Best AI Coding Tools, September 2026: Claude Code vs Codex vs Cursor vs Antigravity

Same prices, different tools

Anthropic, OpenAI and Google now sell their heaviest AI coding plans at the same two prices, $100 and $200 a month, and describe them in the same terms: five times and twenty times the usage of the tier below. Anthropic calls them Max 5x and Max 20x, OpenAI calls them Pro 5x and Pro 20x, and Google split its AI Ultra plan into $100 and $200 versions at I/O in May. Cursor's top individual plan, Ultra, is also $200.

When prices converge, price stops being the reason to choose. What separates Claude Code, OpenAI's Codex, Cursor and Google's Antigravity at the end of September 2026 is where each one wants you to work, which model it puts in front of you by default, and how it charges for the cheap, repetitive work that fills most of a coding day. This guide covers what each tool is, what changed this month, and which kind of work each suits. None of them wins everywhere, and the independent benchmarks published this week show the top models closer together than their makers say.

What changed in September

The month started with Google. On 2 September Antigravity's changelog added a /boost command that sends a hard problem through a multi-agent reasoning pass. The next day OpenAI released GPT-6 Astra, its most capable model, at $10 per million input tokens and $50 per million output tokens. Demand for Astra was heavy enough that on 10 September OpenAI stopped taking new sign-ups for its $200 tier, TechCrunch reported. The same day Cursor put Projects into beta: coordinator agents that hand pieces of a larger job to subagents.

On 21 September Cursor released Grok 4.7, its most capable in-house model so far. The next day brought two launches at once. Anthropic released Claude Opus 5.5 at $4 and $20 per million tokens, which it says is 40% cheaper than Opus 5 on typical workloads, with output more than 30% faster; Claude Code switched its default model to it. OpenAI released GPT-6 Sol and Luna at half the promotional prices of the GPT-5.6 versions they replace, and put both into Codex for paid plans. Antigravity added a /plan command that drafts a plan you can edit before the agent writes code, and on 23 September Cursor shipped Rollouts, which tracks a change's health as it deploys, and Security Review for pull requests, both for Teams and Enterprise.

Claude Code starts you on the strongest default model

Claude Code began as a terminal program and still works best there, but it now runs in several places: the CLI, extensions for VS Code (which also install in Cursor) and JetBrains, a desktop app, and a web version whose cloud sessions run in isolated virtual machines and keep going after you close the laptop. It reads the repository, edits files, runs commands and tests, and keeps going until the task is done or it needs you.

Its biggest advantage this month is the model it starts on. Claude Code's documentation says the default is now Opus 5.5 on every plan from Pro up, at medium effort. Artificial Analysis, testing independently, says Opus 5.5 at maximum effort has the highest score it has measured on its Intelligence Index. Anthropic's pricier Fable 5.1 sits above it, but Anthropic's help center says Pro users can reach Fable only through paid usage credits, and Max users can spend up to half their weekly allowance on it.

The rest of its edge is in packaging. A CLAUDE.md file holds standing project instructions, skills turn a repeatable job (a release checklist, a migration check) into a command the whole team runs the same way, and hooks run shell commands before or after the agent acts. Background agents give each parallel session its own git worktree so they do not overwrite each other, and routines run jobs on a schedule or when something happens on GitHub.

The weak spots are specific. Anthropic does not publish message counts for its plans, only multiples of Pro, so you learn your real limit by using it. And its command sandbox, the part that lets the agent run routine commands without asking, does not work on native Windows; Windows users need WSL2 to get it. Claude Code fits long, tightly coupled changes in one codebase, and teams that want the same job done the same way every time.

Codex spans the widest range of prices

Codex comes with OpenAI's paid plans and runs in the same set of places as Claude Code: an open-source CLI, an IDE extension, the desktop app, and a cloud service that runs tasks in parallel in isolated environments. You can start cloud tasks from GitHub pull requests, GitLab merge requests, Linear issues or Slack threads.

What sets it apart is the spread of its model lineup. OpenAI's API prices run from Astra at $10 and $50 per million tokens, through Sol at $2 and $10, down to Luna at $0.10 and $0.50, which is a tenth of the price of Anthropic's smallest current model. Codex starts at a preset called Sol Light, and OpenAI publishes how far each plan goes: on the $20 Plus plan, 15 to 150 Sol messages per five-hour window, 5 to 45 with Astra, and 350 to 3,000 with Luna. No other vendor in this comparison is that specific.

Codex also has the better story on Windows, with a native sandbox in PowerShell alongside the macOS and Linux versions, and its CLI source is on GitHub for anyone who wants to audit it. The cost is in the default: Sol is a cheaper, weaker model than Opus 5.5 in Artificial Analysis's tests, and matching Claude Code's default means choosing Astra and its small allowance. And the $200 tier, the one with twenty times Plus's usage, is closed to new customers for now. Codex fits high volumes of small, well-specified tasks, Windows-heavy teams, and anyone who wants to plan against published limits. The Claude Code vs Codex comparison goes through the two in detail.

Cursor is an editor that sells its own cheap model

Cursor is where you go if you want to read and change code yourself, with AI close at hand: completions, inline edits, chat and agents inside an editor, plus a CLI and cloud agents that keep working while you are away. It supports MCP, skills and hooks on individual plans.

