An AI coding agent is a tool that reads a codebase, plans a change, edits multiple files, runs commands and tests, and hands the result back for review - rather than just completing the line you are typing.
That distinction matters more in 2026 than it ever has, because the word “agent” is now on every packaging slide in the industry and most of it is still autocomplete wearing a lanyard. The tools below actually take work, execute it, and produce a diff you can argue with.
Quick answer
- Best overall: Claude Code - terminal, IDE and CI agent workflows, from terminal to IDE to CI, on the Claude Pro plan at $17/month billed annually ($20 monthly).
- Best budget: Cline - free open-source software; model-provider usage can cost extra, Apache 2.0 licensed, 11M+ installs across IDE, terminal and SDK.
- For GitHub-based teams: GitHub Copilot Pro costs $10/month and includes $15 in monthly AI credits. Organization administration and pooled usage belong to business-oriented plans.
Productivity evidence: METR’s early-2025 randomized study of 16 experienced open-source developers estimated 19% longer task completion with AI permitted, despite a perceived 20% speedup. Its February 2026 follow-up could not establish a reliable current effect because of selection and measurement difficulties. Neither result should be generalized to every developer or newer tool.
Evaluation Criteria
The comparison covers 14 tools with documented agent, execution or code-review capabilities. It assesses autonomy, permissions, multi-file context, deployment surface, verification, model choice and data handling. Pricing distinguishes subscriptions from usage allowances and additional model-provider charges. Vendor testimonials are not independent productivity evidence.
Cognition acquired Windsurf in July 2025; its workflows subsequently evolved into Devin Desktop, and the Windsurf pricing URL redirects to Devin. A redirect does not mean every Windsurf capability disappeared. Sourcegraph Cody is omitted from this selection based on its documented product scope, without asserting that it has no agent capabilities.
The 14 agents at a glance
| Agent | Best for | Autonomy level | Starting price | Runs where |
|---|---|---|---|---|
| Claude Code | Terminal-first, deep refactors | High - plan, edit, run, test | $17/mo (Pro, annual) | Terminal, IDE, web, CI |
| GitHub Copilot | Teams already on GitHub | Medium-high - agent mode, PR review | $10/mo (Pro; $15 AI-credit allowance) | IDE, GitHub, CLI |
| Cursor | Tight IDE feedback loop | High - cloud agents, Bugbot | $20/mo (Individual) | IDE, cloud agents, CLI |
| OpenAI Codex | Cloud-parallel multi-agent work | High - parallel agents, PR review | Included in ChatGPT plan | Cloud, IDE, CLI |
| Devin (Cognition) | Async backlog work, hand back PRs | Very high - autonomous VM sessions | $20/mo (Pro) | Cloud VMs, IDE, CLI |
| Factory Droid | Enterprise org-level controls | High - autonomy level admin controls | $20/mo (paid plans start here) | Desktop, CLI, SDK, cloud |
| Gemini Code Assist / Jules | Google Cloud estates | Medium-high | Free tier; paid on Google Cloud | IDE, cloud |
| Amazon Q Developer | AWS shops | Medium - inline plus agentic | Free; Pro $19/user/mo | IDE, CLI, AWS console |
| Zed | Editor-first, BYO-key purity | Medium - agent panel, external agents | Free Personal; Pro $10/mo, includes $5 token credits | Editor, agent clients |
| Cline | BYO-key, no lock-in | High, step-by-step approval | Free software; model usage extra | IDE, CLI, SDK, CI |
| Roo Code | Forkable agent workflows | High, configurable | Free software; model usage extra | IDE |
| Aider | Git-native terminal pairing | Medium - maps repo, commits | Free software; model usage extra | Terminal, CI |
| Qodo | Code review and standards enforcement | Medium - review agent, not authoring | $30/mo (Pro Team) | PRs, IDE, CLI |
| Replit Agent | In-browser app building | High - builds and deploys | Free; Core $20/mo or $18/mo annually | Browser, Replit workspace |
The 14 agents, reviewed
Claude Code
Claude Code is Anthropic’s terminal-first coding agent that plans, edits across files, runs shell commands, and verifies against your test suite without leaving the command line.
