Anthropic shipped Claude Opus 4.8 on May 28, 2026, the newest revision of its flagship publicly available model and its second Opus upgrade in under two months. The release holds the same price as Opus 4.7 while adding a faster low-cost processing tier, a parallel-agent feature for large coding jobs, and what the company frames as a measurable improvement in the model’s willingness to admit when it is uncertain.
The compressed timeline is part of the story. Opus 4.7 is barely settled into developer workflows, and the quick turnaround signals an upgrade rhythm built around incremental refinement rather than the longer gaps that defined earlier model generations. Anthropic is candid about the scale of the jump, describing Opus 4.8 in its announcement as a “modest but tangible improvement” on its predecessor — unusually plain language for a frontier launch, and a tone the company carries through the rest of the release.
What changed in Opus 4.8
The headline gains sit in coding, agentic reliability, and reasoning. Anthropic reports improvements across its internal benchmark suite, and launch coverage put the model at a reported 88.6% on SWE-bench Verified, a standard test of real-world software fixes. Independent testers cited cleaner tool use and fewer wasted steps, with one partner reporting an 84% score on Online-Mind2Web, a benchmark for browser-based agent tasks.
The practical read is consistency rather than raw capability. Several enterprise testers pointed to the same trait: Opus 4.8 carries tasks end to end with fewer dropped threads, asks sharper clarifying questions inside Claude Code, and flags problems in its own inputs and outputs that earlier models left for users to catch. Anthropic also notes that Opus 4.8 fixes the comment-verbosity and tool-calling quirks that surfaced in Opus 4.7, which suggests this release is as much about cleanup as expansion.
Trained to flag its own uncertainty
The change Anthropic emphasizes most is honesty. The company trains its models to avoid making claims they cannot support, but acknowledges a persistent failure mode across the field: models that jump to conclusions and confidently report progress on thin evidence. Anthropic says Opus 4.8 is roughly four times less likely than Opus 4.7 to let flaws in its own code pass without comment.
That framing matters for the agentic work Anthropic is positioning the model toward. A coding agent running unattended for hours is only useful if it surfaces its own mistakes, and a model that overstates progress is a liability at scale. Anthropic’s alignment team reported that Opus 4.8 reaches higher marks on prosocial traits such as supporting user autonomy, and that its rates of misaligned behavior — including deception or cooperation with misuse — are substantially lower than Opus 4.7 and close to those of Claude Mythos Preview, the company’s more capable restricted model. The full assessment is documented in the Opus 4.8 system card.
New features shipping alongside the model
Opus 4.8 launches with three updates that extend how the model is used rather than what it knows.
Available in research preview, dynamic workflows let Claude plan a large task, run hundreds of parallel subagents in a single session, and verify its outputs before reporting back. Anthropic says the combination allows Claude Code to handle codebase-scale migrations across hundreds of thousands of lines from kickoff to merge, using a project’s existing test suite as the pass bar. The feature is open to Claude Code users on Enterprise, Team, and Max plans.
Effort control and a Messages API change
A new effort control on claude.ai and in Cowork lets users decide how hard Claude works on a response. Higher settings prompt deeper and more frequent reasoning; lower settings return answers faster and consume rate limits more slowly. The control is available on every plan. On the developer side, the Messages API now accepts system entries inside the messages array, so developers can update Claude’s instructions mid-task — adjusting permissions, token budgets, or environment context as an agent runs — without breaking the prompt cache.
Pricing and a cheaper fast mode
Regular usage is priced identically to Opus 4.7, at $5 per million input tokens and $25 per million output tokens. The notable economic shift is fast mode, which runs the model at 2.5 times the usual speed for $10 per million input tokens and $50 per million output tokens. Anthropic says that fast-mode rate is three times cheaper than it was on previous models, a change aimed at high-volume agentic workloads where latency and per-token cost compound quickly. Opus 4.8 defaults to high effort, with “extra” and “max” tiers available for harder or long-running tasks.
What Anthropic signaled about Mythos
The release doubles as a status note on what comes next. Anthropic reiterated that a small group of organizations is already using Claude Mythos Preview for cybersecurity work under Project Glasswing, and that models at that capability level need stronger cyber safeguards before a wider release. The company said it expects to bring Mythos-class models to all customers in the coming weeks. Opus 4.8 is available everywhere now, with the model identifier claude-opus-4-8 on the Claude API and deployment across Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
