean more search noise, more token use, and more opportunities for the same weak source to appear several times.

What “DeepSearch” means in practice

The current xAI API docs I fetched do not present DeepSearch as a stable API tool name. They document web_search, x_search, code_execution, document search, Collections Search, reasoning effort, and grok-4.20-multi-agent research. The old frontmatter for this article uses “DeepSearch” because that label is familiar to Grok users, but I would treat it as a user-interface label or informal shorthand unless your current Grok account shows it.

That distinction changes how you write prompts. Instead of writing, “Use DeepSearch,” write:

• “Use live Web Search.”
• “Search only sources published from January 1 through July 15, 2026.”
• “Search these domains first, then expand only if needed.”
• “Use X Search only for discovery and public reaction.”
• “Return the source URL, publication date, and the exact claim supported by each source.”

Those instructions survive a UI rename.

How a research run works

A reliable Grok research run follows a loop: plan, search, inspect, compare, calculate, write, and audit. The model may repeat parts of the loop, but your prompt should make the checkpoints explicit.

1. Frame the question. Define the object, place, period, audience, and decision.
2. Set the evidence rule. Name preferred source types and minimum source quality.
3. Search broadly. Find the main entities, terms, reports, and competing explanations.
4. Search laterally. Open the source’s own claims, funding, methodology, and corrections.
5. Bind claims to sources. Ask which exact source supports each sentence.
6. Check dates. Separate current facts from historical context and stale pages.
7. Run calculations. Use code execution for totals, ranges, rates, and sensitivity checks.
8. Write with uncertainty. Use labels such as confirmed, supported, disputed, estimated, and unknown.
9. Run an adversarial pass. Ask what would disprove the answer and which premise may be wrong.
10. Deliver a decision-ready report. Put the short answer first, then methods, evidence, caveats, and next actions.

OpenAI’s current Deep Research help page describes a similar control pattern: the user defines the outcome, chooses sources, reviews a plan, watches progress, can interrupt the run, and receives a report with citations or source links. Google’s current Gemini help page also describes a research plan, source selection, and an editable plan before the report starts. The common lesson is simple: review the plan before you trust the result.

Grok models and deep research competitors compared

Grok’s strongest research advantage is the combination of live Web Search, X Search, code execution, and a documented Multi-Agent option. Competitors often offer a more packaged research report experience or deeper connections to their own productivity ecosystems. The right choice depends on whether you need broad web research, social-source analysis, private work data, or a polished report.

Product or model	Current 2026 research path	Source and tool control	Reasoning or orchestration	Best fit	Main caution
Grok 4.5	xAI API or Grok surfaces with tools enabled	Web Search, X Search, code execution, files, Collections Search, citations	Single-agent reasoning with low, medium, or high effort; high is the documented default	Research briefs, data checks, current events, technical analysis	Search access and UI controls can differ by plan and surface; verify the live tool state
Grok 4.20	xAI API	Web Search, X Search, code execution, structured outputs, document tools	Reasoning and agentic tool calling	Fast evidence gathering and long-context work	Provider performance claims still need independent testing
Grok 4.20 Multi-Agent	xAI API beta	Built-in search and document tools	Multiple agents gather and debate; a leader synthesizes	Multi-part research, consensus mapping, competing explanations	More agents can raise cost, latency, duplication, and synthesis risk
ChatGPT Deep Research	ChatGPT Deep Research tool	Public web, selected sites, uploaded files, connected read-only apps	Plan review, live progress, interruptible research, documented report	Work-product reports and authenticated business sources	The result still needs claim-level fact checking; access and limits vary by plan and country
Gemini Deep Research	Gemini Apps Deep Research	Google Search by default, selected Google services, files, NotebookLM notebooks	Research plan, editable before starting, report export and optional visuals	Research tied to Google Search and Google Workspace material	Search defaults can shape the source pool; specify sources and language
Microsoft 365 Copilot Researcher	Microsoft 365 Copilot for enterprise	Web and Microsoft Graph data, plus connectors and work files	Researcher and Analyst agents inside a governed work ecosystem	Internal company research and Microsoft 365 workflows	Requires the right Microsoft 365 subscription and permissions
Claude	Claude and Anthropic research tools	Capabilities depend on the current product and workspace	Anthropic publishes current research and agent work, but I did not find a stable public “Deep Research” product page in the sources checked	Writing, analysis, coding, and careful synthesis	Don’t assume a named research mode or tool set without checking the current account

The table is a product snapshot, not a leaderboard. In a February 2026 evaluation of commercial chatbots as news intermediaries, the researchers found that retrieval quality, regional coverage, and false-premise resistance affected results. That is exactly why I don’t treat “has citations” as the same thing as “is well sourced.”

A second 2026 arXiv study on AI forecasting found that model diversity mattered: combining models with complementary errors could be more useful than repeatedly sampling a single model. I apply that idea in prompt design. If the decision matters, ask Grok for a primary-source pass, a skeptical pass, and a counterevidence pass. Don’t just ask for three copies of the same answer.

The prompt anatomy I use for serious research

A strong Grok research prompt has six parts: question, scope, sources, method, output, and stop rule. If one is missing, the model has to guess, and its guesses become hidden assumptions.

1. State the question as a decision

“Research electric vehicles” is a topic. “Which battery chemistry should a fleet operator in India consider for city delivery vans purchased in 2026?” is a decision.

Add the audience and the time window. A research answer for a journalist is different from one for a procurement manager, a student, or a regulator.

2. Separate discovery sources from proof sources

Use Web Search or X Search to find leads. Use company filings, government pages, standards bodies, peer-reviewed papers, and named research reports to prove important claims.

Definition: lateral reading means leaving a source to check who produced it, what other credible sources say, and whether the original evidence supports the claim.