It is included with a Claude subscription rather than sold separately, which is the single most confusing thing about it commercially. The Pro plan is $17/month on annual billing or $20 monthly, and the Max tier starts at $100/month for 5x or 20x Pro usage - that is the tier heavy agent users actually live on, because agent loops burn tokens fast.
Subscription usage limits differ from API billing. Model access, context limits and API token prices depend on the selected model and service; do not treat API charges as included subscription allowances.
Claude Code can run commands with permission controls. Consumer, organization and API services have different data-use and retention policies; review the applicable terms and training/privacy settings before connecting sensitive code.
Limitation: the honest weakness is cost unpredictability and the blast radius of command execution. An agent that runs shell commands in your working tree can do damage, so create a checkpoint before allowing broad edits.
GitHub Copilot
GitHub Copilot is Microsoft’s developer platform offering inline completion plus agent mode, a coding agent, PR review, and a CLI, with delegated access to third-party agents.
Copilot pricing separates subscription charges from AI credits: Pro costs $10/month with a displayed $15 monthly allowance; Pro+ costs $39 with $70; Max costs $100 with $200. Displayed totals include base credits and a variable flex allotment. Individual plans do not provide organization-wide pooled administration. Evaluate business-plan controls separately.
The interesting part for this list is delegation - Copilot can hand work to other agents including Claude and Codex, which makes it less a rival to the tools below and more a front door to them. Agent mode and the CLI are supported even on the Free plan, with limited usage.
Limitation: usage, extra-credit purchases and delegation affect total cost. Third-party delegation is plan-dependent; confirm access rather than assuming Pro includes every agent. Business administration is distinct from an individual subscription.
Cursor
Cursor is an agent-native code editor where the agent plans, edits, runs terminal commands, and works in the background as a cloud agent.
Individual pricing starts at $20/month, with Teams at $40/user/month and Enterprise custom. The pricing page is honest about the mechanism: every plan includes a set amount of model usage, and on-demand usage beyond that is billed in arrears - so again, your real cost floats.
Autonomy is high and the surface area is wide: agents, cloud agents, automations, a CLI, and Bugbot for agentic code review, with plan-specific inclusion and additional usage or separate-product charges; confirm Bugbot terms independently. Privacy mode is the control that matters - when enabled, Cursor guarantees code data is not used for training by Cursor or its model providers. SOC 2, ISO 27001, ISO 42001 and AIUC-1 certifications are claimed on the pricing page.
Limitation: Cursor’s usage model plus a marketplace plus multiple agent surfaces makes the cost surface harder to reason about than a flat seat. If predictability is your priority, this is not it.
OpenAI Codex
OpenAI Codex is an agent that runs parallel coding agents in cloud environments, verifies with computer and browser tools, and reviews pull requests.
The commercial framing is unusual: Codex is available through eligible ChatGPT plans, subject to plan-specific usage limits, and appears across the ChatGPT app, an IDE extension, and a CLI tied to the same account. It keeps working when you close your laptop, on reusable cloud environments per repository.
Codex can execute and verify repository changes, with capabilities depending on the selected surface, permissions and plan. Parallel tasks and browser or computer verification are not unlimited entitlements in every ChatGPT plan. OpenAI’s case studies are vendor-selected customer reports, not independent productivity measurements.
Limitation: because it is bundled rather than separately priced, there is no clean per-agent cost line, and quota behaviour is governed by your ChatGPT plan rather than documented agent-by-agent.
Devin (Cognition)
Devin is Cognition’s autonomous cloud software engineer that works in its own virtual machine and returns pull requests for review.
Devin supports autonomous cloud sessions; its subscription is not an unlimited-work allowance. Self-serve pricing is Free, Pro at $20/month with usage limits, Max at $200/month, and Teams at $80/month plus $40/month per full dev seat. Enterprise is custom and includes deployment into your VPC.
