Antigravity vs Cursor vs GitHub Copilot, September 2026: Which Editor, and Whose Models

Cursor now belongs to a company that trains models

On 14 August, Cursor put up a short post: it had officially been acquired by SpaceX. The post promised "access to the largest fleet of GPUs in the world" and pointed to Grok 4.6, released two days earlier, as "an early look at what we can now build together." On 21 September came Grok 4.7, which Cursor's documentation describes as "a frontier model from Cursor and SpaceXAI."

That puts all three of the big AI editors in the same position. Google builds Antigravity around Gemini. GitHub is owned by Microsoft, which now ships a coding model of its own in Copilot's menu. And Cursor, which built its business largely on other labs' models, now sells models it helps train.

This matters because of a second change that landed this year: all three now meter agent work by the token, at a rate that depends on the model. Once that is true, the editor's pricing decides which model is cheap to use by default. The argument of this piece is that the three differ most on exactly that point. Cursor tilts hardest toward its own models, Google gives its models away at the door and sells capacity on top, and GitHub comes closest to neutral, charging list price for everyone's.

All three now bill by the token

GitHub stated the reason most plainly. When it announced the switch in April, it wrote that "agentic usage is becoming the default" and that under the old scheme "a quick chat question and a multi-hour autonomous coding session can cost the user the same amount." A flat price per request made sense when a request was a chat reply. It stops making sense when one request can start an agent that reads a repository for an hour.

Cursor reached the same place in stages. Its help center now says that "all Auto modes bill at the list price of the model each request is routed to," so even letting Cursor pick the model for you is metered at that model's rate. Google restructured Antigravity's plans at I/O in May, merging separate Gemini quotas into one shared pool that you spend across models according to their price ratios.

The practical result is the same in all three. A heavy model burns your allowance faster than a light one, and the editor decides how big each allowance is and which models draw on it.

Cursor's two usage pools

Cursor's individual plans are a free Hobby tier, Pro at $20 a month, Pro+ at $60 and Ultra at $200. Teams costs $40 per user for a Standard seat or $120 for a Premium seat with five times the usage, and Enterprise is custom.

What each plan buys is split into two pools that reset monthly. The Cursor Models pool covers Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5, with what Cursor calls "significantly more included usage." The Other Models pool covers Anthropic, OpenAI, Google and other third-party models, charged at each model's API price. Cursor does not publish the size of either pool. Its docs offer rough guidance instead: daily Tab users "typically stay within included usage," daily Agent users typically spend $60 to $100 a month in total usage, and power users running several agents often pass $200.

Teams pay one more thing. Cursor charges a Cursor Token Rate of $0.25 per million tokens on third-party model requests for Teams and Enterprise customers, including requests made with your own API key. Its own models are exempt: "First-party Cursor models, including Grok and Composer, are exempt." That is the tilt in one sentence.

The house models are cheap enough to make the tilt tempting. Composer 2.5, released on 18 May and built on the same open-source checkpoint as Composer 2 (Moonshot's Kimi K2.5), costs $0.50 and $2.50 per million input and output tokens in its standard version. The Fast version, which is the default, costs $3 and $15. Artificial Analysis scored it 62 on its Coding Agent Index in May and measured $0.07 per task in standard mode and $0.44 in Fast, against $4.10 for Claude Opus 4.7 and $4.82 for GPT-5.5. Both of those rivals have since been superseded, so treat the gap as a snapshot. Grok 4.7 costs $2 and $6 per million tokens, with a 256,000-token standard context window and a 500,000-token long-context mode.

Third-party models are still there. Cursor's model list includes Claude Opus 5.5 at Anthropic's $4 and $20, Gemini 3.8 Flash, and OpenAI's GPT-5.6 Sol. When we checked on 25 September, the list did not yet include OpenAI's GPT-6 models, which GitHub Copilot already lists.

What Antigravity costs, and what it runs

Google launched Antigravity 2.0 at I/O on 19 May as a standalone desktop app for macOS, Linux and Windows. The product page now lists a family: the 2.0 app ("your command center to manage multiple local agents in parallel"), the original Antigravity IDE, a terminal CLI, and an SDK for building custom agents on the same agent runtime.

Its free tier is the most interesting of the three. Google's pricing page lists an Individual plan at $0 with unlimited Tab completions and access to Gemini 3.8, 3.7 and 3.6 Flash, Gemini 3.1 Pro, Claude Sonnet and Opus 4.6, and gpt-oss-120b, under "basic weekly rate limits." (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 Claude models on that list are older than the Opus 5.5 that Cursor and Copilot offer.

Paid capacity comes bundled with a Google AI subscription. Google's May plan update describes the $20 Pro plan as the baseline, a new $100 AI Ultra plan with five times Pro's quota, and the top Ultra plan, cut from $250 to $200, with twenty times. The same update removed AI credits from the base plans; credits are now a top-up for when you run out. Organizations can use Antigravity through Google Cloud on consumption-based pricing.

