Photo by Lukas on Pexels Source

Updated October 23, 2026: rebuilt the comparison with plans and prices checked on each vendor’s own page, SpaceX’s takeover of Cursor and its new Grok bundle, Windsurf’s rename to Devin Desktop, Gemini CLI’s replacement by Antigravity CLI, and October 2026 agent benchmark scores.

Short answer for October 2026: Claude Code is the strongest coding agent you can buy, GitHub Copilot is the easiest choice for teams that already live on GitHub, and Devin is the one to pick if you want to hand over whole tickets and only review the pull requests. Cursor is still an excellent agent-first IDE, but since SpaceX bought it, its plans only include Cursor’s own Grok and Composer models; Claude and GPT now cost extra at API list price. The sleeper is OpenAI Codex: it sits 7.6 points behind Claude Code on the most useful independent agent benchmark at about 13% of the API cost per task.

All prices below were checked on the vendors’ own pricing pages on October 23, 2026. If you read the March version of this article, skip to what changed: most of these products look different now.

The Verdict: Which AI Coding Assistant to Pick

  • Solo developer who wants the most capable agent: Claude Code on a Max plan ($100 or $200 a month). It leads the harness-level benchmark below, and it now runs in your terminal, VS Code, JetBrains, a desktop app, the browser and Slack. It is also the most expensive per task.
  • Solo developer or small team watching costs: Codex through ChatGPT Plus ($20). GPT-6.1 Sol in Codex scored 54.1 against Claude Code’s 61.7, but cost $1.50 per task instead of $11.42.
  • Team on GitHub, or an enterprise that wants one approved vendor: Copilot Business ($19 a seat) or Enterprise ($39). Claude, GPT, Gemini and Grok models sit behind one bill, and the cloud agent works straight from issues and pull requests.
  • Team with a backlog of well-specified tickets: Devin. Its cloud agent takes a ticket and returns a PR, and its new Fusion harness matched Claude Code’s top score at 30% lower cost per task.
  • Developer who wants to stay in an editor and is happy with Grok: Cursor Premium ($30). Treat Grok 4.7 and Composer 2.5 as your daily models and budget extra for Claude or GPT.
  • Google shop: Antigravity, included in Google AI Pro ($19.99).
  • AWS shop that wants specs before code: Kiro ($20 to $200).

Plenty of teams pay for two: one interactive tool for the work they steer, one delegation tool for the tickets they don’t want to touch.

Comparison Table: Plans and Prices (Checked October 23, 2026)

ToolFree tierIndividual plansTeam / businessHow usage is metered
Claude CodeNo (Free plan excludes it)Pro $20 ($17/mo annual), Max $100 (5x) or $200 (20x)Team Standard $25/seat ($20 annual), Premium $125 ($100 annual); Enterprise $20/seat plus usage at API ratesFive-hour session limits plus weekly caps, shared with the Claude apps
OpenAI CodexYes (GPT-6 Luna in the desktop app)Go $8, Plus $20, Pro $100 / $200 / $500Business $25/user ($20 annual, 2+ users); Enterprise customPlan allowance shared with ChatGPT Work; buy credits to extend
GitHub CopilotYes (2,000 completions, limited chat and agent)Pro $10, Pro+ $39, Max $100Business $19/seat, Enterprise $39/seatGitHub AI Credits at $0.01 each, billed at model API rates; completions unlimited
CursorYes (limited)Plus $8 (Grok only), Premium $30, Super $100, Ultra $200Teams from $30/seat; Enterprise customWeekly pool for Cursor’s own models; Claude, GPT and Gemini at API list price
Devin (incl. Devin Desktop, formerly Windsurf)Yes (light quota, unlimited Tab)Pro $20, Max $200Teams $80/month plus $40 per dev seat; Enterprise customDaily and weekly quotas; extra usage at API pricing
Google Antigravity (plus Jules)Yes (basic weekly limits)Google AI Pro $19.99, Ultra $99.99 (5x) or $199.99 (20x)Gemini Enterprise Standard/Plus from $30/seat, or pay-as-you-goRate limits plus a flexible AI credit pool
Kiro (AWS)Yes (50 credits)Pro $20, Pro+ $40, Pro Max $100, Power $200Same per-user tiers for teams of 2 to 500; Enterprise custom1,000 to 10,000 credits a month; add-ons at $0.04 per credit

Sources: Claude, Codex, Copilot, Cursor, Devin, Antigravity and Google AI plans, Kiro. US prices, before tax.

