Anthropic's Claude Opus 5 Matches Fable 5 at Half the Cost, Keeps Opus 4.8 Pricing Intact

Anthropic shipped Claude Opus 5 on Friday, July 24, 2026, positioning it as a model that closes in on the capability of its own top-tier Fable 5 system while charging roughly half the token price. The company is calling it the most capable Opus model it has built, and it now replaces Claude Opus 4.8 across the product line.

The pricing story is where things get interesting. Opus 5 holds steady at $5 per million input tokens and $25 per million output tokens, identical to what Opus 4.8 charged. Anthropic frames this as delivering near-Fable-5 intelligence without raising Opus-tier prices, rather than cutting them outright, a distinction some early coverage blurred by describing the pricing as simply "unchanged" versus "the same as 4.8."

Benchmark gains are substantial. On SWE-bench Verified, Opus 5 hit 96.0%, and it scored 79.2% on the tougher SWE-bench Pro. Its SWE-bench Multimodal result jumped from Opus 4.8's 38.4% to 59.4%, a sizable leap for a single model generation.

On ARC-AGI-3, a benchmark designed to probe novel reasoning rather than memorized patterns, Opus 5 scored 30.16% at high effort. Anthropic and independent outlets both describe this as several times higher than any previously reported score, though exact multiples cited range from three to four times the prior leader.

Agentic coding takes center stage

Anthropic built Opus 5 specifically around coding, cognitive reasoning, and long-running autonomous agents. On Frontier-Bench v0.1, a test of agentic terminal coding, Opus 5 scored 43.3% at maximum effort, ahead of Fable 5 (33.7%), GPT-5.6 Sol (34.4%), and more than double Opus 4.8's 21.1%, all while costing less per task.

On Artificial Analysis's GDPval-AA v2 evaluation, Opus 5 posted an Elo of 1,861, well clear of Fable 5's 1,747 and GPT-5.6 Sol's 1,736. On Zapier's AutomationBench, Anthropic reports Opus 5's pass rate running about 1.5 times the next-best model at equivalent cost per task.

Science and reasoning gains

The model also picked up ground in domain-specific reasoning. Anthropic reports a 10.2 percentage point improvement over Opus 4.8 on organic chemistry tasks involving spectroscopy-based structure inference, and a 7.7 point gain on protein function prediction from sequence variation.

In an unusual test, Anthropic ran Opus 5 against all six 2026 International Mathematical Olympiad problems with no tools or agent scaffolding. A three-model judge panel scored all 24 generated solutions as correct.

Claude Opus 5 keeps the 1 million-token context window from prior Opus releases and is now the default model on Claude Max and the top-tier option on Claude Pro. It is also live on AWS through the Claude Platform, giving enterprise customers a direct deployment path alongside Anthropic's own consumer and API access.