The trade-off is visibility. Google does not publish Antigravity's quotas in tokens, so you cannot work out in advance what a session costs on a Pro plan. You find the limit by reaching it. The changelog shows a product moving quickly: a /boost command that runs "a multi-agent reasoning pipeline" (2 September), Gemini 3.8 Flash for Enterprise users (3 September), Windows Subsystem for Linux support and live status cards for subagents (22 September), and a /plan command that drafts "a plan you can read and edit before it writes any code."

GitHub Copilot has the widest model menu

Since 1 June, every Copilot plan bills in GitHub AI Credits, and one credit is one cent. The plans: Pro at $10 a month with 1,500 credits, Pro+ at $39 with 7,000, Max at $100 with 20,000, Business at $19 a seat with 1,900, and Enterprise at $39 a seat with 3,900. Code completions and next edit suggestions do not use credits. Chat, agents and code review do, and code review on private repositories also consumes GitHub Actions minutes.

The distinctive part is the menu. GitHub's supported models list includes Claude Opus 5.5, Sonnet 5 and Fable 5.1, GPT-6 Sol, Luna and Astra, Gemini 3.5 to 3.8 Flash, Grok 4.5 to 4.7, Moonshot's Kimi models and Microsoft's own MAI-Code-1.1-Flash. And the rates are list rates: Opus 5.5 at $4 and $20 per million tokens, GPT-6 Sol at $2 and $10, Grok 4.7 at $2 and $6, the same Grok price Cursor charges. There is no separate pool for Microsoft's model; everything draws on the same credits. Of the three, Copilot is the one where the editor's owner has least visibly tilted the meter.

Copilot has also started hosting other companies' agents. In JetBrains IDEs, where agent mode has been generally available since July 2025, you can now pick Claude as the agent (in preview since June), and since July Codex is available as an agent provider in public preview.

What one agent session costs

Take a hypothetical agent session that sends 2 million input tokens and gets back 200,000 output tokens, with no caching. At list prices:

Three things fall out of that table. First, Composer's default Fast mode costs more per token than Grok 4.7, so the cheapest model in Cursor is only cheapest if you switch it off Fast. Second, on Copilot Pro, the 1,500 included credits cover that Opus 5.5 session about once before you start paying. Third, where the spend lands matters as much as the total: in Cursor, the Grok and Composer rows draw on the pool with more included usage, while the Opus row draws on the smaller one.

Real sessions cost less than this. Agents resend the same context over and over, and cached input is much cheaper: GitHub lists Opus 5.5 cached input at $0.20 per million, a twentieth of the fresh rate. And price per token is not cost per task. A model that finishes in fewer steps can beat a cheaper one that wanders, which is why the Artificial Analysis per-task figures are more useful than any rate card. Antigravity is missing from the table because Google does not price its individual plans in tokens.

The agent features have converged

When Antigravity first shipped, its pitch was that it let you supervise several agents at once. That is now table stakes. Cursor's changelog shows Projects (10 September, in beta), where a coordinator agent delegates to subagents for "a feature, a migration, or a full app," plus cloud agents that can run on machines you manage (2 September), and two new bots on Teams and Enterprise plans on 23 September: Rollouts, which watches each pull request as it deploys, and Security Review, which reports exploitable bugs in pull requests. Copilot has agent mode in the editor, a cloud agent, a CLI and code review. Antigravity has parallel local agents, subagents and scheduled tasks.

The differences that remain are about where the work lives. Copilot's center of gravity is the pull request on GitHub. Cursor's is its own editor and cloud. Antigravity's is a desktop app sitting on top of your local machine, with Google Cloud behind it for organizations.

Which one to pick

Pick GitHub Copilot if your team already reviews code in GitHub pull requests and wants to choose models freely, at published list prices, starting at $10 a seat. The catch is that frontier models drain the included credits quickly, so budget for overage or set spending limits.

Pick Cursor if you want the most complete AI editor and are happy for Grok 4.7 or Composer 2.5 to be your everyday models, with Opus or GPT for the hard cases. Go in knowing that the pricing is built to make that split attractive, that teams pay extra on third-party tokens, and that Cursor does not publish its pool sizes.

Pick Antigravity if you want to start free, prefer Gemini, or run on Google Cloud. Expect opaque quotas and older Claude models than the other two offer.

For long, tightly coupled changes across a codebase, any of the three pairs well with a terminal agent; see Claude Code vs Codex.

Whichever you use, the useful thing to do this week is open last month's usage export and see which models actually spent the money. Then run one real ticket through two of them and compare the cost per merged change, not per token. For SEO and Ads teams, the editor is where title-template, markup and redirect fixes get made; the case for making them should come from Search Console and Google Ads figures computed by code, which is how AIO Copilot works.

Frequently Asked Questions