Anthropic released Claude Opus 4.8 today, May 28 — in the company's words, "an upgrade to our Opus class of models, with stronger performance across coding, agentic tasks, and professional work, and the consistency to handle long-running work." Pricing lands at $5 per million input tokens and $25 per million output, with a fast mode at $10 and $50. Anthropic says it's "available everywhere today."

The launch claims cluster around one theme: agents that don't quit partway. Anthropic's headline: "On our Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost." The page also calls it "the strongest computer-use and browser-agent model we've tested, scoring 84% on Online-Mind2Web," and says it "delivers the highest score recorded on our Legal Agent Benchmark, and is the first model to break 10% overall on the all-pass standard." The stated focus is "reliability for agent products in translation, deep research, slide-building, and analysis."

Two smaller notes from the page. The model ships with effort-setting controls and dynamic workflows for larger tasks, and defaults to high-effort responses. And Anthropic names its own recent misses in print: the release "improves on Opus 4.6 and fixes the comment-verbosity and tool-calling issues we saw with Opus 4.7." Vendors almost never list their own regressions on a launch page. When one does, the rest of the page has earned a slightly longer look.

Finished work is checked work

The claim worth the most attention isn't a leaderboard number. It's this one: Anthropic says Opus 4.8 is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked."

Read that like a hiring manager, not an engineer. When you hand someone a multi-step task — reconcile these statements, assemble this filing, build this deck — you aren't paying for raw intelligence. You're paying for a person who finishes, and who flags their own mistakes before the work reaches a customer. Delegation runs on those two properties. The failure mode that kills small-business automations isn't a dumb model; it's a confident answer with a silent flaw that nobody notices until a quote, an invoice, or a filing has already gone out.

Put Anthropic's two central claims side by side and you can see the product they're actually selling. "Only model to complete every case end-to-end" is about finishing. "Four times less likely to let flaws pass unremarked" is about noticing. An agent with both is the difference between a tool you babysit and a task you hand off. The industry has been moving along this axis for several releases now — long-horizon reliability over peak cleverness — and point releases are where the shift shows up first, an argument we made when Opus 4.7 shipped. Opus 4.8 is that argument compressed into one launch page.

The usual caveat applies double here: every number above is Anthropic grading Anthropic, on evaluations Anthropic built — Super-Agent and the Legal Agent Benchmark are the company's own. Day-one vendor claims are directionally useful and independently unverified. What you can verify is behavior on your own work, which is what the list below is for.

What to do with this

  • Re-score the processes you shelved for reliability reasons. If a multi-step workflow got parked because agents stalled midway or needed constant supervision, this is precisely the property the vendors now claim to have fixed. Re-test one candidate with the 30-day playbook — a bounded pilot, not a rollout.
  • Make self-review part of the agent's job description. Whatever model you run, add an explicit step: the agent audits its own output against a short checklist before handoff, and logs what it flags. Anthropic building this into the model signals where products are headed; one extra prompt step approximates it today.
  • Track two numbers, not demo quality. Completion rate — how often the agent finishes without a human rescue — and caught-flaw rate: of the errors found in its output, how many did the agent catch itself? Those two metrics answer the question we put to every agentic tool: what happens when step four fails?
  • Match the model to the step, not the whole job. At $5/$25 per million tokens, Opus-class pricing buys the checking and the finishing; cheaper models still make sense for the routine steps in between. Scoping which step needs which grade of model is most of the work in our builds.

If you want an honest read on which of your handoffs an agent could now finish — and check — that's a 30-minute conversation.

Source: Introducing Claude Opus 4.8 — Anthropic, May 28, 2026