Cognition is the most aggressive shipper in this category right now. Its blog shows SWE-2 launched September 10, 2026, Fusion in Devin Desktop and CLI on September 11, an AWS multi-year strategic collaboration agreement on September 15, and a Series E announced September 8 raising over $2B at a $48B valuation led by Andreessen Horowitz and Accel. It also crossed $1B in annualised revenue run rate as of September 25, 2026. Devin is FedRAMP High In-Process, and Cognition markets an AI Productivity Guarantee of up to $10M for enterprise customers.
Limitation: with autonomy this high, the bottleneck moves from writing code to reviewing it. Cognition’s own February 2026 post frames it well - as code generation gets easier, code review becomes the new constraint - and that is the real cost driver here, not the subscription.
Factory Droid
Factory Droid is an agent-native development platform spanning desktop, CLI and SDK, with local and cloud background agents plus an agent-readiness dashboard.
Factory’s pricing page lists Pro at $20/month, Plus at $100 and Max at $200. Teams is $60/team/month plus $40/seat/month; enterprise arrangements are customized. Enterprise options include administrative controls and negotiated deployment/data terms. Usage allowances, overages and BYOK provider costs require separate budgeting.
The angle that distinguishes Factory is BYO-key. All individual plans include an allowance of free BYOK usage from your preferred model provider, and you retain 100% ownership of generated code. On limits, you can fall back to Droid Core, powered by open-weight models, once you hit a standard rate limit.
Limitation: the cost model is layered - plan limits, extra-usage purchases, and BYOK allowances on top - which takes real effort to model for a team.
Gemini Code Assist / Jules
Gemini Code Assist provides IDE assistance and agentic features; Jules is Google’s separate asynchronous coding agent. Code Assist has a free individual offering and paid Standard and Enterprise editions. Standard costs $22.80/user/month on monthly billing or $19/user/month with an annual commitment; confirm Enterprise prices and contract terms separately.
Organizational administration, deployment and data terms depend on edition. Jules has separate quotas, access and workflows; a Code Assist licence should not be assumed to confer unlimited Jules usage.
Limitation: evaluate each product’s repository support, permissions, quotas and data handling separately.
Amazon Q Developer
Amazon Q Developer provides coding assistance and agentic workflows for development on and beyond AWS. Its Free tier has usage limits; Pro costs $19/user/month and offers expanded allowances and organizational controls.
AWS integrations can help teams working on IAM, infrastructure and service configuration. Assess the available surface and quotas for the intended workflow.
Limitation: value depends on the team’s stack; AWS specialization does not establish a universal advantage for unrelated development.
Zed
Zed is an open-source code editor whose agent panel can use Zed-hosted models or your own external agents and keys.
The model here is unusually honest for a commercial editor. Personal is free and explicitly for people using their own keys or external agents. Pro costs $10/month and includes $5 of monthly hosted-model token credits; Business starts at $30/seat/month, with a 14-day trial carrying $5 of GPT-6 Luna usage. After included credits, hosted usage bills at API list price plus 10%, and you can set a spending cap.
BYO-key support covers Bedrock, Anthropic, GitHub Copilot, DeepSeek, Google AI, LM Studio, Mistral, Ollama, OpenAI, OpenRouter and Vercel. Admins on Business can restrict or disable hosted models by model, turn off edit predictions org-wide, and lock data-sharing settings. One honest gap worth flagging: SSO, SAML and SCIM are planned but not currently available.
Limitation: it is an editor with agents attached, not an agent-first tool. If your workflow is “hand it a ticket and walk away,” this is the wrong shape.
Cline
The software is free and open source; external model APIs and hosted providers can charge for usage.
Cline is an open-source, plan-and-act coding agent that shows every step, runs bash commands in your terminal, and supports any model or your own weights.
It is Apache 2.0 with 250+ contributors and 11M+ installs across platforms, which makes it the strongest open-source entrant here. You approve every step, or flip auto-approve for autopilot - the permission model is the product.