ToolOwnerWhere you use itModels included
Claude CodeAnthropicTerminal, VS Code, JetBrains, desktop app, web, SlackOpus 5.5, Sonnet 5.5, Haiku 5.5; Fable on Max
CodexOpenAIDesktop app, CLI, IDE extension, web, iOSGPT-6.1 Sol, GPT-6 Astra, GPT-6 Luna
GitHub CopilotGitHub (Microsoft)VS Code, Visual Studio, JetBrains, Copilot CLI, Copilot app, GitHub.comClaude, GPT, Gemini, Grok, Kimi and Microsoft models
CursorSpaceX (Anysphere)Cursor desktop app, CLI, web, iOS, SlackGrok 4.7, Grok 4.6, Grok 4.5, Composer 2.5
DevinCognitionDevin Desktop, Devin Cloud, CLI, SlackSWE-2 plus OpenAI, Anthropic, Google and SpaceXAI models
AntigravityGoogleAntigravity app, IDE extensions, CLI; Jules for async GitHub workGemini 4 Argon on Pro and Ultra
KiroAWSKiro IDE, CLI, webAuto router, Claude Opus 5 and Sonnet 5, GPT-5.6, open-weight models

What Changed Since March: Dead, Merged and Renamed

Windsurf is gone as a brand. On June 2, 2026, Cognition relaunched Windsurf as Devin Desktop, an IDE built around an agent command center. Its Cascade agent was replaced by a Rust rewrite called Devin Local, with legacy Cascade available through July 1. windsurf.com/pricing now redirects to devin.ai/pricing.

Gemini CLI was retired for consumers. Google announced Antigravity CLI on May 19 and, on June 18, 2026, stopped serving Gemini CLI and the Gemini Code Assist IDE extensions for Google AI Pro, Ultra and free users. Organizations on Gemini Code Assist Standard or Enterprise licenses keep access, and Gemini CLI still works with paid API keys.

Cursor belongs to SpaceX. The roughly $60 billion deal was announced in June, and Cursor confirmed it closed on August 14. In mid-October the old $20 Pro plan disappeared from the pricing page, replaced by a bundle sold through X that includes Grok and X Premium perks. Subscribers on legacy Pro, Pro+, Ultra and Teams plans keep them as they are.

Copilot’s premium requests are gone. Since June 1, 2026, every Copilot plan runs on GitHub AI Credits, drawn down by input, output and cached tokens at each model’s API rate. Base plan prices did not change.

Devin’s ACU pricing is off the self-serve page. The $20-plus-ACUs Core plan and the $500 Teams plan from the March version are no longer listed. Current plans use quotas that refresh daily and weekly, with extra usage at API pricing.

Claude Code is no longer terminal-only. The March article called that its biggest weakness. It now ships native VS Code and JetBrains extensions, a desktop app and a web version.

Benchmarks: The Harness Matters as Much as the Model

The March version led with SWE-bench Verified numbers, mostly vendor-reported and run on different scaffolds. The better comparison now is the Artificial Analysis Coding Agent Index, which scores the agent harness and model together. Version 2.0 averages three benchmarks (FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA) over three attempts per task, and reports the pay-per-token API cost of each run. Scores as displayed on October 23, 2026:

Agent + modelIndex scoreAvg. API cost per task
Claude Code + Opus 5.5 (max)61.7$11.42
Devin Fusion CLI + Opus 5.5 with GPT-6 Luna61.5$8.00
Claude Code + Sonnet 5.5 (max)59.3$13.30
Muse Code + Muse Wonder Full (max)58.9not listed
Codex + GPT-6 Astra (max)54.7$7.04
Codex + GPT-6.1 Sol (max)54.1$1.50
Antigravity CLI + Gemini 4 Argon (high)53.2$5.86
Opencode + GLM-5.348.9$5.23
Grok Build + Grok 4.7 (xhigh)46.6$9.48

Three things stand out. Claude Code with Opus 5.5 leads, but it is also near the top on cost. Devin’s Fusion harness runs the same Opus 5.5 as a lead model with the cheaper GPT-6 Luna doing execution, and lands within 0.2 points for 30% less, which is the clearest proof yet that scaffolding is worth real money. And Codex with GPT-6.1 Sol is the value outlier of the whole table.