3. Make the model show a claim ledger

A claim ledger turns a smooth paragraph into auditable rows:

Claim	Evidence	Source date	Source type	Confidence	What could change it
The claim itself	Quote, table, result, or calculation	Publication or update date	Primary, secondary, academic, social	High, medium, low	Missing data or new evidence

4. Ask for a stop rule

Research can keep expanding forever. Add a stop rule such as: “Stop after you have three independent primary or high-quality secondary sources for each major claim, or clearly mark the claim unresolved.”

5. Ask for uncertainty in plain language

Don’t request a fake confidence percentage. Ask the model to explain whether the uncertainty comes from stale data, conflicting sources, missing methodology, small samples, regional gaps, or ambiguous definitions.

6. Ask for a final audit

The final audit should check:

• unsupported numbers;
• citations that don’t support the sentence beside them;
• sources outside the date window;
• copied reporting presented as independent confirmation;
• false premises in the original question;
• conclusions stronger than the evidence;
• recommendations that depend on unverified assumptions.

1. Consensus map prompt

Use this prompt when you need to know where credible sources agree, where they disagree, and why the disagreement exists. It’s especially useful for policy, product categories, public health claims, climate questions, and emerging technology debates.

Full prompt

Research the question below as a consensus map, not as a single definitive answer.

Question: [insert the precise question]
Audience: [who will use this]
Geography: [country, region, or global]
Time window: [start date] through July 15, 2026

Use live Web Search. Search primary sources first, then high-quality secondary reporting and recent academic work. Search at least three distinct source families, such as government or regulator pages, original research or filings, and independent reporting. Do not count syndicated copies of the same report as independent confirmation.

Build a claim ledger with these columns:
1. Claim
2. Sources that support it
3. Sources that challenge it
4. Exact disagreement
5. Source date
6. Method or evidence type
7. Confidence: high, medium, or low
8. What evidence would change the conclusion

Separate:
- broad consensus;
- qualified consensus;
- active dispute;
- unresolved because evidence is missing.

For every statistic, provide the definition, denominator, date, geography, and at least two independent sources. If two sources repeat the same underlying dataset, mark them as one evidence family. Do not average incompatible figures.

Start with a five-sentence answer. Then show the map, the disagreement drivers, the strongest evidence, the weak evidence, and a short conclusion for my audience. End with five questions a human researcher should verify before publishing.

When to use it

Use the consensus map for questions where “the answer” depends on definitions. For example, “Do AI agents improve productivity?” could mean time saved, output volume, revenue, quality, or worker experience. A good map makes those definitions visible.

I also use it when a search result page is full of confident claims that all point back to one press release. The prompt forces Grok to trace the evidence family rather than count repetition as consensus.

Expected output structure

Ask for:

1. Short answer.
2. Scope and definitions.
3. Broad consensus.
4. Qualified consensus.
5. Active disputes.
6. Evidence table.
7. Source-quality audit.
8. What would change the conclusion.
9. Editorial or business implication.

Example walkthrough

Suppose the question is: “Are AI web agents ready for online shopping in 2026?” Grok should first define “ready.” Does it mean it can complete a purchase, compare products, avoid unsafe actions, explain its choice, or all five?

The 2026 arXiv study on agent-ready websites gives a useful structure for this kind of question. It separates agent interpretability, agent executability, and agent decision reliability. Its controlled experiment compared a baseline site with an agent-ready version and reported 134 passes out of 150 for the agent-ready condition versus 74 out of 150 for the baseline. The abstract and full HTML paper both report those figures, while the paper also labels the evidence as a controlled proof of concept. A careful report should not turn that into “all websites should expect an 89.3% success rate.”

The consensus map should therefore say: there is promising evidence that clearer machine-readable structure helps; there is not proof that every commercial website will see the same effect.

Variations

• Policy version: Require laws, regulator guidance, court decisions, and implementation dates.
• Scientific version: Require study design, sample size, preregistration, effect size, and replication status.
• Market version: Require company filings, customer evidence, independent analyst work, and a distinction between forecast and reported result.
• News version: Require an original event source, a local source, and a correction check.

Pro tip

Ask Grok to draw the map around claims, not brands. “OpenAI says X, Google says Y, and xAI says Z” is a company comparison. “The evidence for lower latency is strong, while the evidence for better factuality is mixed” is a research comparison.

2. Assumption testing prompt

Use this prompt before you accept the premise of a research question. It is the best first step when the question contains a loaded verb such as “caused,” “proved,” “best,” “safe,” “leading,” or “guaranteed.”

Full prompt

Act as an assumption auditor before answering.

Research question: [insert question]
Decision I’m trying to make: [insert decision]
Audience: [insert audience]
Date boundary: Use only information available through July 15, 2026.

First, rewrite the question in neutral language. Then list every factual, causal, definitional, geographic, temporal, and value assumption hidden inside it.

For each assumption, return:
- assumption;
- why the question depends on it;
- evidence that supports it;
- evidence that weakens it;
- whether it is verified, plausible, disputed, or unsupported;
- the smallest change to the question that would make it answerable.

Use Web Search for current sources. Use X Search only to identify claims or reactions that need primary-source verification. Do not treat a post, press release, vendor page, or repeated article as proof of causation.

If the question contains a false premise, say so clearly before continuing. Do not silently repair the premise. Provide two versions of the answer:
A. answer under the original premise, if possible;
B. answer under the corrected premise.

Return a compact assumption table, then a researched answer, then a list of unresolved assumptions. Put citations beside the claims they support. For every number, include its date, unit, denominator, and source family.

When to use it

Use it for competitor research, investigations, health questions, political claims, and any request that starts with “Why did…” or “How much better is…”. Those forms often assume a fact that has not been established.

The February 2026 commercial chatbot evaluation is a strong reason to use this prompt. The researchers tested false-premise questions and found that models could perform well on clean questions yet struggle when a subtle invented detail was embedded in the question. The lesson isn’t that one model is always bad. The lesson is that a polished prompt can steer a capable model toward a false starting point.

Expected output structure

• Neutral restatement.
• Assumption inventory.
• Evidence for and against each assumption.
• Corrected research question.
• Answer under the original question.
• Answer under the corrected question.
• Remaining uncertainty.