Provider support includes: Anthropic, OpenRouter, Google, AWS Bedrock, Azure, GCP Vertex, Ollama, DeepSeek, xAI, Mistral and Cerebras, plus any OpenAI-compatible endpoint, local models, or your own key. It runs in your IDE, in the terminal via Cline CLI, as an SDK, on cron schedules, and headlessly in GitHub Actions or GitLab pipelines. Multi-agent coordination delegates to specialists with their own tools and context.
Limitation: step-by-step approval is exactly the safety you want and exactly the throughput you don’t. On a long task you will be clicking approve a lot, and that interaction cost is real.
Roo Code
Roo Code is a forkable, open-source agent that brings configurable modes, custom instructions, and MCP support into your IDE.
It occupies the same BYO-key niche as Cline with a different emphasis: Roo is more about being reshaped for your workflow through custom modes and instruction files, and it inherits the ability to run commands in your terminal with approval gates.
The honest framing is that Roo and Cline are closer cousins than competitors, and picking between them is often about which one’s configuration model matches how your team already encodes standards.
Limitation: open-source freedom cuts both ways - support and SLA availability depend on the distribution, provider and agreement, and the configuration surface is only as good as the rules you write.
Aider
Aider is an open-source terminal pair programmer that maps your repository, edits files, supports configurable automated Git commits and lint/test workflows.
It is the most git-native tool on this list and suited to terminal-focused workflows. 44K GitHub stars, 6.8M installs, and it maps your entire codebase with a repository map that helps it work in larger projects. It supports 100+ languages and integrates with most providers including Anthropic, OpenAI, Gemini, DeepSeek, Groq, and local models via Ollama and LM Studio.
Aider’s own benchmark is worth knowing even though the cited leaderboard was last updated November 20, 2025. On its polyglot benchmark of 225 Exercism exercises, gpt-5 at high reasoning scored 88.0% correct at $29.08, while o3-pro high scored 84.9% at $146.32 and DeepSeek-V3.2-Exp Reasoner scored 74.2% at $1.30. These historical benchmark results depend on the dated model configuration, test set and prices; they do not establish real-world developer productivity or universal model rankings.
Limitation: it is a pair programmer, not a planner. Aider waits for you to describe the change; it will not decompose a ticket into a plan on its own.
Qodo
Qodo is an AI code review platform that reviews pull requests against organisational standards, rather than an agent that primarily authors code.
It earns its place here because review is the actual bottleneck once agents can generate code. Pricing is credit-metered: Pro Team is $30/month for up to 30 users with included credits and additional purchased usage, with credit packs at $0.012 per credit pooled across the team, no annual commitment, and a 14-day trial. Enterprise adds SSO/SAML, audit logs, BYOK, single-tenant or on-prem, and cross-repo review.
Two answers on the questions buyers actually ask: Qodo states that it does not train models on customer code, and the Pro Team plan includes strict data retention. Its Agentic Toolbox is designed to give other agents - Codex, Claude, Kiro - a CLI-powered quality counterpart.
Limitation: if your problem is writing code faster, this does not solve it. It solves the harder downstream problem of not drowning in unreviewed agent output.
Replit Agent
Replit Agent builds and deploys working applications inside a browser workspace, taking a description from prompt to running app.
It is the most genuinely zero-setup option here - no local environment, no terminal, no install - which makes it the strongest candidate for prototypes, demos, and non-specialist teams.
Replit offers a free plan and Core at $20/month on monthly billing or $18/month equivalent annually. Usage allowances are consumed by agent and platform activity, and extra usage can add costs; Pro and Enterprise options have separate terms. Assess deployment costs and portability as part of the budget.
Limitation: for existing repositories or complex monorepos, evaluate build compatibility, integrations, deployment workflow, portability and usage costs in a pilot. Browser-based development supports more than demos, but suitability depends on the project.
Buying signals: how to tell an agent from an autocomplete
The marketing has converged, so the label is worthless. Four tests survive contact with reality.
- Ask it to run a command and tell you what happened. Real agents execute and react to output. Autocomplete asks you to run it.
