Claude Opus 5: Anthropic’s Cheaper Flagship Pushes Fable 5 on Coding and Knowledge Work

Anthropic is positioning Claude Opus 5 as a cheaper alternative to Claude Fable 5, claiming top scores in coding and knowledge work while keeping token rates at half of Fable 5’s rates. Opus 5 costs $5 per million input tokens and $25 per million output tokens, while Fable 5 is listed at $10 per million input tokens and $50 per million output tokens.

The model becomes the default for Claude Max and the most capable model available on Claude Pro. Anthropic says the 1 million-token context window remains unchanged, while a new Fast Mode increases speed by 2.5x and doubles the price.

According to Anthropic’s benchmarks, Opus 5 leads Frontier-Bench v0.1 agentic terminal coding with 43.3 percent, ahead of Fable 5 at 33.7 percent, GPT-5.6 Sol at 34.4 percent, and Opus 4.8 at 21.1 percent. On GDPval-AA v2, a knowledge-work benchmark, Opus 5 posts an Elo score of 1,861, ahead of Fable 5 at 1,747 and GPT-5.6 Sol at 1,736.

Opus 5 does not lead every test. GPT-5.6 Sol tops DeepSWE v1.1, while Fable 5 and Mythos 5 outperform Opus 5 on health tasks and legal benchmarks, respectively.

The most striking result in the article is Opus 5’s ARC-AGI-3 score of 30.2 percent, a benchmark focused on novel problem-solving without memorized patterns. The Decoder reports that Opus 4.8 scored 1.5 percent and GPT-5.6 Sol scored 7.8 percent, making Opus 5 nearly four times higher than the next-best model listed.

The article also notes there is no Fable 5 result for ARC-AGI-3. It remains unclear whether such a large benchmark lead will translate into everyday use.

The article cautions that token rates alone do not determine real task cost, because token efficiency can change the final bill. Opus 5 offers five effort settings—low, medium, high, xhigh, and max—and Anthropic recommends broad use of low and medium for many tasks, while starting with xhigh for coding and agentic work.

Anthropic says Opus 5 can check and improve its own work through iteration and build tools when needed, including a reported Frontier-Bench task where it wrote a computer vision pipeline to reconstruct a machine part. On safety, Opus 5 allows source code vulnerability research but blocks several cyber uses, and its cyber classifiers reportedly trigger about 85 percent less often than Fable 5’s.