Cursor and Copilot are not in the index. Both are multi-model front ends, so their quality depends mostly on which model you select. The cost column is API-equivalent spend, not what you pay on a subscription.

Claude Code: Strongest Agent, Highest Bill

Claude Code started as a terminal tool and is now available on every surface Anthropic ships: the CLI, VS Code and JetBrains extensions, a desktop app with visual diffs and side-by-side sessions, the web, and Slack. The terminal is still where it is strongest.

Models. Claude Opus 5.5 launched September 22, 2026 at $4/$20 per million input/output tokens, which Anthropic says makes it 40% cheaper to run than Opus 5 on typical workloads. Anthropic reports 54.6% on FrontierCode at default effort, ahead of GPT-6 Astra’s 53.3%. Sonnet 5.5 and Haiku 5.5 followed, and Fable 5.1, Anthropic’s priciest model, is available on Max plans. Context windows go up to 1M tokens.

Agent features. A lead agent can spawn multiple Claude Code agents that work on different parts of a task, assign subtasks and merge the results. Background agents let you run several full sessions and watch them from one screen, and cloud scheduled tasks (formerly routines) keep running when your laptop is off. Claude Code now runs in auto mode by default on Pro, Max and Team plans, sessions can run on your own infrastructure (public beta), and Code Review is generally available for Team and Enterprise.

Pricing. Claude Code is not in the Free plan. Pro is $20 a month ($17 billed annually). Max is $100 for 5x Pro usage or $200 for 20x, and includes $100 or $200 a month in API credits. Team Standard seats cost $25 ($20 annually) and Premium seats with 5x the usage cost $125 ($100 annually). Enterprise is $20 a seat plus usage at API rates.

The trade-off. You get Anthropic models only, and limits reset on five-hour windows with weekly caps on top. Heavy agent days hit them. If you want the highest ceiling and can pay for it, nothing else is clearly better.

Related: Software Factories: When AI Agents Build Software Without Human Review

Cursor: Still a Great Agent IDE, Now With a Grok-First Price List

Ownership. Cursor’s maker Anysphere is now part of SpaceX. Cursor says the deal gives it access to SpaceX’s GPU fleet to train its own models, and Grok 4.6 shipped from Cursor and SpaceXAI together. Grok 4.7 followed, priced from $2/$6 per million tokens.

Product. Cursor 3 added an agent-first workspace built from scratch, with the classic VS Code-fork editor one click away. Local and cloud agents sit in one sidebar, whether you started them from desktop, mobile, web, Slack, GitHub or Linear, and you can move a session from cloud to local and back. The October changelog added cloud agent canvases, an Environments API, environments with up to 300 repositories, remote control of local agents from the iOS app, and an app marketplace for Origin, Cursor’s own code hosting. Bugbot still handles PR review, and teams can turn on Cursor Router, which Cursor says cut costs 30% to 50% for early enterprise users.

Pricing. Free, Plus ($8), Premium ($30), Super ($100) and Ultra ($200). Plus only covers Grok models and excludes cloud agents. Premium and up add cloud agents, Bugbot and third-party models. The included weekly pool covers Grok 4.7, 4.6, 4.5 and Composer 2.5 only. Claude, GPT and Gemini are billed at API list price from prepaid credit, and Teams pay an extra Cursor Token Rate of $0.25 per million tokens on them. Team seats match the individual tiers at $30, $100 or $200.

The trade-off. If your Cursor habit was Claude Sonnet all day for $20, that deal is gone for new subscribers. Existing legacy plans keep their monthly pools, so don’t migrate unless the Grok pool fits how you work.

When Cursor Beats Claude Code

Cursor wins when you want to stay in an editor: Tab completions, a proper diff view with full LSP navigation, cheap included usage on its own models, and per-task model choice. Claude Code wins when the job is “build this feature end to end” and you want frontier Anthropic models included in a flat subscription.