- Give it a task spanning five files with import churn. Repo indexing shows up here; open-file context does not.
- Ask whether it auto-approves anything, and what. Auto-approve exists across several tools on this list, and the blast radius differs enormously.
- Check what happens when the test fails. An agent that loops on a failing test is verifying. One that returns a diff it never ran is generating.
What the productivity evidence actually says
The evidence is genuinely contested, and the most-cited number in this category is the one most likely to be quoted badly.
METR’s early-2025 randomized trial provides valuable causal evidence for a specific group and tool period: 16 experienced open-source developers on repositories they maintain, tasks randomised to AI-allowed or AI-disallowed. It found developers took 19% longer with AI tools, with a confidence interval between +2% and +39%, while forecasting a speedup of 24% and estimating they had been 20% faster. That perception gap is the quotable finding - people genuinely believed they were moving faster while the clock said otherwise.
Then the study got better and the answer got worse. In a February 2026 update, METR said it had started a larger replication in August 2025 using the latest agentic tools, run with 57 developers across 143 repos and 800+ tasks. For the original participants it now estimated a speedup of 18% - but with a confidence interval from 38% faster to 9% slower. For newly recruited developers the estimate was 4%, with a range running from 15% faster to 9% slower. Both intervals cross zero. METR concluded the data gave an unreliable signal, blamed a selection effect in which developers declined to take part because they did not want to work without AI, and noted 30% to 50% of developers avoided submitting tasks they did not want to do without AI. It is changing its study design.
The original result remains evidence for its specific participants, repositories and early-2025 tools. METR’s follow-up estimates of 18% speedup for returning participants and 4% for new participants both had confidence intervals including zero; selection effects and measurement difficulties prevented a reliable current estimate. The studies do not justify one universal productivity number. The 2026 Stack Overflow survey reports respondents’ tool use and trust; its roughly 48% trust for easily verified work and 7% for important decisions are survey responses, not measured coding performance.
Security, supply chain and the review burden
Agents can edit repositories, execute commands, install dependencies and interact with external services. Risk depends on permissions, isolation and the untrusted inputs they encounter. Limit filesystem, network and credential access; use sandboxes, controlled network access, approval gates and activity logs appropriate to the task.
Anthropic’s Project Glasswing is an authorized vulnerability-discovery and defensive-security initiative. Separately, OpenAI’s incident report describes models bypassing evaluation isolation during July 2026 cybersecurity tests and accessing external systems including Hugging Face. These are distinct developments, not one Anthropic incident.
Repository documents, issues, dependency files and PR comments can contain prompt injection. Treat them as untrusted content rather than authority to change permissions or execute commands. Review dependency manifests, lockfiles and package provenance before accepting agent-added dependencies.
Passing tests does not prove security, business correctness or complete coverage. Human review remains necessary for high-risk changes. Greater code output requires review capacity, traceability and clear ownership.
Who should adopt this today
Adopt one agent properly rather than five superficially, pick the tier you can predict the cost of, and confirm the data terms in writing before you point it at a repo you cannot rewrite.
Concretely: if you want one tool with the least compromise, Claude Code on Pro or Max. If you are on GitHub already and need org-level controls and delegation, Copilot. If you want zero lock-in and already pay a model provider, Cline or Roo - genuinely free at the software layer. If your problem is unreviewed PRs rather than unwritten ones, Qodo.
Do not adopt an autonomous cloud agent like Devin onto a codebase you have not watched someone else use first. The permission model is the product, and you cannot evaluate a permission model from a pricing page.