Example walkthrough

Question: “Why did Grok 4.5 become the fastest research model in July 2026?”

The hidden assumptions include: that it is the fastest, that “research model” has a defined benchmark, that July 2026 data is available, and that the reason is known. The official xAI page describes Grok 4.5 as fast and efficient, and TechCrunch reported the company’s claims about token efficiency. That is not the same as an independent proof that it is the fastest across all research tasks.

A corrected version could ask: “What did xAI and independent reporting claim about Grok 4.5’s speed and token efficiency in July 2026, and what independent evidence is available?” That question is narrower and safer.

Variations

• Causal audit: Require a timeline, possible confounders, and evidence that the cause preceded the outcome.
• Investment audit: Separate reported results, management guidance, analyst estimates, and your own scenario assumptions.
• Health audit: Add “do not diagnose; identify what requires a qualified professional.”
• Legal audit: Ask for jurisdiction, effective date, primary law, and a clear note that the answer is not legal advice.

Pro tip

Put “Do not silently repair my premise” near the top. If it appears at the bottom, the model may already have built the answer around the assumption.

3. Source verification and lateral reading prompt

Use this prompt when you have a claim or URL and need to know whether the source really supports it. It turns Grok from a summarizer into a source auditor.

Full prompt

Audit the claim and source below using lateral reading.

Claim: [paste the exact claim]
Source URL: [paste URL]
Why it matters: [decision or publication context]
Date boundary: Use current information through July 15, 2026.

Do not summarize the source first. Verify it.

1. Identify the publisher, author, publication date, update date, funding or commercial interest, and stated methodology.
2. Find the original source behind the claim, such as a filing, dataset, paper, government release, transcript, or direct announcement.
3. Search for at least two independent sources that discuss the same claim.
4. Check whether the headline, social post, or summary exaggerates the underlying evidence.
5. Check definitions, denominators, sample sizes, time periods, geography, and omitted qualifiers.
6. Check corrections, retractions, archived versions, and material updates.
7. Classify the claim as supported, partly supported, misleading by omission, disputed, or unsupported.

Return:
- verdict in two sentences;
- claim-to-evidence table;
- source lineage;
- independent corroboration;
- contradictions or missing context;
- a corrected version of the claim;
- confidence and what remains unverified.

Cite every material finding. Do not count sources that copy one press release as independent. Do not infer that a source is reliable merely because it ranks highly in search.

When to use it

Use it for vendor claims, benchmark screenshots, “study shows” headlines, viral posts, statistics in press releases, and pages with no visible methodology.

I use this before quoting a model vendor’s benchmark. The source may be real and still be incomplete. The benchmark could use an internal test set, a particular prompt scaffold, a subset of tasks, or a scoring rule that makes the number look better than it is.

Expected output structure

1. Verdict.
2. Publisher and incentive check.
3. Original evidence.
4. Two independent checks.
5. Claim lineage.
6. Missing qualifiers.
7. Corrected wording.
8. Confidence and publication recommendation.

Example walkthrough

Claim: “Grok 4.5 is an Opus-class model that is twice as token efficient.”

Grok should separate the claim into pieces. “Opus-class” is a comparison label from Elon Musk and the TechCrunch report. “Twice greater token efficiency” is an xAI claim reported by TechCrunch. A source audit should not convert those phrases into an independent benchmark result. It should ask whether token efficiency means tokens per task, output length, cost per completed task, or something else.

For the price claim, the xAI model detail page, xAI pricing page, and TechCrunch article agree on the $2 input and $6 output prices for the listed Grok 4.5 API path. That is a cross-checked price point. It still does not prove total research cost, because Web Search invocations, reasoning tokens, and the number of tool calls can change the bill.

Variations

• Academic paper audit: Add peer-review status, dataset release, code availability, preregistration, and whether the result is a preprint.
• Company benchmark audit: Add test-set ownership, prompt disclosure, model version, temperature, tool access, and independent replication.
• Social post audit: Add account identity, original media, timestamp, location, and whether the post is firsthand.

Pro tip

Paste the exact sentence, not only the URL. That keeps the audit tied to the claim you intend to publish.

4. Competitive intelligence prompt

Use this prompt when you need a decision-ready comparison instead of a feature list. It works for AI tools, software, vendors, products, and market categories.

Full prompt

Create a competitive intelligence brief for:

Category: [category]
My product or organization: [description]
Decision: [what I need to choose or change]
Target users: [audience]
Markets: [geography]
Evaluation period: [start date] through July 15, 2026

Research competitors using current Web Search. Use official product pages for stated capabilities and independent reporting, user evidence, filings, analyst work, or technical documentation for validation. Use X Search only for discovery and public reaction. Mark vendor claims as vendor claims.

Compare no more than [number] competitors on:
- target user;
- primary job to be done;
- verified features;
- integrations and data access;
- research or search workflow;
- citations and export options;
- privacy or governance information;
- pricing only when published by a current primary source;
- known limitations;
- evidence of adoption or customer use;
- switching cost;
- strongest differentiator;
- strongest weakness.

Build a table with source links and dates for every row. Separate verified facts, reported claims, analyst interpretation, and my own inference. Do not rank competitors using an invented score. Instead, give three recommendations for different priorities: lowest risk, fastest workflow, and deepest research.

End with:
1. what I can decide now;
2. what needs a live trial;
3. a seven-day test plan;
4. the assumptions that could make this comparison wrong.

When to use it

Use it before buying a research tool, choosing a model, writing a “best tools” article, or planning a migration. A feature matrix alone hides the thing buyers care about: which workflow works for their data, time, team, and risk level.

Expected output structure

• Executive decision.
• Market definition.
• Competitor table.
• Evidence notes.
• Three priority-based recommendations.
• Trial plan.
• Risks and unknowns.

Example walkthrough

For a comparison of Grok, ChatGPT Deep Research, Gemini Deep Research, and Microsoft 365 Copilot Researcher, the prompt should distinguish public web research from private work-data research.

