Playbooks · Coding
Cursor's freeze cost, measured
Written November 10, 2026
Cursor's newest OpenAI model costs $7.98 per Terminal-Bench test and passes 37.88% of them. The model OpenAI never shipped to Cursor costs $1.72 and passes 55.05%. The gap is not a rumour — it is a measured benchmark, and it comes from a freeze that started in August.
Why Cursor is stuck on GPT-5.6
On August 28, two weeks after SpaceX closed its purchase of Cursor, OpenAI said it would wind down the contract that supplies its models to Cursor. Reporting of that announcement, from CNBC and InfoWorld, gives four facts: the proposed shutoff is November 12, 2026; OpenAI said it will not give Cursor future models during the wind-down; the stated reason is that OpenAI "cannot be confident that SpaceX will use our technology within our terms of service"; and Cursor CEO Michael Truell said OpenAI models serve about 5% of Cursor user traffic, and that Cursor was talking with OpenAI to resolve it.
The November 12 date is a proposal, and the exact day could still move. Some coverage states it as settled; the reporting it rests on does not. The freeze itself is not in doubt: OpenAI shipped GPT-6 and then GPT-6.1 everywhere else while Cursor stayed on the GPT-5.6 family. Cursor's Models & Pricing page lists three OpenAI models — GPT-5.6 Luna, Sol and Terra — and no GPT-6 model.
The benchmark
Vals.ai runs every model through Terminal-Bench 4.0 with the same harness, Mini-SWE-agent, and publishes accuracy, cost, tokens and wall time per test. The board was last updated November 5, 2026. Terminal-Bench 4.0 is 66 long, real-deliverable tasks, each scored by its own verifier, with each model's score the mean of three full runs.
| Model | In Cursor? | Accuracy | Cost / test | Tokens / test | Cost per solved test |
|---|---|---|---|---|---|
| GPT-6.1 Sol | No | 55.05% | $1.72 | 2.86M | $3.12 |
| GPT-5.6 Sol | Yes | 37.88% | $7.98 | 10.49M | $21.07 |
| GPT-6.1 Terra | No | 49.49% | $2.16 | 16.82M | $4.36 |
| GPT-5.6 Terra | Yes | 22.73% | $5.60 | 13.66M | $24.64 |
| GPT-6 Luna | No | 13.64% | $0.35 | 15.38M | $2.57 |
| GPT-5.6 Luna | Yes | 11.62% | $0.73 | 19.00M | $6.28 |
GPT-6.1 Sol scores 17.17 points higher than GPT-5.6 Sol and costs 4.6 times less per test. The price cut explains only half of it: GPT-6.1 Sol is about twice as cheap per token, and it used 2.86 million tokens per test against GPT-5.6 Sol's 10.49 million. Terra is the reverse case — GPT-6.1 Terra burns more tokens than GPT-5.6 Terra and still costs less, because it is half the price or less per token, and it more than doubles the score.
On the same board, Kimi K3.1 Pro scores 63.64% at $4.57 per test — the highest score of the models compared here, at about 38% of Claude Opus 5.5's $12.04. And GPT-5.6 Sol is not a bargain inside Cursor either: Claude Sonnet 5.5 scores 26.26 points higher for $2.77 more per test, and costs less per solved test.
What a Cursor team can do
First, the per-token price is the wrong number to compare. GPT-6.1 Sol is twice as cheap per token and 4.6 times as cheap per test; Terra is the case where more tokens still means a cheaper test. Judge a model on cost per solved test, not on its list price.
Second, a personal OpenAI API key is only a partial way out. Cursor's bring-your-own-key page says OpenAI keys cover standard, non-reasoning chat models, and that custom keys only work with chat models — Tab keeps using Cursor's built-in models. On Teams and Enterprise, eligible third-party requests carry a $0.25 per million token Cursor Token Rate, own key included; models with a cache-read price strictly below $0.10 per million are exempt. GPT-6.1 Sol's cache reads cost exactly $0.10, so it would pay the rate, if your key can reach it in Cursor at all. Check the model picker before planning a team migration around it.
Third, the models that replace GPT-5.6 Sol inside Cursor are already there. Where GPT-6.1 Sol does run: the OpenAI API ($2 / $10 per million in / out), Codex on ChatGPT Plus, and GitHub Copilot, where GitHub lists it as generally available. Rules, prompts and agent instructions tuned for GPT behave differently on Claude or Gemini — re-run your hardest tasks on the replacement while the old model still answers.
Limits
- These are per-test figures on Terminal-Bench 4.0, not a monthly bill. A test is one long task in a sandboxed terminal; your own workload will differ.
- The November 12 shutoff is OpenAI's proposed date, reported by CNBC and InfoWorld. It could move, and the wind-down could be resolved. Treat it as a deadline to plan around, not a settled fact.
- Cost per solved test is a division of Vals.ai's cost per test by its accuracy. The table figures are as reproduced from the Vals.ai board by MegaBrain and four other independent reproductions: the board renders its numbers with JavaScript, so they did not appear in a direct text read of the board itself.
- GPT-6.1 Sol was slower per test than GPT-5.6 Sol — 45 minutes 12 seconds against 33 minutes 46 seconds on average. The cheaper, stronger model takes longer.
Related playbooks
- Move Cursor work before the OpenAI shutoff
- OpenAI's Cursor cutoff day came and went
- Claude Code 2.1.317: schedule warnings and cost reports
- Cursor Agent SDK: define an agent as a folder
Sources
- Cursor's Best OpenAI Model Costs $7.98 Per Coding Test — MegaBrain
- Terminal-Bench 4.0 Leaderboard and Methodology — Vals AI
- Models & Pricing — Cursor Docs
- What OpenAI's Cursor Cutoff Means for Your Team — HackerNoon
Back to: Coding playbooks hub · Also: Run your own agent bake-off on one real task · Check every package name your agent invents