Anthropic has shipped Claude Opus 4.8, and the easiest version of the story is that the model is better at coding and professional work than Opus 4.7. That is true, at least by Anthropic’s own benchmark table. It is also the less useful part of the launch.
The useful part is that Anthropic is treating high-end agent work like something that needs a throttle. Opus 4.8 arrives with explicit effort controls and cheaper fast mode. It also adds mid-conversation system messages plus Claude Code workflows that can fan out into many background agents.
That is not just a model release. It is a runtime-management release wearing a benchmark jacket.
Why it matters: The agent market is moving from “can the model do it?” to “can the system finish the work without burning a budget hole through the floor?” Opus 4.8 is built around that second question.
The model bump is the setup
Anthropic says Opus 4.8 is available now at the same standard price as Opus 4.7: $5 per million input tokens and $25 per million output tokens. Fast mode is priced at $10 per million input tokens and $50 per million output tokens, with Anthropic claiming it runs 2.5 times faster and is three times cheaper than fast mode on previous models.
Those numbers matter more than the headline benchmark deltas. Coding agents do not fail only because the model misses a clever refactor. They fail because long sessions get expensive. Instructions change midstream, while orchestration can turn one prompt into a swarm of tool calls.
The API release notes show how Anthropic is responding. Opus 4.8 supports the claude-opus-4-8 model ID, a 1 million token context window by default, and a 128,000 token max output. It also supports mid-conversation system messages, which lets developers change system instructions during a long session while preserving prompt cache hits.