GitHub Copilot: The Default for GitHub Teams

Copilot’s advantage is still ecosystem integration, but it is no longer the weak model option. Its model list now includes Claude Opus 5.5, Sonnet 5.5 and Fable 5.1, GPT-6.1 Sol and GPT-6 Astra, Gemini 4 Argon, Grok 4.7, Kimi K3 and Microsoft’s MAI-Code-1.1-Flash.

Agents. Agent mode runs in the IDE and in Copilot CLI. The GitHub Copilot app, available on all plans, is a desktop app for running parallel agent sessions, each on its own worktree and branch, in Interactive, Plan or Autopilot mode. On GitHub.com, the coding agent (now called the cloud agent in GitHub’s docs) picks up issues, edits code, runs tests and linters in an ephemeral cloud environment, and can run as a scheduled or event-triggered automation. Custom agents, agent skills, MCP servers and hooks shape what it does. You can also hand tasks to Anthropic’s Claude agent or OpenAI Codex from the same interface (public preview, Pro+ and Max on individual plans).

Pricing. Free, Pro ($10), Pro+ ($39) and the new Max ($100). Business is $19 a seat and Enterprise $39. Each plan includes AI Credits worth $0.01 each: 1,500 a month on Pro, 7,000 on Pro+ and 20,000 on Max (base plus a flex allotment GitHub says may change), and 1,900 or 3,900 per Business or Enterprise user, pooled across the org. Code completions and next edit suggestions don’t consume credits.

The trade-off. Credits are billed at model API rates, so Pro’s 1,500 credits are $15 of frontier-model usage. Long agent sessions on Opus or Astra burn through that fast. For a team on GitHub, the governance (org-wide pooling, admin spend limits, model policies) is the real product.

Related: VS Code 1.109: How the IDE Became a Multi-Agent Development Platform

Devin: Delegation, Now With an IDE Attached

Devin by Cognition is still the most autonomous option: describe a task in Slack, the web app or a ticket, and Devin Cloud plans it, writes the code, tests it and opens a PR. What changed is everything around it.

One product family. Devin Desktop (the former Windsurf) makes an agent command center the default view: a Kanban board of every local and cloud agent. It supports the Agent Client Protocol, so Codex, Claude Agent and OpenCode run inside it too. Devin CLI and Devin Review round out the lineup.

Models and harness. SWE-2, Cognition’s own model, launched September 10, post-trained from Moonshot’s Kimi K3. Cognition reports 50.0% on FrontierCode 1.1 Main, within a point of Fable 5.1 at 64% lower cost, and it’s free to use on Pro. Fusion pairs a frontier “lead” model for planning and review with a cheaper “sidekick” for execution. That is the harness that tied Claude Code in the benchmark table above. Cognition says it crossed $1 billion in annualized revenue in September.

Pricing. Free, Pro ($20) and Max ($200). Teams costs $80 a month plus $40 per full developer seat, for up to 200 users. Enterprise adds VPC deployment and SSO. Quotas refresh daily and weekly, and extra usage is billed at API pricing.

The trade-off. The old ACU guessing game is gone, but autonomy still pays off only on tickets with clear acceptance criteria. For architectural judgment calls, you want a human in the loop, and an interactive tool does that better.

OpenAI Codex: The Value Pick

Codex is included in every ChatGPT plan. Free and Go ($8) get GPT-6 Luna in the desktop app. Plus ($20) adds Codex on the web, in the CLI, the IDE extension and iOS, plus automatic code review and Slack, with GPT-6.1 Sol and GPT-6 Luna. Pro comes at $100, $200 or $500, currently has no five-hour limit, and the $500 tier gets Ultrafast mode. Business is $25 per user monthly or $20 annually.

GPT-6.1 Sol arrived on September 29, and GPT-5.5 was retired from Codex on October 14. OpenAI estimates a Plus user gets 15 to 160 local messages per five hours with GPT-6.1 Sol, 5 to 45 with GPT-6 Astra, and 350 to 3,000 with Luna.

The trade-off. Codex’s ceiling sits below Claude Code’s on the index, and Astra limits on Plus are tight. But a 54.1 score at $1.50 per task makes Codex the cheapest way to run a frontier-class agent all day, especially if you already pay for ChatGPT.

