GPT-5.5 Leads SWE-bench Verified by 1.1 Points. Opus 4.7 Leads SWE-bench Pro by 5.7. Pick the Right One.
After two weeks of post-launch data accumulation, the GPT-5.5 versus Claude Opus 4.7 coding comparison has settled into a clear split. Neither model dominates everywhere. Which benchmark you weight determines which model you pick.
The Numbers
| Benchmark | GPT-5.5 | Claude Opus 4.7 | Winner |
|---|---|---|---|
| SWE-bench Verified | 88.7% | 87.6% | GPT-5.5 (+1.1) |
| SWE-bench Pro | 58.6% | 64.3% | Opus 4.7 (+5.7) |
| Terminal-Bench 2.0 | 82.7% | 69.4% | GPT-5.5 (+13.3) |
| AA Intelligence Index | 60 | 57 | GPT-5.5 (+3) |
| LM Arena ELO | ~1504 | 1504 | Tied |
Why the Benchmarks Diverge
SWE-bench Verified uses 500 GitHub issues, Python only. It has been public since 2023. Every frontier model has been trained on its surrounding ecosystem — repositories, issues, pull requests, documentation. Scores above 85% are almost certainly benefiting from proximity to the training data, if not direct contamination.
SWE-bench Pro runs 1,865 tasks across 41 repositories in Python, Go, TypeScript, and JavaScript. It was released after GPT-5.4 and Opus 4.6 were trained, which means newer models have far less overlap with their training distributions. GPT-5.5’s 58.6% versus Opus 4.7’s 64.3% on Pro is a 5.7-point gap, compared to the 1.1-point Verified gap. The harder, newer benchmark favors Opus 4.7 more clearly.
Terminal-Bench 2.0 is different again — it tests command-line reasoning, multi-step orchestration, and tool coordination in a bash environment. GPT-5.5 wins by 13.3 points here. This benchmark correlates with agentic CLI work: pipelines, scripts, automated deployments.
What the Split Means for Agent Selection
The right model depends on the task:
Choose GPT-5.5 when:
- The work is agentic and terminal-bound (CLI orchestration, script generation, automated ops)
- You need broad composite intelligence (AA Index score leads by 3 points)
- Your task is not primarily long-form code review or multi-repo refactors
Choose Claude Opus 4.7 when:
- Production coding quality is the primary concern (SWE-bench Pro is the harder test)
- You need multi-language support (Pro’s Go, TypeScript, JavaScript coverage matters)
- Human preference tasks are in the loop (Arena ELO is tied but Opus leads expert knowledge-work tasks)
The 5.7-point Pro lead is not a rounding error. At 64.3% versus 58.6%, Opus 4.7 resolves roughly one in every 17 additional bugs that GPT-5.5 fails on. For a team running hundreds of agent-submitted PRs per day, that is a meaningful false-positive rate difference.
Pricing Reality
GPT-5.5 is available through ChatGPT and Codex. Opus 4.7 is $5/$25 per million input/output tokens. OpenRouter’s analysis of production agentic workloads found Opus 4.7 runs about 27% more expensive per task in the typical agentic context window range, partially offset by 72% fewer output tokens from GPT-5.5. Total cost-per-resolved-issue at scale likely favors GPT-5.5 for Terminal-Bench-style work and Opus 4.7 for complex refactors where first-pass success rates matter more than token volume.
The April 2026 coding leaderboard has two winners. The question is which benchmark you’re actually running in production.