Anthropic's Opus 5.5 Tops Cursor's Coding Benchmark at 40% Lower Cost

Cursor has enabled Opus 5.5, Anthropic’s newest flagship model. It now leads Cursor’s internal coding benchmark, with Cursor and Anthropic reporting stronger task completion, lower token use, and lower cost per completed task than Opus 5.

At maximum reasoning effort, Opus 5.5 scored 57.8% on CursorBench. Anthropic’s Opus 5.5 details also report gains in agentic coding and knowledge work, plus a 40% cost reduction against Opus 5 on typical workloads.

CursorBench rewards more reasoning

CursorBench evaluates coding agents on ambiguous, multi-file tasks drawn from real Cursor sessions. These tasks require models to inspect repositories, choose tools, edit related files, and validate changes, which reflects the work performed by editor agents in large codebases.

Opus 5.5 at default effort scored 0.7 percentage points above Fable 5.1 at maximum effort. Raising Opus 5.5 to Max added another 5.3 points, though the larger reasoning budget can increase latency and credit consumption.

Efficiency reaches the tool loop

Anthropic compares Opus 5.5’s default CursorBench score with GPT-5.6 Sol’s 41.7%, a difference of 10.8 percentage points, and estimates that Opus 5.5 costs about one-third as much per task in that comparison. Launch partners also reported fewer tokens, tool calls, and agent steps:

• GitHub: Tests across Copilot CLI and VS Code placed Opus 5.5 among the models using the fewest tokens and steps. In VS Code, it completed more terminal tasks than Opus 5 in less than half as many steps.
• Kiro: The model used 40% fewer calls and half as many tokens.
• Box: Opus 5.5 used one-third as many tokens as Opus 5 and produced 40% less output without reducing accuracy.
• Factory: Opus 5.5 at medium effort matched Opus 5 at high effort while using 20% to 25% fewer tokens.

Anthropic also reports that Opus 5.5 ported HAProxy from C to Rust in 9.5 hours, at 51% lower cost than Fable 5.1’s 12-hour run. In another test, it audited 200,000 lines of code in under three hours; Opus 5 required more than 20 hours. These vendor and partner results use different workloads and harnesses, so they provide operational examples rather than a standardized cost comparison.

Rivals still hold some leads

Astra remains ahead on AutomationBench and Terminal-Bench-Science. Anthropic also says its internal experience shows a narrower gap between Opus 5.5 and Claude Fable 5.1 than the published benchmark scores suggest.

Small leaderboard differences can disappear when repositories, prompts, tools, and validation rules change. For production agents, completion rate should be evaluated alongside token use, tool calls, wall-clock time, retries, and the amount of human correction required.

Match the mode to the job

Cursor lists Opus 5.5 in its model picker with standard and Max modes. Max allocates the larger reasoning budget used for the 57.8% CursorBench result. On Cursor’s usage-based plans, Opus models draw from the third-party model pool, so sustained Max sessions consume credits faster than Sonnet-class sessions.

1. Use default effort for multi-file implementation, debugging, and repository analysis where strong reasoning must stay within a practical budget.
2. Use Max effort for long-running migrations, cross-repository refactors, code audits, and tasks where retries or supervision would cost more than additional inference.
3. Use a Sonnet-class model for small edits, rapid iteration, and latency-sensitive work that does not require a large reasoning budget.

Track cost per completed task

Opus models have historically paired high capability with high operating costs. Anthropic’s claimed 40% reduction, combined with partner reports of fewer tokens and tool calls, could make Opus 5.5 practical for a broader set of agent workloads.

Teams evaluating the model should measure the full cost of finishing representative tickets. A model with a higher token price can still cost less when it uses fewer tokens, completes tasks in fewer steps, and requires fewer retries or manual corrections.