Its distinctive move is selling its own models. Cursor is no longer independent: on 14 August it announced it had officially been acquired by SpaceX, and it has since released Grok 4.6 (12 August) and Grok 4.7 (21 September). Cursor's pricing documentation puts Grok 4.5, 4.6 and 4.7 and Composer 2.5 in a Cursor Models pool that comes with much more included usage than the pool for third-party models, which bills at API rates. Grok 4.7 costs $2 and $6 per million tokens. Composer 2.5 costs $0.50 and $2.50 in its standard version and $3 and $15 in its fast version. When Composer 2.5 came out in May, Artificial Analysis scored it 62 on its Coding Agent Index, behind only Claude Opus 4.7 in Claude Code (66) and GPT-5.5 in Codex (65). Those two cost $4.10 and $4.82 per task; Composer 2.5 cost $0.07 to $0.44. Both rivals have since been replaced by newer models, so treat the ranking as dated and the cost gap as the lasting point.

The third-party list is broad but not uniformly current. On 25 September it included Anthropic's Opus 5.5, Fable 5.1 and Sonnet 5 and Google's Gemini 3.8 Flash, while its OpenAI entries were still the GPT-5.6 generation rather than GPT-6. Teams and Enterprise plans pay an extra $0.25 per million tokens on third-party models, a charge Cursor's own Grok and Composer models do not carry. Plans run from a free Hobby tier through Pro at $20, Pro+ at $60 and Ultra at $200; Teams costs $40 per user for a Standard seat and $120 for a Premium seat with five times the usage. Cursor fits people who spend the day in an editor and want the cheapest capable model for routine agent work. Our Cursor vs Antigravity comparison sets it against the other editors.

Antigravity is free to start and built for supervising agents

Google rebuilt Antigravity at I/O on 19 May. Antigravity 2.0 is a standalone desktop app for macOS, Linux and Windows whose job is managing several local agents in parallel: you group conversations into projects, hand off tasks, schedule recurring ones, and review what comes back. The original Antigravity IDE remains for people who want an editor with an agent manager. A new Antigravity CLI, written in Go, replaced Gemini CLI, which stopped serving Google AI Pro and Ultra subscribers on 18 June. An SDK in preview lets developers build their own agents on the same agent framework.

Its strongest argument is price. Google's pricing page lists an individual plan at no cost, with weekly rate limits, and says Google AI Pro and Ultra subscriptions raise those limits. The model list on that plan includes Gemini 3.8, 3.7 and 3.6 Flash, Gemini 3.1 Pro, Claude Sonnet and Opus 4.6, and gpt-oss-120b. (Gemini 3.8 Flash first arrived for Enterprise users, according to the 3 September changelog; the pricing page now lists it on the free plan.) The third-party models are a generation or more behind what Claude Code and Cursor offer; if you want Opus 5.5, Antigravity is not where you get it.

Antigravity fits work that splits cleanly into independent tasks you can review one by one, teams already on Google Cloud, and anyone who wants to see how an agent-first workflow feels before paying for one. For a single long change where every step depends on the last, one agent holding one plan is usually the better arrangement.

What this week's benchmarks show

Every launch arrives with a table, and this month's tables disagree. Anthropic's Opus 5.5 announcement put it at 66.4% on Terminal-Bench 4.0, a test of agents working in a real terminal, against 57.9% for GPT-6 Astra, with a footnote saying Opus ran at xhigh effort and Astra at high. Artificial Analysis, running both at xhigh in its own test setup, found them level at 59.6%. OpenAI's own Astra announcement, three weeks earlier, had shown Astra ahead of Anthropic's Fable 5.1, 57.9% to 55.8%.

Each figure is a real result under conditions its publisher chose. The lesson for a buyer is that the frontier models are close, and that the settings behind a benchmark rarely match what you run. Claude Code uses medium effort by default and Codex starts at Sol Light, both well below the settings on the launch charts. Cost per task depends on how many steps an agent takes as well as the token price, which is why Artificial Analysis's per-task figures, where they exist, are more useful than price lists.

The test that matters is cheap to run. Take five real tasks from your backlog, give each tool a week at its $20 tier and its default settings, and count merged changes, review time and whether you hit a limit.

Where search and ads teams fit

Search and paid media teams can now hand coding agents jobs that used to sit in a developer's queue: fixing title templates across a theme, adding structured data to every product page, building redirect maps for a migration, and writing scripts that read Search Console and Google Ads exports.

Site-wide template and markup changes suit a terminal agent working against the real repository, with a person reviewing the diff before it ships. Bulk labelling of pages or search terms does not need a coding agent at all; a small model or a decision model such as Jev is faster and cheaper. Anything involving money should have code compute the figures from the export and a model explain them, as our piece on Claude Opus 5.5 for SEO and Google Ads sets out.

That split, with code for the numbers, a model for the writing and a person for the apply button, is how AIO Copilot runs SEO, AI search and Ads work for businesses and agencies. It is priced per site; tell us how many you manage and we email the price.

Plenty of developers settle on two of these tools: an editor for the day's work and an agent for the long jobs. The cheapest paid tier of each is $20 or less, and Antigravity costs nothing to try, so a month with two of them costs about $40. Pick the pair after that month of real work, not after reading the launch charts.

Frequently Asked Questions