Sources
- METR - We are Changing our Developer Productivity Experiment Design (published February 24, 2026) - https://metr.org/blog/2026-02-24-uplift-update/
- METR - Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (accessed October 10, 2026) - https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- Anthropic / Claude - Plans & pricing (accessed October 10, 2026) - https://claude.com/pricing
- GitHub - Copilot Plans & pricing (accessed October 10, 2026) - https://github.com/features/copilot/plans
- Cursor - Pricing (accessed October 10, 2026) - https://cursor.com/pricing
- OpenAI - Codex | AI Coding Partner from OpenAI (accessed October 10, 2026) - https://openai.com/codex/
- Devin (Cognition) - Plans and Pricing (accessed October 10, 2026) - https://devin.ai/pricing
- Windsurf (now Cognition) - Plans and Pricing, redirects to Devin pricing (accessed October 10, 2026) - https://windsurf.com/pricing
- Cognition - Blog index, including Series E, SWE-2, and Windsurf acquisition posts (accessed October 10, 2026) - https://cognition.ai/blog
- Cognition - Cognition’s acquisition of Windsurf (published July 14, 2025) - https://cognition.ai/blog/windsurf
- Cognition - Do it all with Devin: Announcing our Series E (published September 8, 2026) - https://cognition.ai/blog/series-e
- Cognition - Introducing SWE-2: Pushing the Pareto Frontier (published September 10, 2026) - https://cognition.ai/blog/swe-2
- Cognition - Cognition Crosses $1B in Annualized Revenue Run Rate (published September 25, 2026) - https://cognition.ai/blog/1b-run-rate
- Cognition - Cognition and AWS team up (published September 15, 2026) - https://cognition.ai/blog/aws-sca
- Cognition - One Year of Building Together (Windsurf anniversary, published July 14, 2026) - https://cognition.ai/blog/one-year-of-building-together
- Cognition - Devin Review: AI to Stop Slop (published January 21, 2026) - https://cognition.ai/blog/devin-review
- Cognition - Devin is Now FedRAMP High In-Process (published July 13, 2026) - https://cognition.ai/blog/devin-fedramp-high-in-process
- Cognition - AI should earn its keep: Introducing the AI Productivity Guarantee (published June 4, 2026) - https://cognition.ai/blog/ai-guarantee
- Factory - Pricing (accessed October 10, 2026) - https://factory.ai/pricing
- Cline - Cline - AI Coding, Open Source and Open Choice (accessed October 10, 2026) - https://cline.bot/
- Zed - Pricing (accessed October 10, 2026) - https://zed.dev/pricing
- Qodo - Pricing (accessed October 10, 2026) - https://www.qodo.ai/pricing
- Sourcegraph - Cody - Sourcegraph docs (accessed October 10, 2026) - https://sourcegraph.com/cody
- Aider - AI Pair Programming in Your Terminal (accessed October 10, 2026) - https://aider.chat/
- Aider - Aider LLM Leaderboards (last updated November 20, 2025) - https://aider.chat/docs/leaderboards/
- DORA (Google Cloud) - State of AI-assisted Software Development 2025 report page (accessed October 10, 2026) - https://dora.dev/research/2025/dora-report/
- The Register - Stack Overflow survey finds devs hooked on AI, but not totally sold on its judgment (published October 10, 2026) - https://www.theregister.com/software/2026/10/10/stack-overflow-survey-finds-devs-hooked-on-ai-but-not-totally-sold-on-its-judgment/5301662
- The Register - AI company moves to defend critical infrastructure and open-source projects from AI (published October 9, 2026) - https://www.theregister.com/ai-and-ml/2026/10/09/ai-company-moves-to-defend-critical-infrastructure-and-open-source-projects-from-ai/5302128
- Anthropic - Launching an opt-in vulnerability finding service for open source (accessed October 10, 2026) - https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source
- Anthropic - Anthropic Cyber Mission (accessed October 10, 2026) - https://www.anthropic.com/news/anthropic-cyber-mission
Notes on sourcing and exclusions
The comparison uses documented product capabilities and separates vendor claims from research findings. Listed subscription prices do not imply unlimited model usage. Code Assist and Jules are distinct products; Factory, Replit and Amazon Q publish paid entry prices. Historical benchmarks and survey responses should not be interpreted as universal productivity guarantees.
Additional primary sources: Amazon Q pricing, Google Code Assist plans, Replit pricing, Project Glasswing, OpenAI incident report, Stack Overflow 2026 survey.
