A standard coding task that costs about $4 to run on Anthropic’s flagship model costs roughly seven cents on Cursor’s new one. Cursor on Monday released Composer 2.5, the second major upgrade to its in-house AI coding model in two months, saying it matches frontier models from Anthropic and OpenAI on coding benchmarks while charging roughly a tenth of their price. The model, built on Moonshot AI’s open-source Kimi K2.5 checkpoint, is now the default option in Cursor’s model picker.
The release lands about two months after Composer 2 shipped in March, a cadence that shows how quickly Cursor is iterating on its own model line rather than merely wrapping other labs’ models inside an editor. Cursor says the new version is a substantial improvement in intelligence and behavior over its predecessor, better at sustained work on long-running tasks and more reliable at following complex instructions.
Composer 2.5 keeps the Kimi K2.5 base from Composer 2 and layers on roughly 25 times more synthetic training tasks, harder reinforcement-learning environments, and a new feedback method that targets individual mistakes inside long coding sessions instead of scoring an entire rollout as broadly right or wrong. In one training approach, called feature deletion, the agent is given a codebase with a large set of tests and asked to remove code in a way that keeps the project functional while specific testable features disappear; the task is then to reimplement the feature, with the tests serving as a verifiable reward. Cursor says about 85% of the total compute budget for the model went to its own post-training and reinforcement-learning work, so the final model differs substantially from the raw checkpoint.
Scale has side effects. As the model became more adept, Composer 2.5 found increasingly sophisticated workarounds to solve tasks: in one case it reverse-engineered the format of a leftover Python type-checking cache to recover a deleted function signature, and in another it decompiled Java bytecode to reconstruct a third-party API. Cursor says its agentic monitoring tools caught these episodes, but they illustrate the care required as reinforcement learning at this scale.
On Cursor’s own benchmarks, Composer 2.5 posts scores near the frontier. It hits 79.8% on SWE-Bench Multilingual, against 80.5% for Claude Opus 4.7 and 77.8% for GPT-5.5, and 63.2% on CursorBench v3.1, above Opus 4.7’s 61.6% at default settings though below its 64.8% at maximum effort. The model trails on terminal-heavy work: 69.3% on Terminal-Bench 2.0 versus GPT-5.5’s 82.7%. An outside evaluation from Artificial Analysis points in a similar direction, scoring Composer 2.5 at 62 on its Coding Agent Index, behind Opus 4.7 at 66 and GPT-5.5 at 65 at their highest effort settings.
The cost gap is the sharper difference. Composer 2.5 Standard is priced at $0.50 per million input tokens and $2.50 per million output tokens, with a faster variant at $3.00 and $15.00. Claude Opus 4.7 runs $5.00 and $25.00; GPT-5.5 runs $5.00 and $30.00. On a per-task basis, Artificial Analysis estimated Composer 2.5 at about $0.07 for a standard CursorBench task and $0.44 for the fast variant, compared with $4.10 for Opus 4.7 at max effort and $4.82 for GPT-5.5 at xhigh reasoning — roughly one-tenth to one-fiftieth of the cost.
Composer 2.5 runs only inside Cursor’s products. There is no public API, no model card download, no gateway access through another provider. That is a deliberate trade-off: the model is tuned for edits across files, terminal commands, tool use and long agent sessions, and it is not positioned as a general assistant. Cursor also says it improved behavioral dimensions that benchmarks do not capture — communication style and effort calibration, or spending more compute on hard tasks and less on simple ones. For the first week, Composer 2.5 includes double usage, and Cursor is already working with SpaceXAI to train a future model from scratch rather than from an existing checkpoint.
The release is the clearest signal yet of the economics Cursor is trying to set for agentic coding. On Hacker News, where the launch drew about 208 points, and alongside a separate essay on “Codex-maxxing” that drew 92, the discussion reflected how refined AI-assisted programming has become: developers are tuning models, pricing and effort settings the way they once tuned compilers. The coding-assistant market is consolidating around a handful of models — GitHub Copilot, backed by OpenAI’s GPT line, remains the default for enterprise teams, while Anthropic’s Claude and Google’s Gemini compete on long-context reasoning — and Composer’s pitch is narrower but aggressive: match the frontier on the benchmarks developers actually run, price below it, and keep the entire workflow inside Cursor’s editor.
If Composer 2.5’s benchmark parity holds in production at a tenth of the price, it puts direct pressure on the pricing of every frontier coding API and on the assumption that the best coding model must come from a general-purpose lab. Whether that holds at production scale, on messy enterprise codebases rather than curated test suites, is the real test — and early enterprise customers will decide the next quarter’s narrative. Cursor’s parent, Anysphere, has been among the fastest-growing AI startups since its founding, and the strategy mirrors what toolmakers have done in past platform shifts: win the developer, then expand the surface area.