The current OpenAI help page says Deep Research can use the public web, selected sites, uploaded files, and connected read-only apps, with a proposed plan and citations or source links. Google’s help page says Gemini Deep Research uses Google Search by default and can also use Gmail, Drive, files, and NotebookLM notebooks. Microsoft’s current enterprise page positions Researcher and Analyst inside a Microsoft 365 environment connected to work data and connectors.

That produces a useful conclusion: Grok is attractive when X Search and flexible tool use matter; ChatGPT is attractive when a packaged report and connected sources matter; Gemini is attractive for Google-centered research; Copilot is attractive when the evidence is already inside Microsoft 365. None of those statements means one tool wins every task.

Variations

• SEO tool comparison: Add crawling limits, keyword database region, API access, and export format.
• Enterprise procurement: Add SSO, audit logs, data retention, admin controls, and contract terms.
• Content production: Add citation usability, fact-check turnaround, brand controls, and editorial review time.

Pro tip

Ask for a trial plan. A comparison that ends with “it depends” is unfinished; a trial plan turns the uncertainty into something you can test.

5. Research synthesis from PDFs and private files prompt

Use this prompt when the evidence lives in reports, filings, PDFs, spreadsheets, or internal documents rather than on the open web. It reduces the risk that Grok fills gaps from general web knowledge.

Full prompt

Analyze the attached files as a closed evidence set first.

Research goal: [question]
Decision or output: [what I need]
Files attached: [list files]
Date boundary for outside sources: through July 15, 2026

Phase 1: Search only the attached files. Identify the relevant passages, tables, figures, definitions, caveats, and document dates. Return a file-grounded claim ledger with file name, page or section, exact supporting passage, and confidence.

Phase 2: List gaps, contradictions, ambiguous terms, and calculations that require checking. Do not use outside knowledge to silently fill gaps.

Phase 3: If outside Web Search is needed, search only for the named gap. Prefer primary or official sources. Clearly label every outside source and explain why it was needed.

Phase 4: Use code execution for arithmetic, totals, rates, trend lines, or comparisons. Show formulas and inputs. Do not invent missing values.

Return:
- answer in five sentences;
- source map by file;
- claim ledger;
- contradictions;
- calculations with assumptions;
- outside-source additions;
- unresolved gaps;
- a review checklist for a human owner.

If a document is unreadable, incomplete, outdated, or internally inconsistent, stop and say exactly what failed.

When to use it

Use it for annual reports, regulatory filings, research literature, internal surveys, policy documents, due diligence, and technical manuals. xAI’s current Files documentation says attached documents can trigger an agentic attachment-search workflow, and Collections Search can query persistent document collections.

Expected output structure

1. File-only answer.
2. Source map.
3. Claim ledger.
4. Contradiction list.
5. Code-assisted calculations.
6. Outside evidence.
7. Unknowns.
8. Review checklist.

Example walkthrough

Imagine you attach a 2026 annual report, two quarterly filings, and an analyst spreadsheet. Grok should first answer from those files. If one file says “active users” and another says “monthly active users,” it should not merge them without checking the definitions.

If you ask for a growth rate, the prompt should make Grok show the numerator, denominator, date range, and whether the comparison uses reported or rounded figures. A report can look precise while mixing fiscal years, calendar years, and trailing twelve-month figures.

If you attach a spreadsheet, ask for code execution and a reproducible calculation. A line such as growth = (new - old) / old is easier to audit than a confident sentence with no working.

Variations

• Literature review: Add study design, cohort, intervention, outcome, limitations, and DOI.
• Financial filing: Add reported versus non-GAAP numbers, segment definitions, and restatement checks.
• Internal policy review: Add policy owner, effective date, superseded version, and control mapping.

Pro tip

Tell Grok when not to search the web. A closed evidence set is useful when you want to know what your own documents actually say, not what the model can find elsewhere.

6. Quantitative analysis and uncertainty prompt

Use this prompt when the research question includes numbers, forecasts, rates, rankings, or a comparison that could be distorted by rounding. It turns prose into an auditable calculation.

Full prompt

Perform a quantitative research check on the question below.

Question: [insert question]
Data or sources: [attach data or provide URLs]
Time window: [dates]
Geography and unit: [define]
Decision threshold: [if any]

Use Web Search to verify the source definitions and current values. Use code execution for all calculations. Do not calculate with numbers that have different definitions, dates, currencies, units, or denominators without flagging the mismatch.

Return:
1. data dictionary;
2. source and date for every input;
3. cleaned data table;
4. formula for each metric;
5. calculation output;
6. sensitivity analysis using reasonable alternative assumptions;
7. uncertainty range or scenario band;
8. plain-language interpretation;
9. what the data cannot establish.

For every statistic, show the denominator and the population it describes. If a source gives a percentage without a sample size, method, or base population, mark it incomplete. If an estimate is not independently verified, label it an estimate. Do not manufacture a confidence interval from insufficient information.

Finish with a one-paragraph answer for a nontechnical reader and a reproducibility checklist.

When to use it

Use it for market sizing, adoption data, benchmark comparisons, forecast models, survey results, and “what is the percentage increase?” questions.

Expected output structure

• Data dictionary.
• Input table.
• Source and date audit.
• Formula block.
• Results.
• Sensitivity scenarios.
• Plain-English interpretation.
• Limits.

Example walkthrough

Suppose you want to compare the cost of Grok 4.5 research with a competitor. The xAI model page lists token prices, while xAI’s pricing page also lists server-side tool charges. TechCrunch reported the model price on July 8, 2026. A careful analysis should not multiply only the input and output token price. It should also ask how many reasoning tokens, Web Search calls, X Search calls, and follow-up turns the task used.

The correct output may be a scenario table:

Scenario	Model tokens	Tool calls	Follow-ups	What it means
Short lookup	Low	Low	None	Fast and cheap, shallow evidence
Source audit	Medium	Medium	One	Better traceability, higher usage
Multi-agent brief	Higher	Higher	Several	Broadest review, highest cost and latency

The point is not to pretend that a published per-token rate equals a final project cost.

