Cursor bets on cheap open-weight post-training. The numbers make a strong case.

Cursor shipped Composer 2.5 on May 18. It's the fourth Composer model in seven months, which is a pace that would have sounded like fiction two years ago. The release post leads with benchmarks. The actual story is underneath them.

Composer 2.5 is built on Kimi K2.5 — an open-weight checkpoint from Moonshot AI. That base accounts for roughly 15% of the total training compute. Cursor spent the other 85% on their own post-training pipeline: reinforcement learning, continued pretraining, 25× more synthetic coding tasks than Composer 2, and a targeted text-feedback technique that applies corrections to specific spans within a generation rather than grading the entire output pass/fail. Standard RL post-training grades the whole rollout and backpropagates from there. What Cursor is describing, per their announcement post, is localized feedback — the model learns which part of a completion was wrong, not just whether the output cleared a threshold.

That last part isn't in any headline. It should be.

Source spread

• cursor.com/blog/composer-2-5 — builder. Primary source; benchmark tables, pricing, post-training methodology.
• The New Stack — builder. Independent coverage of the compute split and training pipeline details.
• DataCamp — builder. Clean benchmark comparison table with Opus 4.7 and GPT-5.5 context.
• ChatForest — skeptic. Flags CursorBench as first-party and notes where Opus 4.7 retains the edge.

Pros & cons

What's real:

• Three external benchmarks. On two (SWE-Bench Multilingual and Terminal-Bench 2.0), Composer 2.5 is essentially tied with Opus 4.7. On CursorBench v3.1, it leads.
• Standard tier is genuinely 10× cheaper on input: $0.50/M vs $5.00/M. The output ratio is similar.
• The jump from Composer 2 to 2.5 on Terminal-Bench — from 61.7% to 69.3% — is a 7.6-point gain in a single version. That's large.
• Open-weight base means Cursor is no longer entirely dependent on Anthropic or OpenAI for model supply. That's a different kind of infrastructure resilience.

What deserves skepticism:

• CursorBench v3.1 is Cursor's benchmark. The company built it, runs it, and reports the scores. Not invalid — but treat it as supplementary to the external evals.
• Fast tier pricing ($3.00/M input, $15.00/M output) is not 10× cheaper. It's comparable to other premium models. "10× cheaper" only holds at the standard tier.
• "Improved communication style and effort calibration" is a quality claim that doesn't appear in any benchmark. I'll believe it when I use it.
• Composer 3 will ship. Four models in seven months. Factor in migration costs if you build tightly against this API version.

Further reading

• Introducing Composer 2.5 — Cursor — primary source, benchmark tables and methodology
• Cursor bets on cheaper coding with Composer 2.5 and Kimi K2.5 — The New Stack — independent coverage of the compute split and training pipeline
• Composer 2.5: benchmarks, pricing, and how it compares — DataCamp — clean comparison tables with Opus 4.7 and GPT-5.5
• Technical report: Composer 2 — Cursor — prior methodology context for evaluating the technique claims