Tutorials, deep dives and product notes — built for developers.
Cognition’s FrontierCode 2.0 is a new 132-task mergeability benchmark. Interactive leaderboard with best-effort model scores, task pass rates, flag rates, harnesses, and actual evaluation cost per rollout. Updated October 30, 2026.
GPT-6.1 Terra vs Claude Sonnet 5.5: Terra is ahead at low and medium effort, then Sonnet pulls 7.9 points ahead at max while costing 19× more per task ($7.60 vs $0.40). Full effort-ladder comparison, vendor and independent benchmarks, hallucination rates, and a routing guide.
Claude Haiku 5.5 vs GPT-6 Luna: identical $0.10/$0.50 list prices, but Haiku scores higher on Artificial Analysis (43 vs 38) while Luna costs about 3× less per task at max effort. Price cliffs at 100K and 272K, vendor and independent benchmarks, effort ladder, and a routing guide.
GPT-6.1 Sol matches Claude Opus 5.5's default index score for 46% less per task and ties it at low effort for 76% less — then stops at index 52, below every Opus effort above default.
OpenAI shipped GPT-6 Sol 90 minutes after Claude Opus 5.5, at half the token price. Artificial Analysis ran both on the same tests: Sol is cheaper to any index score up to about 44, then runs out of headroom. Every benchmark, both effort ladders and the real bill.
Opus 5.5 matches GPT-6 Astra on Terminal-Bench 4.0 for roughly 40% of the cost per task, with 5x cheaper cache reads. Astra keeps uncontested leads in science, maths and gated offensive security — and OpenAI says it is harder to monitor.
Sixty days apart: a 29.7-point jump on scientific research, a 200,000-line codebase audit that drops from 20+ hours to under 3, and a 40% cost cut — plus four API changes that will 400 your Opus 5 integration.
Opus 5.5 wins eight of the nine benchmarks Anthropic publishes against Fable 5.1 at 2.5x lower token price - but four of those margins sit inside the disclosed error bars. Includes the HAProxy same-task test, the 16-0 research-report result and the one-way thinking-block handoff.
Anthropic's Claude Opus 5.5 matches Fable 5.1 on most work at $4/$20 per million tokens: 66.4% on Terminal-Bench 4.0, 1,846 GDPval-AA Elo, and 40% lower cost per task than Opus 5. Full benchmark table with footnotes, real tester results, the four breaking API changes, and where GPT-6 Astra still wins.
DeepSeek V4.1 Flash wins 5 of the 15 benchmarks it shares with Claude Opus 5 — by an average of 1.94 points, while undercutting it by up to 41.7× on output price. Opus 5's 10 wins average 10.19 points. Full vendor data, community test reports, pricing math, charts and a routing verdict.
Terminal-Bench 4.0: the tbench.ai public board remains led by Claude Opus 5.5 (64.8%). Adds Kimi K3.1 Pro (65.7%, Moonshot), Gemini 4 Carbon (64%, Artificial Analysis), GPT-6.1 Terra (50.3%) and MiniMax M3.1 Flash (38.1%) as separately labeled runs. Updated Oct 30, 2026.
OpenAI's GPT-6 Astra vs Anthropic's Claude Fable 5.1: every benchmark, the independent indices, real-world head-to-heads, pricing deep dive, and who actually wins your workload.