Variations

• Forecasting: Add base rate, forecast horizon, update schedule, and calibration check.
• Survey analysis: Add sample design, weighting, margin of error, and nonresponse risk.
• A/B comparison: Add unit of analysis, randomization, exposure window, and stopping rule.

Pro tip

Ask for “what this data cannot establish.” That sentence often stops a descriptive statistic from becoming a causal claim.

7. Timeline and event reconstruction prompt

Use this prompt when the answer depends on what happened first, when a decision changed, or which source appeared at each stage. It is useful for product launches, regulatory actions, lawsuits, safety incidents, and breaking news.

Full prompt

Reconstruct a sourced timeline for:

Event or topic: [topic]
Geography: [location]
Start date: [date]
End date: July 15, 2026
Audience: [audience]

Use current Web Search and prioritize primary records: official announcements, filings, court documents, regulator notices, transcripts, data releases, and dated company pages. Use reputable reporting to fill context, not to replace primary records. Use X Search only when a post is itself part of the event or points to a first-hand record.

For every timeline entry, return:
- date and time zone if available;
- event;
- entity responsible;
- exact source;
- source publication or update date;
- whether the item is confirmed, reported, alleged, or inferred;
- what changed because of the event.

Then identify:
1. first public signal;
2. first primary confirmation;
3. major corrections or reversals;
4. disputed dates;
5. gaps in the record;
6. causal claims that the timeline alone cannot prove.

Do not backdate a claim. Do not treat a later summary as evidence that the earlier event was known at the time. End with a concise narrative and a fact-check list.

When to use it

Use it when a normal summary would flatten a complicated sequence. A timeline shows whether a company announced a feature, corrected it, restricted it, or changed the terms later.

Expected output structure

• Five-sentence overview.
• Chronological table.
• Confirmed versus alleged events.
• Corrections and reversals.
• Disputed dates.
• Causal limits.
• Fact-check list.

Example walkthrough

For current Grok research, a timeline can separate the July 2026 Grok 4.5 release from March 2026 Grok 4.20 and Multi-Agent release notes, and from earlier product names that may still appear in search results. xAI’s release notes list the month-by-month additions, while the current model pages show the models available in July.

That distinction helps prevent a common mistake: describing a March capability as if it launched with Grok 4.5 in July, or treating an old UI label as a current API feature.

Variations

• Legal timeline: Add docket number, jurisdiction, filing date, hearing date, and ruling status.
• Safety timeline: Add incident, mitigation, verification, and independent assessment.
• Product timeline: Add availability surface, model slug, plan restriction, and deprecation risk.

Pro tip

Ask for time zones. “July 8” can mean different release days for different users, and the difference matters when you’re tracing what was known when.

8. Counterevidence and red-team prompt

Use this prompt after you have a draft conclusion. It should try to break the conclusion without inventing objections.

Full prompt

Red-team the conclusion below using evidence, not contrarian performance.

Conclusion to test: [paste conclusion]
Research question: [question]
Evidence already used: [paste links or claim ledger]
Decision at stake: [decision]
Date boundary: through July 15, 2026

First, restate the conclusion and list the minimum evidence it requires. Then search for:
- credible counterexamples;
- boundary conditions;
- failed replications or negative results;
- contradictory primary sources;
- changes in definitions or measurement;
- survivorship bias;
- selection bias;
- regional or language gaps;
- incentives that could distort reporting;
- cases where the conclusion is true but not useful for the decision.

For each challenge, return:
- challenge;
- source;
- exact evidence;
- whether it weakens, narrows, or does not affect the conclusion;
- what additional test would resolve it.

Do not create a false balance. Weight evidence by method, source quality, recency, and independence. If the conclusion survives, rewrite it in a narrower, evidence-matched form. If it fails, say what should replace it.

End with a stoplight verdict: green, amber, or red, plus the reason in three sentences.

When to use it

Use it before publishing a strong headline, choosing a vendor, relying on a safety claim, or turning a benchmark result into a recommendation.

Expected output structure

1. Conclusion under test.
2. Evidence required.
3. Counterevidence table.
4. Weight of each challenge.
5. Revised conclusion.
6. Stoplight verdict.
7. Next test.

Example walkthrough

Conclusion: “More reasoning always improves research quality.”

The current xAI docs show that reasoning effort controls the amount of reasoning for Grok 4.5, while Multi-Agent effort controls the number of collaborating agents. That verifies the control exists. It does not prove that high effort always improves factual reliability.

A 2026 arXiv paper on agent meltdowns found unsafe behavior in agents responding to simulated environmental errors and reported that increasing thinking effort did not reduce meltdown rates in its experiments. That paper studied agent safety under errors, not ordinary research quality, so it is counterevidence to an absolute “always,” not proof that reasoning is bad.

The narrower conclusion is stronger: more reasoning can help with complex analysis, but it can also increase cost, latency, and the number of actions an agent takes. Use it when the task needs it, and set a stop condition.

Variations

• Safety review: Add misuse cases, red-team test, human escalation, and rollback.
• Editorial review: Add headline accuracy, quote context, missing caveats, and reader harm.
• Business review: Add downside scenario, switching cost, and reversible pilot.

Pro tip

Ask the model to distinguish “weakens the conclusion” from “changes the decision.” A small factual caveat may not change the practical choice, while a single safety issue might.

9. Consensus forecasting and scenario prompt

Use this prompt when you need to reason about what may happen next without pretending that a forecast is a fact. It is designed for planning, not fortune-telling.

Full prompt

Build a scenario forecast for:

Question: [binary or clearly defined outcome]
Resolution date: [date]
Geography: [location]
Decision user: [who will act on it]
Information cutoff: July 15, 2026

Use Web Search for current evidence. Use X Search only for public signals and treat it as noisy evidence. Identify the base rate, current evidence, key drivers, opposing indicators, and unknowns.

