Anthropic ships Claude Opus 4.8, with effort dials and parallel subagents
Opus 4.8 lands at the same price as 4.7 with reported gains on coding, agent, and computer-use benchmarks, plus a new effort control and a Claude Code feature that fans out hundreds of subagents. Anthropic also teased a higher-tier "Mythos" class arriving in the coming weeks.
Builders, integrators, prompt engineers · 2 min read
Anthropic upgraded Opus without touching the sticker. Claude Opus 4.8 is available today via the Claude API as claude-opus-4-8, at $5 / $25 per million input/output tokens — identical to Opus 4.7. Fast mode is $10 / $50, which Anthropic says is roughly three times cheaper than fast mode on prior models.
The headline change for builders is the effort control. Opus 4.8 defaults to "high," which Anthropic claims spends a similar number of tokens as Opus 4.7's default while scoring better on coding. Above that sit "extra" ("xhigh" in Claude Code) and "max," which trade more tokens for more quality. Rate limits in Claude Code have been raised to absorb the upper tiers.
A few other API and harness changes worth noting:
- System entries inside the messages array. The Messages API now accepts mid-conversation system updates without breaking the prompt cache or laundering them through a user turn. Anthropic pitches this for updating permissions, token budgets, or environment context inside a running agent — a small but real ergonomic win for long-horizon harnesses.
- Dynamic workflows in Claude Code (research preview, Enterprise/Team/Max). Claude plans the work, fans out "hundreds of parallel subagents" in a single session, then verifies before reporting back. Anthropic's stated target is codebase-scale migrations across hundreds of thousands of lines, using the existing test suite as the bar.
- Honesty / self-checking. Anthropic says Opus 4.8 is "around four times less likely" than 4.7 to let flaws in its own code pass without flagging them. Independent reproduction will matter here, but for agent loops where the model grades itself, even directionally fewer false greens is meaningful.
The benchmark commentary is mostly partner quotes rather than a clean table we can verify, but the numbers Anthropic does cite include 84% on Online-Mind2Web for computer-use / browser agents, and a partner claim of 61% lower token cost than Opus 4.7 on multimodal document reasoning inside Databricks Genie. A footnote also notes Anthropic re-ran OSWorld-Verified and revised Opus 4.7's score to 82.3% — worth keeping in mind when comparing older charts.
The roadmap line to watch: Anthropic says it's working on cheaper models with Opus-like capability, and plans to release a higher-intelligence class above Opus. Claude Mythos Preview is already in the hands of a small group through "Project Glasswing" for cybersecurity work, gated on stronger cyber safeguards, with general availability promised "in the coming weeks." If that ships on schedule, Opus 4.8 may be a short-lived flagship.
