Claude Opus 4.6

Qwen3.8-27B trails Claude Opus 4.6 on Terminal Bench, GPQA Diamond, and HLE benchmarks (73 vs 78.2, 89.2 vs 91.3, and 30.8 vs 40, respectively).

Claude Opus 4.6 has a maximum output of 128K tokens, and Claude Sonnet 4.6 has a maximum output of 64K tokens.

According to reports by The Register and Decrypt, the agent involved was Anthropic Claude Opus 4.6 run by Cursor.

Gemini 3 Pro scored 31.1% on ARC-AGI-2, Claude Opus 4.6 scored 68.8%, and GPT-5.2 scored 52.9%.

For a code review task (3000 input tokens, 800 output tokens), Gemini costs about $0.016 and Claude about $0.035; in a 500,000 input token scenario, Gemini costs about $2.05 and Claude about $5.11.

On the GDPval-AA benchmark for human preference evaluation, Claude Opus 4.6 scored 1606 Elo and Gemini 3.1 Pro scored 1317 Elo.

Claude Opus 4.6's Agent Teams feature allows multiple Claude instances to work together in different roles (planner, executor, reviewer).

In the ATLAS evaluation, the leading model within the up-to-128K-token range is Gemini 3.1 Pro Preview, and the leading model within the up-to-1M-token range is Claude Opus 4.6.

Order effect: swapping the order of two answers, GPT-5.4 changed its verdict 11.4% of the time on one benchmark and 10.5% on another; Gemini 3.1 Pro and Claude Opus 4.6 were between 2.0% and 3.8%.

Manual Extended Thinking (thinking.type: "enabled" plus budget_tokens) is already marked as outdated on Opus 4.6 and Sonnet 4.6, is not accepted at all by the new models, and is now controlled by the effort parameter, with a default of medium.

According to a benchmark run headlessly via Antigravity on the same seven-constraint article planning task (three trials each, n=6, measured on 2026-09-27), Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17/17 on machine-verified constraint adherence, but the median wall-clock time was 23 seconds for Gemini versus 67 seconds for Opus.

Earlier models Claude Opus 4.6 and GLM-5.2 had a success rate of zero on these exploitation tasks.

In the IntLawNER benchmark, Claude Opus 4.6 achieves the best result with a micro F1 of 0.873.

Readers should independently verify whether "GPT-5.4" and "Claude Opus 4.6" are model versions officially released by their respective vendors, and avoid treating unofficial or speculative names as established facts.

Claude Opus 4.6 produced 30 times more errors than GPT-5.4, averaging 2.5 errors per trajectory vs. 0.08, with the difference being statistically significant (p < 0.0001, Cliff's δ = 0.613).

The task success rates of three frontier models are very close, all falling between 84% and 89%.

Qwen3 27B's scores on several programming and agent benchmarks surpassed those of Claude Opus 4.6.

Alibaba's official figures show Qwen3.8 Flash scoring 62.5 on SWE-Bench Pro, leading Claude Opus 4.6 by approximately 9 points; 73.9 on Cowork Bench versus DeepSeek's 45.1; and 55.7 on Job Bench, exceeding Opus 4.6 by approximately 20 points.

On Live Code Bench, Qwen3.8-27B scored 90.3 and Claude Opus 4.6 scored 88.8.

In February 2026, Anthropic, OpenAI, and Google released three coding models—Claude Opus 4.6, GPT-5.3 Codex, and Gemini 3.1 Pro—within 15 days of each other.