Anthropic launches Claude Opus 4.8 with effort controls

TL;DR:

• Anthropic has released Claude Opus 4.8, a benchmark upgrade over Opus 4.7 offered at the same price ($5/$25 per million input/output tokens).
• The launch ships with user-facing “effort” controls, a “dynamic workflows” feature for Claude Code that runs hundreds of parallel subagents, and faster, cheaper fast-mode inference.
• Anthropic says a more powerful Mythos-class model is being held back behind cybersecurity safeguards, with general release expected “in the coming weeks”.

Anthropic has upgraded its flagship to Claude Opus 4.8, framing it as a “modest but tangible” step beyond Opus 4.7 rather than a generational leap. The pricing is unchanged, and the company is leaning on reliability and judgement gains — early testers report the model is better at pushing back on unsound plans and flagging its own uncertainty — rather than a single headline benchmark.

What is actually new

Three changes matter more than the benchmark table. First, an effort control now sits alongside the model selector on claude.ai and in Claude Code, letting users trade speed against depth of reasoning. Second, “dynamic workflows” — in research preview for Enterprise, Team and Max plans — lets Claude Code plan a task, run hundreds of parallel subagents and verify outputs before reporting back, which Anthropic says enables codebase-scale migrations across hundreds of thousands of lines. Third, fast mode for Opus 4.8 runs at 2.5× speed and is now three times cheaper than on previous models.

Anthropic also says Opus 4.8 is roughly four times less likely than its predecessor to let coding flaws pass unremarked, and that its alignment team recorded lower rates of misaligned behaviour. Those claims are the company’s own and have not been independently verified.

Looking forward

The more strategically interesting line is about what is not shipping. Anthropic says its Mythos-class model — higher in capability than Opus and already used by a small number of organisations for cybersecurity work under “Project Glasswing” — needs stronger cyber safeguards before wider release. That deliberate hold-back, expected to lift in the coming weeks, is the same episode that reportedly eased a stand-off with the White House. For UK firms standardising on Claude Code, the immediate takeaway is incremental: better agentic reliability at the same price, with a more capable tier visible on the horizon.