Opus 5
A reportedly forthcoming flagship large language model from Anthropic, rumored to succeed the Fable model and representing the fifth generation of the Claude family's Opus series.
Core Facts
Opus 5.5 reaches the level of Fable 5.1 on most tasks, but with approximately 40% lower operating costs and faster speed.
Anthropic states that Opus 5.5 can reduce total costs for typical tasks by approximately 40% and improve output speed by more than 30%.
Claude Opus 5.5 generates output more than 30% faster than Opus 5.
Timeline (last 90 days)
DeepSeek V4.1 Flash version 1017 has 763 billion total parameters and uses the MIT open-source license.
The blogger pointed out four discount factors in the benchmark comparison: different test frameworks (TerminalBench 1017 ran within the ClaudeCode framework while the baseline ran within the DeepSeek framework), Artificial Analysis's own test of the old Flash on TerminalBench scored only 26.8% rather than 31.2%, the competitor comparison used Opus 5 and GPT 5.6 Sol rather than the latest models, and the old model's NL2 repo score is inconsistent across different tables (64.0 vs 65.4).
When Opus 5 was released, Anthropic removed more than 80% of Claude Code's own instructions.
The Opus 5 guide explicitly states that the model already checks itself, and that writing "please check again" in the prompt only increases cost without improving results.
On the Shopping task, all writers (including Opus 5) fail to reach 24 points.
The Opus series has consistently taken a high-end路线 and has deep accumulated expertise in complex reasoning and long-chain tasks.
The settings page supports configuring different models for different features, for example using GPT-7 for quick email Q&A, Opus 5 and S6 Luna for design, Opus 5.5 for complex tasks such as video topic selection, and Whisper 1 for transcription.
Anthropic's three steep jumps in merge volume correspond respectively to the releases of Fable 5 and Mythos 5 (June 9), Opus 5 (July 24), and Fable 5.1 and Mythos 5.1 (September 1).
The author speculates that when using Opus 5 as a coding agent in experiments, it confidently gave incorrect conclusions about correlations, which may be the result of the combined effect of RLVR and RLHF.
Across 14 tests covering approximately 145 skills, Jev found the correct skill in under 5 seconds on average, while using Opus 5 by default required approximately 30 seconds.
40 more timeline events
All Facts (20)
Opus 5.5 reaches the level of Fable 5.1 on most tasks, but with approximately 40% lower operating costs and faster speed.
80%VerifiedAnthropic states that Opus 5.5 can reduce total costs for typical tasks by approximately 40% and improve output speed by more than 30%.
70%VerifiedClaude Opus 5.5 generates output more than 30% faster than Opus 5.
70%VerifiedV4.1 Flash surpassed Opus 5 on the DeepSWE benchmark by a narrow margin of 0.2 points.
70%VerifiedAccounting for real usage scenarios, the actual cost advantage of Opus 5 over Fable is approximately 20% to 25%, with Fable costing about $2.75 per task and Opus 5 about $2.03.
65%VerifiedOpus 5's pass rate at Low Effort is approximately 60%, surpassing Opus 4.8's 59% at the Max level, for which the per-task cost is as high as $13.
65%VerifiedClaude Opus 5's capabilities are close to the level of the previous-generation flagship Opus, but with significantly reduced inference costs.
65%UnverifiedThe release party claimed that the cost of DeepSeek-V4.1-Flash is only 1/40 that of Opus 5 and GPT-5.6 Sol.
70%Partially VerifiedThe per-task cost of GPT-6 Sol High on Automation Bench is only one-tenth that of Opus 5.
65%UnverifiedOpus 5 can handle multiple technical layers in a single generation, including 3D scene construction, collision detection, user input response, and audio triggering.
60%UnverifiedOpus 5 is the flagship model of Anthropic's Claude series, known for its longer context window (capable of handling codebases of hundreds of thousands of tokens) and more cautious output style.
60%UnverifiedClaude Opus 5 scored 30.2% on ARC-AGI-3, compared to only 1.5% for the previous generation.
60%UnverifiedClaude Opus 5 scored 1861 on the knowledge work benchmark GDPval, above the human average.
60%UnverifiedDeepSeek V4.1 Flash version 1017 has 763 billion total parameters and uses the MIT open-source license.
50%UnverifiedThe blogger pointed out four discount factors in the benchmark comparison: different test frameworks (TerminalBench 1017 ran within the ClaudeCode framework while the baseline ran within the DeepSeek framework), Artificial Analysis's own test of the old Flash on TerminalBench scored only 26.8% rather than 31.2%, the competitor comparison used Opus 5 and GPT 5.6 Sol rather than the latest models, and the old model's NL2 repo score is inconsistent across different tables (64.0 vs 65.4).
50%UnverifiedWhen Opus 5 was released, Anthropic removed more than 80% of Claude Code's own instructions.
50%UnverifiedThe Opus 5 guide explicitly states that the model already checks itself, and that writing "please check again" in the prompt only increases cost without improving results.
50%UnverifiedOn the Shopping task, all writers (including Opus 5) fail to reach 24 points.
50%UnverifiedThe Opus series has consistently taken a high-end路线 and has deep accumulated expertise in complex reasoning and long-chain tasks.
50%UnverifiedThe settings page supports configuring different models for different features, for example using GPT-7 for quick email Q&A, Opus 5 and S6 Luna for design, Opus 5.5 for complex tasks such as video topic selection, and Whisper 1 for transcription.
50%