Return three scenarios:
1. base case;
2. upside or higher-probability case;
3. downside or lower-probability case.

For each scenario, provide:
- conditions required;
- evidence for;
- evidence against;
- leading indicators;
- what would falsify it;
- likely timing;
- confidence expressed in words, not a fake precise percentage.

If you give numerical probabilities, explain the reference class, assumptions, and sensitivity. Do not use a model-generated probability as if it were an empirical frequency.

Then create a weekly update template that records what changed, which evidence was new, and whether the forecast should move. Separate facts, forecasts, and recommendations. End with a decision rule for acting under uncertainty.

When to use it

Use it for launch planning, market changes, policy implementation, technology adoption, and research questions with a known future resolution.

Expected output structure

• Question and resolution rule.
• Base rate.
• Evidence ledger.
• Three scenarios.
• Leading indicators.
• Falsifiers.
• Update template.
• Decision rule.

Example walkthrough

The 2026 arXiv forecasting study “Diversity is the Strength of the AI Crowd” is useful here. It studied whether different model forecasts add value when their errors are not perfectly correlated. The practical prompt lesson is to ask for diverse perspectives, not just repeated samples from one model.

For a question such as “Will a new AI research feature become broadly available by October 1, 2026?” one pass should inspect official release notes, one should inspect public product pages and regional availability, and one should look for restrictions, beta language, or unresolved dependencies. The answer should not say “yes” because a launch post sounds confident. It should list what has to happen next.

Variations

• Product forecast: Add user adoption, distribution, pricing, and support dependencies.
• Policy forecast: Add legislative, regulatory, implementation, and litigation paths.
• Science forecast: Add replication, peer review, grant, and commercialization gates.

Pro tip

Set a resolution date and a falsifier. A forecast without a date or a condition that could prove it wrong is just an opinion with formatting.

10. Publish-ready research report prompt

Use this prompt when you need a final article, memo, or briefing that readers can verify. It combines the other nine methods into one editorial workflow.

Full prompt

Write a publish-ready research report on:

Topic: [topic]
Primary question: [question]
Audience: [audience]
Length: [range]
Tone: direct, human, conversational, and precise
Date boundary: use only 2026 information through July 15, 2026, unless historical context is needed to explain a current 2026 event

Research protocol:
1. Start with a short answer in two sentences.
2. State the scope, definitions, and assumptions.
3. Use primary sources first, then independent reporting, academic research, and analyst or industry reports.
4. Use at least two independent sources for every material statistic. If that is not possible, omit the number or label it unverified.
5. Do not count syndicated copies of the same report as independent sources.
6. Use X Search only for discovery, public reaction, or a first-hand post that can be authenticated.
7. Use code execution for calculations and show the inputs and formula.
8. Build a claim ledger before drafting.
9. Run a false-premise audit and a counterevidence pass.
10. Mark each key claim as confirmed, supported, disputed, estimated, or unresolved.

Required output:
- headline;
- short answer callout;
- definition of key terms;
- comparison table where relevant;
- method and source policy;
- findings in descending order of importance;
- counterevidence and limits;
- practical implications;
- FAQ with four to six direct questions and answers;
- dated linked source list;
- final fact-check checklist.

Writing rules:
- Keep paragraphs to three sentences or fewer.
- Put citations beside claims, not only at the end.
- Never fabricate a source, statistic, model name, benchmark, price, or feature.
- If a product label is not present in the current primary documentation, say that it is unverified rather than guessing.
- Do not hide uncertainty in a smooth sentence.

When to use it

Use it for a blog article, analyst memo, client briefing, or internal research note. It is also useful as a second pass after you have run a separate research session and want Grok to turn the evidence into a clean deliverable.

Expected output structure

• Short answer.
• Definitions.
• Comparison table.
• Findings.
• Evidence and limits.
• FAQ.
• Sources.
• Fact-check checklist.

Example walkthrough

For this article, the prompt would require Grok to distinguish the old “Grok 3 prompts” framing from the current 2026 model family. The xAI release notes say Grok 4 was released in July 2025, Grok 4.20 and Multi-Agent went live in March 2026, and Grok 4.5 appears in the July 2026 section. The current model pages then show the models and tools available on the API.

The report should say that Grok 4.5 and Grok 4.20 are current documented API models, while “DeepSearch,” “Think,” and “Big Brain” should be verified against the user’s current product surface before being treated as stable names. That is less flashy than guessing. It is also more useful six months later.

Variations

• SEO article: Add search intent, related questions, one comparison table, and a clear definition in the first 100 words.
• Executive memo: Add recommendation, risk, owner, deadline, and decision gate.
• Academic brief: Add method, inclusion criteria, limitations, and full citations.
• Customer-facing guide: Remove unnecessary model jargon and preserve only the checks a reader can perform.

Pro tip

Make the model draft the claim ledger first and the prose second. A beautiful article is easy to write after the evidence is organized. It’s hard to audit after the prose has smoothed away the uncertainty.

Best practices for Grok deep research prompts

The most reliable Grok research workflows are narrow, source-aware, and explicit about what counts as evidence. Prompt length alone is not the goal. A short prompt with a clear evidence policy beats a long prompt full of adjectives.

Define the research object

A research prompt should name exactly what is being compared, measured, or explained. Include the model slug, product surface, geography, date, and unit when any of those can change the answer.

For Grok, say “Grok 4.5 on the xAI API with Web Search enabled” if that is what you mean. Don’t say “Grok” when the answer could differ between the app, X, Grok.com, and the API.

Use source tiers

Source tiers tell Grok how to use different kinds of evidence. My default order is:

• primary documents and official documentation;
• peer-reviewed papers, preprints, and original datasets;
• regulator, court, standards, and government records;
• independent reporting from named outlets;
• analyst or industry reports with disclosed method;
• company statements and vendor pages, labeled as claims;
• expert commentary;
• social posts, forums, and snippets as leads.

A vendor page is still useful. It is simply not independent proof of the vendor’s own claim.

