There is a particular irony at the heart of the modern AI coding industry: the most successful tools built on top of Anthropic and OpenAI's models are, by design, the companies most motivated to destroy their dependence on them.
On Thursday, Cursor — the coding assistant built by Anysphere Inc. — made its most decisive move yet toward independence, releasing Composer 2, a proprietary model it claims now rivals the frontier.
The numbers are striking. On CursorBench — drawn from real engineering tasks averaging 352 lines of code across eight files — Composer 2 scored 61.3, a 39% leap over its predecessor and enough to overtake Claude Opus 4.6, which scored 58.2. On Terminal-Bench 2.0, the agent evaluation benchmark for command-line performance, it again bests Anthropic's flagship. Only GPT-5.4 Thinking, at 63.9, remains ahead.
The more consequential number may be price. At $0.50 per million input tokens on the standard tier, Composer 2 costs a tenth of Claude Opus 4.6. Even the fast variant — set as the product default at $1.50/$7.50 per million tokens — significantly undercuts both dominant rivals.
Co-founder Aman Sanger attributes this efficiency to "extreme restraint" in training: Composer 2 is code-only, built exclusively on programming data across a 200,000-token context window. Its reinforcement learning targets long-horizon tasks requiring hundreds of sequential actions — the grinding, unglamorous work of real software engineering.
The Street Verdict: Impressed, But Watching
Where the press release ends, the developer forums begin — and the picture is more complicated.
Several engineers report being genuinely astonished. A Rust developer singled out the model's handling of complex borrowing and lifetime issues as superior to anything they had tested, calling it "impressive for the price point." Some report downgrading from expensive Ultra and Pro plans entirely after running evaluations. The Cursor team's active solicitation of feedback has itself been read as a signal of confidence.
But dissent is pointed and specific. The claim of superiority over Claude Opus 4.6 is disputed outright by users who argue the model falters on complex reasoning tasks. Code quality complaints are recurring: excessive nested conditionals, vague inline comments, a tendency to over-engineer simple problems. Developers describe being impressed in controlled testing, then pulling back from committing it to production work.
The deepest friction, however, is not with the model — it is with the platform. A strand of distrust runs through the community rooted in past allegations against Cursor: opaque model routing, suspected throttling, and benchmarks users cannot independently verify. For these developers, Composer 2 is not evaluated in isolation; it inherits the credibility debt of the product it ships inside.
This is the tension Cursor must now navigate. It has built a model credible enough to be taken seriously and priced aggressively enough to reshape the market. But in an industry where trust is as load-bearing as performance, a strong benchmark score only opens the door. Keeping it open is a different problem entirely.
Sources: https://cursor.com/blog/composer-2