Related: GPT-5.3-Codex vs. Claude Opus 4.6: The Coding Agent Wars

Google Antigravity: What Replaced Gemini CLI

Google put its coding bets into one product. The Antigravity 2.0 desktop app and Antigravity CLI, written in Go, share one agent harness. The CLI keeps Gemini CLI’s agent skills, hooks, subagents and extensions (now plugins) and adds background multi-agent runs.

Pricing starts with a free individual tier with basic weekly limits. More capacity and Gemini 4 Argon come with Google AI Pro at $19.99 or Ultra at $99.99 (5x) or $199.99 (20x). Businesses buy through Google Cloud, with Gemini Enterprise Standard and Plus from $30 a seat or a pay-as-you-go option. Jules, Google’s asynchronous GitHub agent, is still around: 15 tasks a day free, 100 on Pro and 300 on Ultra.

Antigravity CLI with Gemini 4 Argon scored 53.2 on the index at $5.86 per task, close to Codex on quality. If you already pay for Google AI Pro, you are getting a competitive coding agent included.

Others Worth Knowing

Kiro (AWS) is built around spec-driven development, with a spec mode next to ordinary prompting. Plans run from 50 free credits to Pro ($20, 1,000 credits), Pro+ ($40), Pro Max ($100) and Power ($200, 10,000 credits), with add-on credits at $0.04.

Muse Code (Meta) is a terminal agent that launched on August 5 with an aggressive price: a contributor tier at about $0.10/$0.20 per million tokens, in exchange for letting Meta train on your prompts. Running Muse Wonder Full, it placed fourth on the index.

Opencode is the open-source option for running any model you like. Grok Build is SpaceXAI’s own coding agent and is included in the Cursor and X bundle.

Related: Kiro IDE: AWS Built an Agentic IDE That Writes Code from Specs, Not Prompts

The architectures are converging fast: Copilot hosts Claude and Codex agents, Devin Desktop runs Codex and Claude, and Claude Code ships in IDEs. What still separates them in October 2026 is which models your subscription actually covers and what a heavy agent day costs. Price that against your own repo before you commit to an annual plan.

Frequently Asked Questions

Which AI coding assistant scores highest on independent benchmarks in 2026?

Claude Code with Opus 5.5 leads the Artificial Analysis Coding Agent Index v2.0 at 61.7 as of October 23, 2026, with Devin’s Fusion harness close behind at 61.5 for about 30% less cost per task. Codex with GPT-6.1 Sol scores 54.1 at $1.50 per task, the best value in the index. Cursor and GitHub Copilot are not ranked because they are multi-model front ends.

Is Cursor or Claude Code better for coding in 2026?

Claude Code is better for end-to-end agent work with frontier Anthropic models included in a flat subscription, and it now runs in VS Code, JetBrains, a desktop app and the web as well as the terminal. Cursor is better if you want to stay in an agent-first editor and are happy using its own Grok and Composer models, since Claude and GPT usage in Cursor is now billed separately at API list price on top of the $30 Premium plan.

How much does Devin cost in 2026?

As of October 23, 2026, Devin has a Free plan, Pro at $20 a month, Max at $200 a month, and Teams at $80 a month plus $40 per full developer seat for up to 200 users. Enterprise pricing is custom. The old ACU-based Core and Teams plans are no longer listed; quotas refresh daily and weekly, and extra usage is billed at API pricing.

Does GitHub Copilot support models other than GPT?

Yes. Copilot’s model list includes Anthropic Claude models such as Opus 5.5, Sonnet 5.5 and Fable 5.1, Google Gemini 4 Argon, xAI Grok 4.7, Kimi K3 and Microsoft MAI-Code-1.1-Flash alongside OpenAI’s GPT-6 family. Since June 1, 2026, usage is billed in GitHub AI Credits at each model’s API rate, and Pro+ and Max subscribers can also delegate tasks to Anthropic’s Claude agent and OpenAI Codex in public preview.

What happened to Windsurf and Gemini CLI?

Cognition renamed Windsurf to Devin Desktop on June 2, 2026, and replaced its Cascade agent with Devin Local; plans and prices carried over. Google replaced Gemini CLI with Antigravity CLI for Google AI Pro, Ultra and free users on June 18, 2026. Enterprise Gemini Code Assist licenses and paid API keys still work with Gemini CLI.