Ask for source diversity, not source volume

Ten sources can still be one source if they all repeat the same announcement. Ask Grok to group sources by evidence family and identify the original dataset, report, filing, or statement behind the copies.

This is where the 2026 forecasting research is useful: complementary errors can matter more than more samples from the same model. A diverse search process can matter more than a longer list of similar URLs.

Make date boundaries explicit

Use exact dates whenever current information matters. “Latest” is not a date. Use from_date and to_date in API search parameters where available, and write the same boundary in the prompt.

For this article, the boundary is July 15, 2026. That prevents a 2025 launch page from being mistaken for a current 2026 capability page and makes it easier to explain what was available on the update date.

Ask for citations at the claim level

Citations help only when they support the sentence they sit beside. Ask for the exact source, date, claim supported, and any limitation. xAI’s citation documentation distinguishes all citations from inline citations, which is a useful mental model even when you’re using the Grok app.

Ask for a human checkpoint

A human should review the plan, the source set, the key numbers, and the final conclusion before publication or a consequential decision. Current OpenAI and Google deep-research help pages both describe plan review and source control. Grok prompts should adopt the same discipline even when the interface doesn’t force it.

Common failure modes

Grok fails most dangerously when the prompt rewards a smooth conclusion more than a traceable process. These are the problems I watch for first.

It answers from stale model knowledge

A model knowledge cutoff is not a live news feed. xAI’s current page lists February 1, 2026 as the Grok 4.5 cutoff. For a July 2026 product feature, price, regulation, or event, turn on live search and set the date window.

It mistakes a press release for independent evidence

If every article says “the company claims,” the claim is still a company claim. Ask Grok to find the underlying benchmark, method, filing, or external test.

It cites a related page that does not support the sentence

A citation about a product launch may not support a claim about accuracy, safety, market share, or performance. Ask for a claim-to-source mapping and read the linked passage.

It accepts a false premise

The 2026 news-intermediary evaluation found that subtle false premises could sharply reduce model performance. Use the assumption-testing prompt before asking “why” or “how much better.”

It counts repeated sources as consensus

Syndication, copied press releases, and search summaries can create the appearance of agreement. Ask for evidence families and original-source lineage.

It overuses X Search

X Search is valuable for public reaction, first-hand posts, and fast signals. It is also noisy. A post can be deleted, misdated, clipped, or based on a rumor. Use it to find leads, then confirm the claim elsewhere.

It keeps searching after the answer is stable

More browsing can add little value after the major claims have independent support. Add a stop rule: “Stop when each material claim has two independent sources, the main counterargument has been tested, and remaining gaps are named.”

It hides uncertainty behind a precise number

A percentage with no denominator is not a useful statistic. A forecast with two decimal places is not automatically more rigorous. Ask for the data definition, range, assumptions, and what would change the result.

It treats citations as proof of safety

Citations say where information came from. They do not guarantee that an answer is safe, lawful, or appropriate. The Common Sense Media Youth AI Safety Institute labeled Grok an unacceptable risk for teen users in its January 22, 2026 assessment, and TechCrunch reported the assessment on January 27. That is a reminder to treat safety as a separate review question.

Advanced techniques: DeepSearch, Think, Big Brain, and current xAI controls

The current verified xAI controls are Web Search, X Search, code execution, file and Collections Search, citations, structured outputs, reasoning effort, and Multi-Agent Research. “DeepSearch,” “Think,” and “Big Brain” should be treated as interface labels unless the current product documentation confirms them.

DeepSearch: use the underlying controls

If your Grok interface shows a DeepSearch button, use it as a shortcut, then still specify the evidence policy in your prompt. The official API docs do not require a tool called DeepSearch. They describe live Web Search and X Search, plus tool orchestration through the Responses API.

A robust prompt for a DeepSearch-like workflow looks like this:

Use live Web Search and, only when useful, X Search. Search from January 1 through July 15, 2026. Start with [approved domains]. For each material claim, return the source URL, source date, evidence type, exact supporting passage or table, and confidence. Search for one credible counterexample before writing the conclusion. Stop when the claim ledger is complete or state which claim remains unresolved.

That prompt still works if the button is renamed.

Think mode: reasoning effort, not a guarantee

The current xAI API documents reasoning effort for Grok 4.5 as low, medium, or high, with high as the default and reasoning unable to be disabled on that model. xAI’s reasoning guide says low is suited to latency-sensitive or simpler tasks, medium to complex analysis, and high to very challenging multi-step work.

Use low for:

• a short source extraction;
• a small number of citations;
• a simple comparison;
• a follow-up that only clarifies the previous report.

Use medium for:

• conflicting sources;
• data interpretation;
• a multi-constraint recommendation;
• an uncertainty map.

Use high for:

• multi-part research;
• technical proof or derivation;
• a red-team pass;
• complex calculations;
• a source audit with several competing explanations.

Reasoning costs tokens and can increase latency. It also does not replace a source policy. A model can think longer about the wrong source.

Big Brain mode: not verified in the current docs I checked

I could not verify “Big Brain” as a stable xAI API model or documented research control in the current xAI model pages and July 2026 release notes. I would not tell readers to search for a specific Big Brain toggle, model slug, or promised behavior unless their current Grok surface exposes it and the linked help text explains it.

If the UI does show Big Brain, treat it as an experiment:

1. record the exact label and model shown;
2. save the prompt, date, source list, and output;
3. run the same prompt without the label;
4. compare source quality, completeness, errors, latency, and cost;
5. don’t publish a model-spec claim until an official page confirms it.

This is not pedantry. Model names and aliases change. xAI’s current docs explicitly recommend aliases for migration and dated model names for reproducibility. Record the exact model ID when a result matters.

Use structured outputs for a claim ledger

Structured Outputs can force the API response into a defined JSON schema, which is useful for research records. xAI says supported schema features are guaranteed to match the schema, while best-effort keywords still need validation.

