Introducing Claude Opus 4.8 (Anthropic announcement)
anthropic‘s own launch post for claude-opus-4-8, 28 May 2026. Pulled by the hub research pass to put a first-party record under a launch this wiki had only through secondary coverage (claude-opus-4-8-review, claude-opus-4-8-zvi, claude-opus-4-8-launch-tomsguide). It is a vendor announcement, so T3 — the tier is the point: for a product fact like a price or a shipped feature the vendor is the authority, and the trade-press pages restating it are not.
What Anthropic states
Pricing is $5 / $25 per million input/output tokens, with fast mode at $10 / $50 — the same figures the corpus already carried. Fast mode runs at 2.5× the speed and is described as three times cheaper than on previous models. Prompt caching is offered at up to 90% savings and batch processing at 50%.
Two features are new here rather than in the reviews: effort control, letting the user choose how much the model thinks on a task, and dynamic workflows in Claude Code, where “Claude can plan the work and then run hundreds of parallel subagents in a single session.”
The benchmark numbers are the part no secondary source in this wiki carried: Terminal-Bench 2.1 92.3%, OSWorld-Verified 84.9%, Online-Mind2Web 84%, and the claim of being the first model to break 10% overall on the Legal Agent Benchmark’s all-pass standard.
The 4× claim is not what the corpus recorded
This wiki’s Opus 4.8 pages report, from the reviews, that the model is “~4× less likely to make unsupported claims” — an honesty result. Anthropic’s own wording is narrower: around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked. That is a code-review property, not a general truthfulness one.
Both claims are kept. The secondary sources may be summarizing a different measure in the system card, or may have generalized a specific one; nothing here settles it. The discrepancy matters beyond this page because claude-opus-4-8 and claude-opus-4-8-launch-tomsguide both lead with the honesty framing as the launch’s flagship result, and the vendor’s own text does not support that framing in the broad form. Flagged in synthesis.
Related
claude-opus-4-8 · anthropic · claude-opus-4-8-review · claude-opus-4-8-zvi · claude-opus-4-8-launch-tomsguide · claude-fable-5