Leaderboard Check: Where the Models Stand at Month's End
With the month's releases settled, the September 29 leaderboard snapshots tell a coherent story: Claude Opus 5.5 leads Anthropic's pack at 86.9, GPT-6 Astra claims 96.0% on GPQA Diamond, and the open-weight tier keeps closing in.
Leaderboard Check: Where the Models Stand at Month's End
The September 29 benchmark snapshots offer the clearest picture yet of a frontier race being fought on three fronts: raw capability, price-performance, and open availability. Here is where things stand as the month closes.
Anthropic's table. Claude Opus 5.5 tops the BenchLM leaderboard for Anthropic models — 86.9 in the September 29 snapshot, 86.3 as of the October refresh. The 5.5 series displaces its own predecessors quickly — Opus 5 from July and Sonnet 5 from June now occupy the supporting tiers, while the 4-series was retired outright on June 15. Sonnet 5.5 and Haiku 5.5, documented in system cards this month, were still pending in the published tables at the time of writing — both have since entered them, with Sonnet 5.5 at roughly 83.8 (81.3% SWE-bench Pro) after its September 28 GA and Haiku 5.5 following its October 7 release.
OpenAI's table. GPT-6 Astra's 96.0% on GPQA Diamond is the month's headline number — graduate-level science reasoning at a level that would have been frontier-defining in any previous year. The five-level reasoning control complicates direct comparison, since results depend on the effort setting chosen.
The open-weight table. September comparisons rank DeepSeek V4, Kimi K3, GLM-5.3, and Qwen3.8 at the head of the open cohort, evaluated on SWE-Bench coding, GPQA reasoning, price, and license. The gap to closed frontier models on these axes has narrowed to the point where workload fit, not raw capability, decides the choice.
What the tables miss. Benchmarks run on clean, well-specified problems; production work is neither. Latency, reliability, tool-use robustness, and long-context behavior under real retrieval noise remain badly under-measured — and are frequently where actual deployments succeed or fail.
Use the leaderboards to shortlist, then evaluate on your own traffic. That advice has never changed, and this month's numbers — impressive as they are — do not change it either.