Render

Anthropic ships Opus 4.8 with better agentic judgment — same price, sharper execution.

The upgrade matters less for benchmark scores than for what early testers report: the model catches its own mistakes now.

On May 28, Anthropic released Claude Opus 4.8, upgrading from Opus 4.7 with improvements across coding, agentic skills, and reasoning benchmarks. The price stays the same. Fast mode — where the model runs at 2.5× speed — is now three times cheaper. The release also introduces user-controlled effort levels on claude.ai and dynamic workflows in Claude Code for large-scale problems.

The benchmark table looks like every other model release this year: modest gains across SWE-bench, GPQA, MMLU-Pro. The kind of incremental progress that reads as noise unless you’re the one shipping the model. What matters is buried in the testimonials from early testers, and it’s specific: “it asks the right questions, catches its own mistakes, pushes back when a plan isn’t sound.” That’s not capability theater. That’s the model developing something closer to judgment.

Judgment is the hardest thing to benchmark and the most valuable thing to ship. A model that scores 5% higher on a reasoning test but still confidently suggests a broken implementation is not more useful. A model that pauses before committing to a refactor, asks a clarifying question, or flags its own uncertainty — that model changes how much you can delegate. The tester quote from Cursor is the tell: “tool calling is meaningfully more efficient, using fewer steps for the same intelligence.” Fewer steps means the model isn’t flailing. It knows when it doesn’t know.

The legal benchmark result is the other data point worth isolating. Opus 4.8 is the first model to break 10% overall on the “all-pass standard” for substantive legal work. Ten percent sounds low until you understand what all-pass means: every sub-task correct, no hallucinated citations, no dropped context across a multi-step research chain. That’s the difference between a tool that generates plausible-sounding drafts and a tool that an attorney can actually rely on for real work. The quote from the legal tester is direct: “that’s the kind of accuracy lift that translates directly into how much real attorney work our customers can hand off with confidence.”

The skeptical read is that testimonials are selection bias in paragraph form. Anthropic picked the quotes. The testers who found Opus 4.8 worse than 4.7 didn’t make the press release. That’s fair. But the specificity of the feedback — “pushes back when a plan isn’t sound,” “catches its own mistakes,” “carries end-to-end tasks through” — is hard to fake. These are practitioners describing workflow changes, not vibes. The Super-Agent benchmark claim is independently verifiable: Opus 4.8 is the only model to complete every case end-to-end, beating GPT-5.5 at cost parity. If that’s true, it’s not incremental.

The practitioner instruction is narrow: if you’ve been running Opus 4.7 in production and hitting reliability issues on multi-step agentic tasks, test 4.8 this week. Specifically, test it on the tasks where 4.7 required the most retry logic or human correction. The upgrade is free — same API, same price, same context window. If the judgment improvements are real, your retry rate should drop. If it doesn’t, you’ll know in a day. The other move is to experiment with the new effort control on claude.ai. If you’ve been frustrated by Claude over-explaining or under-explaining, you now have a dial. Use it.

The broader frame is that we’re entering a phase where model releases matter less for raw capability and more for operational reliability. The ceiling on what these models can do is rising slowly. The floor on what they do consistently is rising faster. That’s the shift that matters for practitioners. A model that’s 5% smarter but 20% more predictable is the one you can build a business on.