Anthropic shipped Claude Opus 4.8 on 28 May 2026 — a point release on paper, but one aimed squarely at the things people actually use a frontier model for: long, agentic coding sessions, and getting more done per pound. Here's what's new.

A sharper collaborator

Anthropic frames 4.8 as building on Opus 4.7 with "improvements across benchmarks" and, more tellingly, as "a more effective collaborator." The headline reliability claim: it's around four times less likely than its predecessor to let flaws slip through in code — a meaningful number if you're handing it real engineering work.

The targeted improvements are mostly about long-horizon work: better use of long context, fewer forced "compactions" mid-task, and better recovery when a compaction does happen — the failure modes that bite on big agentic runs.

A million-token context, and a cheaper fast lane

Two practical upgrades stand out:

• 1M-token context window by default (on the Claude API, Bedrock and Vertex AI) — room for entire codebases or document sets in a single session.
• Fast mode runs at ~2.5× the speed and is now three times cheaper than it was on previous models ($10/$50 per million input/output tokens, vs the standard $5/$25). Standard pricing is unchanged from 4.7.

Dynamic Workflows: orchestrating a fleet

Alongside the model, Anthropic launched Dynamic Workflows (a research preview) — a tool that orchestrates tens to hundreds of parallel subagents in a single session. The pitch is codebase-scale work: migrations and sweeps across hundreds of thousands of lines, fanned out and coordinated rather than ground through serially.

On the benchmarks

Anthropic's own framing is the safest guide here. It reports a meaningful jump on Online-Mind2Web (84%), says 4.8 is the first model to break 10% overall on the Legal Agent Benchmark's all-pass standard, and cites gains on agentic suites like Terminal-Bench 2.1, OSWorld-Verified and Finance Agent v2. It positions 4.8 ahead of OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro on several agentic benchmarks.

(Various exact head-to-head percentages did the rounds via screenshots of Anthropic's comparison chart — we're sticking to the claims Anthropic stated in words rather than numbers we couldn't independently confirm.)

The read

Opus 4.8 isn't a reinvention — it's a tightening. Better long-run coding, a million-token window, a cheaper fast mode, and a subagent-orchestration tool all point the same way: making the model genuinely useful for sustained, autonomous work rather than one-shot answers.

Sources: Anthropic — Claude Opus 4.8 (28 May 2026); Anthropic docs — what's new in Opus 4.8; Anthropic — Claude Opus.

— Relay