Anthropic releases Claude Opus 4.8 with sharper judgement and dynamic workflows

Anthropic shipped Claude Opus 4.8 on May 28, 2026 — an upgrade to Opus 4.7 across coding, agent work, reasoning, and knowledge work. The release notes frame the model as having sharper judgement, more honesty about its progress, and the ability to work independently for longer; Anthropic reports a more-than-10x reduction in overconfidence versus 4.7, and says it is the first Claude model to score 0% on uncritically reporting flawed results — a calibration posture that maps to the trust-engineering direction the frontier labs have signaled for a year.

Two things shipped alongside read as bigger than the version bump. Claude Code gained Dynamic Workflows — a research-preview mode where Claude writes a script that spawns parallel subagents (capped at 16 concurrent and 1,000 per run), verifies their output, and returns only the final answer — moving the agentic-operator pattern from “build it yourself with the SDK” to a primitive in the IDE. A new effort control lets callers dial how aggressively the model thinks before responding, across claude.ai, the API, and Cowork, and the context window reaches 1M tokens in beta on Vertex AI and Bedrock.

Pricing is unchanged at $5 and $25 per million input and output tokens, with an optional fast mode roughly 2.5x quicker at $10 / $50. The forward signal is the part to watch: Anthropic flagged “Mythos-class” models, still in safety testing, for subsequent weeks. Opus 4.8 is the third Opus point-release since early February — 4.6 on February 5, 4.7 on April 16, 4.8 on May 28 — and the gap between them is narrowing. As with the rest of the buildout this site tracks, the cadence itself is the harder signal: a frontier lab shipping a materially better flagship roughly every six weeks is its own kind of capex.