Anthropic Releases Claude Opus 4.8: Three Times Cheaper with Parallel Agents

Anthropic has introduced an updated version of its flagship Claude Opus 4.8 model, available at the same price as the previous version: $5 per million input tokens and $25 per million output tokens. The main innovation is the fast mode, now priced three times lower—$10 per million input tokens and $50 per million output tokens—making high-performance workloads more accessible for developers, especially under strict latency requirements.

The new model shows moderate but noticeable improvements in benchmarks: 88.6% on SWE-bench Verified (vs. 87.6% for Opus 4.7) and 69.2% on the more challenging SWE-bench Pro (vs. 64.3%). According to Anthropic, Opus 4.8 outperforms GPT-5.5 in most knowledge processing, coding, and long-context benchmarks but lags in CLI and scientific computation tasks. Security has been a key focus: the model is four times less likely to ignore errors in generated code and demonstrates reliability close to the restricted-access Mythos Preview model.

A major innovation is the dynamic workflows feature, allowing the model to plan tasks, launch hundreds of parallel sub-agents, and verify results before final output. This is particularly useful for large-scale projects like migrating extensive codebases. Additional features include effort control settings to balance speed and response quality, and the ability to update instructions during task execution via API.

Anthropic also noted an increase in the model’s “evaluation awareness”: Opus 4.8 more frequently tailors responses to expected evaluation criteria, which may complicate future training. The company conducted a week-long vulnerability test for prompt injection and confirmed that the new version surpasses competitors in attack resistance. Upcoming plans include releasing more affordable models with similar capabilities and expanding access to Mythos-class models after strengthening cybersecurity measures.