Anthropic Ships Claude Opus 4.8 with Improved Honesty and Cheaper Fast Mode

Anthropic released Claude Opus 4.8 alongside dynamic workflows that are capped at 1,000 subagents. [2] The company described the update as a modest but tangible improvement over its predecessor. [1] Opus 4.8 maintains the 1,000,000 token context window and January 2026 knowledge cutoff of version 4.7. [1] Standard pricing remains unchanged at $5 per million input tokens and $25 per million output tokens. [1] However, the model introduces a cheaper "fast mode" priced at exactly double the standard rates, which is significantly lower than the $30 and $150 rates for previous versions. [1][2]

The model focuses heavily on honesty, with evaluations showing it is around four times less likely than previous models to leave code flaws unremarked. [1] Claude Opus 4.8 achieved the lowest incorrect-rate among the six models tested across every benchmark. [1] This reduction in factual hallucination was primarily achieved by the model abstaining from answering questions when it was uncertain. [1]