A simple claim schema might include:

{
  "claim": "string",
  "status": "confirmed | supported | disputed | estimated | unresolved",
  "sources": [
    {
      "url": "string",
      "date": "string",
      "evidence_type": "primary | academic | reporting | social",
      "supports": "string"
    }
  ],
  "counterevidence": ["string"],
  "confidence_reason": "string",
  "next_check": "string"
}

Schema validation does not validate truth. It only keeps the output organized. You still need to open the URLs and verify the claim.

Combine files, web search, and code execution carefully

A strong research workflow can combine private documents, live web sources, and calculations, but the prompt should keep those evidence streams separate. Ask Grok to label a claim as file-grounded, web-grounded, calculated, or inferred.

For example:

• a filing supplies reported revenue;
• Web Search supplies a current competitor price;
• code execution calculates the margin difference;
• the final recommendation is an inference that depends on demand assumptions.

If those four layers are written as one sentence, readers cannot tell what is known and what is modeled.

Use multi-agent roles that disagree productively

Multi-Agent Research works best when each role has a different job and a final leader must reconcile the work. Don’t ask every agent to “research the topic.” Assign roles:

• primary-source finder;
• independent reporting checker;
• counterevidence finder;
• data and calculation checker;
• domain-method reviewer;
• final editor.

Tell the leader to penalize duplicated sources, expose disagreements, and preserve unresolved claims. A multi-agent system that produces one smooth answer without showing disagreement has not done enough synthesis.

Protect long research sessions

Long research conversations need context and cost controls. xAI documents stateful Responses API interactions, prompt caching, and Context Compaction. The current Responses API stores responses for 30 days by default, and xAI documents store: false for workflows that should not keep previous requests on the server.

Use a stable prompt prefix and a prompt cache key when you are iterating on the same research. Compact a long conversation before it becomes unwieldy. Keep private information out of prompts unless your data policy permits it.

FAQ

What is the best Grok model for deep research in 2026?

Grok 4.5 is the best documented single-agent choice for deep research, while Grok 4.20 Multi-Agent is the best documented option for parallel research. xAI’s current pages list Web Search, X Search, code execution, structured outputs, and reasoning for Grok 4.5. The Multi-Agent guide adds several collaborating agents and a leader-agent synthesis step. Choose the smallest model and tool set that can answer the question reliably.

Is Grok DeepSearch the same as Grok 4.20 Multi-Agent?

Not necessarily. The current xAI API documentation I checked does not use DeepSearch as a stable API tool name. It documents Web Search, X Search, document search, reasoning, and Grok 4.20 Multi-Agent Research. If your app shows DeepSearch, check the current in-product description and record the underlying model and tools.

Does Grok 4.5 have live web access by default?

Live information requires search tools or a search-enabled product surface. xAI’s model documentation says Grok has no access to real-time events without search tools enabled. In an API request, add the documented Web Search or X Search tool and ask for citations. In a consumer interface, check whether the search or browsing control is active.

How many sources should a Grok research answer use?

Use enough sources to cover each material claim, not a fixed number for the whole answer. My default is two independent sources for a statistic or high-stakes claim, with at least one primary source when available. For a broad report, ask Grok to group sources into evidence families so ten copied articles don’t count as ten confirmations.

Can Grok citations be trusted?

Citations are useful leads and audit points, not automatic proof. xAI documents all citations and inline citations, but a citation can still be incomplete, tangential, stale, or mismatched to the sentence. Open the source, find the supporting passage, check its date, and ask whether it supports the exact claim.

What should I do if Grok gives conflicting sources?

Do not force a single answer. Ask Grok to define the disagreement, compare dates and methods, identify whether the sources use different populations or units, and state which evidence is stronger. If the disagreement cannot be resolved, publish the disagreement and explain what would settle it.

Sources

• xAI: Models and current API pricing - current page accessed July 15, 2026.
• xAI: Grok 4.5 model card - current page accessed July 15, 2026.
• xAI: Grok 4.5 overview - current page accessed July 15, 2026.
• xAI: Web Search tool - current page accessed July 15, 2026.
• xAI: X Search tool - current page accessed July 15, 2026.
• xAI: Multi-Agent Research - beta documentation accessed July 15, 2026.
• xAI: Reasoning and reasoning effort - current page accessed July 15, 2026.
• xAI: Citations - current page accessed July 15, 2026.
• xAI: Structured Outputs - current page accessed July 15, 2026.
• xAI: Files and document search - current page accessed July 15, 2026.
• xAI: Chat with Files - current page accessed July 15, 2026.
• xAI: Release Notes - July 2026 release notes accessed July 15, 2026.
• xAI: Pricing - current page accessed July 15, 2026.
• OpenAI: Deep research help center - updated July 15, 2026.
• OpenAI: Introducing deep research - February 10, 2026 update listed on the page.
• Google Gemini Apps Help: Use Deep Research - current 2026 help page accessed July 15, 2026.
• Microsoft: Microsoft 365 Copilot for enterprise - current 2026 product page accessed July 15, 2026.
• Anthropic: Research - current research page accessed July 15, 2026.
• Suzgun et al.: Evaluating Commercial AI Chatbots as News Intermediaries - submitted May 21, 2026.
• Aitchison et al.: Diversity is the Strength of the AI Crowd - submitted June 29, 2026.
• Elnaffar and Rashidi: Designing Agent-Ready Websites for AI Web Agents - submitted July 13, 2026.
• Jha et al.: Agent Meltdowns - submitted May 18, 2026.
• Tan and Thien: Structured Chain-of-Thought Prompting for Political Evasion Detection - submitted June 14, 2026.
• Waibl et al.: BELLS-O - submitted June 12, 2026.
• Common Sense Media Youth AI Safety Institute: Grok and @grok on X - updated January 22, 2026.
• TechCrunch: SpaceXAI releases Grok 4.5 - July 8, 2026.
• TechCrunch: Report on Grok child-safety failures - January 27, 2026.
• WIRED: SpaceX listed Grok’s Spicy mode as a risk - May 20, 2026.

---