Anthropic Prices Claude Opus 5 to Undercut Its Own Top Model

Opus 5 claims near-flagship reasoning at half the token rate of Fable 5, holding the same price as its predecessor.

Anthropic released Claude Opus 5, a new flagship it positions as a cheaper alternative to Claude Fable 5 at half the token rates. Opus 5 costs $5 per million input tokens and $25 per million output tokens, matching the earlier Opus 4.8, and carries a 1 million-token context window. A new Fast Mode raises speed 2.5x but doubles the price. It is now the default on Claude Max and the most capable model on Claude Pro.

Anthropic's own benchmarks show a mixed but strong picture. On ARC-AGI-3 Opus 5 scored 30.2 percent, against 1.5 percent for Opus 4.8 and 7.8 percent for GPT-5.6 Sol. On the Frontier-Bench agentic coding test it hit 43.3 percent, ahead of Fable 5's 33.7 percent. On the GDPval knowledge-work benchmark it led at 1,861 Elo. But on DeepSWE, GPT-5.6 Sol led at 72.7 percent, with Opus 5 third. Cyber classifiers reportedly trigger about 85 percent less often than Fable 5's.

The throughline for abundance is cost. Frontier-grade reasoning offered at half a competitor's token rate keeps pushing the price of machine cognition down, which makes capable AI reach more people and more tasks.

The caveats are real. These are Anthropic's own numbers, there is no Fable 5 score on ARC-AGI-3, and token rates alone don't capture total cost, since token efficiency varies between models. It's unclear whether large benchmark leads translate to real-world work.

Source: The Decoder