claude-opus-4-8 · 1M-token context · $5 / $25 per million tokens standard, $10 / $50 in fast mode (2.5× speed) · Available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot.
Opus 4.8 is a focused refinement, not a generational leap. It’s faster, more careful, and notably more honest about its own limits — roughly 4× less likely than Opus 4.7 to let flaws in its own code slip through unflagged — and it ships at the same standard price. If you were on 4.7, the upgrade is essentially free and worth switching to today.
Anthropic has released Claude Opus 4.8, the newest version of its most capable publicly available model. It arrived on May 28, 2026 — just 41 days after Opus 4.7 — an unusually fast turnaround that signals how quickly the frontier-model race is moving right now. For context, Anthropic’s most recent Sonnet and Haiku models were then three and seven months old, respectively, so a six-week Opus refresh counts as a sprint. So what actually changed in this update? Here’s a clear, practical breakdown.
A sharper, more reliable collaborator
Anthropic frames Opus 4.8 as a “modest but tangible” improvement over 4.7, and that honesty is refreshing. This isn’t a generational leap — it’s a focused refinement. The model posts stronger results across coding, agentic tasks, reasoning, and professional knowledge work, but the standout upgrade is something subtler: trust.
Opus 4.8 is much better at knowing what it doesn’t know. Earlier models had a tendency to declare victory on thin evidence, claiming a task was finished when it wasn’t quite there. The new version raises its hand when it’s uncertain and, by Anthropic’s own measurement, is roughly four times less likely than its predecessor to let flaws in its own code slip through unmentioned — the company describes it as the first Claude model to effectively stop uncritically reporting flawed results, with a more than ten-fold drop in overconfidence. For anyone running long, unattended agent workflows, that kind of reliability compounds fast: it shifts human reviewers from line-by-line checking toward higher-level architectural review.
Where the gains show up
Selected results Anthropic highlighted at launch. Benchmarks near saturation move only slightly at the top.
Beyond those, Anthropic reported that Opus 4.8 beats prior Opus models on CursorBench across every effort level, uses fewer tool-calling steps for the same result, set a record on its internal Legal Agent Benchmark (the first model to break 10% overall on the strict all-pass standard), and ranks as its strongest computer-use and browser-agent model to date.
New features shipping alongside the model
Opus 4.8 didn’t arrive on its own. Three additions are worth knowing about.
Decide how hard Claude works
On Claude.ai and Cowork, a new control next to the model selector lets you choose how much effort Claude puts into a response. Higher effort means deeper, more frequent thinking and better answers; lower effort means faster replies that use up your limits more slowly. Opus 4.8 defaults to “high,” with “extra” (“xhigh” in Claude Code) and “max” available for harder jobs. It works on every plan — the same reasoning dial that used to be an API-only setting is now visible to everyone.
Codebase-scale work in Claude Code
Claude can plan a large task, launch hundreds of parallel subagents in a single session (capped at 1,000), verify the results, then report back. Anthropic says it can now carry out codebase-scale migrations across hundreds of thousands of lines of code — from kickoff all the way to merge, using the existing test suite as the bar. You trigger it by including the word “workflow” in a prompt or enabling “ultracode.” It’s available on Claude Code Enterprise, Team, and Max plans; token cost scales with the number of workers, so it’s best reserved for problems that genuinely need parallelism.
Update instructions mid-task
Developers can now place system instructions directly inside the messages array, updating Claude’s permissions, token budgets, or context mid-task without breaking the prompt cache or faking a user turn — handy for hours-long agent runs that used to break the moment context shifted.
Faster, at the same price
Pricing for standard usage hasn’t moved. Fast mode — which runs about 2.5× quicker — now costs three times less than it did on previous models. Here’s the full rate card:
| Mode | Input / output (per million tokens) | Speed | Best for |
|---|---|---|---|
| Standard | $5 / $25 | Normal | Most production workloads |
| Fast mode | $10 / $50 | ~2.5× faster | Interactive tooling, IDE completions, latency-sensitive chat |
The model is available everywhere you’d expect it: the Claude API (as claude-opus-4-8), Amazon Bedrock, Google Cloud’s Vertex AI, Microsoft Foundry, and GitHub Copilot. One nuance worth internalizing: with Opus 4.8, “faster” and “better” are now separate switches — fast mode buys latency, effort control buys reasoning.
Safer under the hood
Anthropic ran its standard pre-release alignment assessment, and the findings are encouraging. The team reported new highs on prosocial measures such as respecting user autonomy and acting in the user’s best interest, along with meaningfully lower rates of deceptive or misuse-friendly behavior than Opus 4.7 — placing it close to the company’s best-aligned model to date, on par with the Mythos Preview.
The Opus 4.8 system card notes that agentic prompt-injection robustness is slightly weaker than Opus 4.7, with red-teaming showing a higher attack-success rate (~9.6% vs 6.0%). Teams running Opus 4.8 in agent pipelines that touch untrusted input should review their sandboxing before scaling up. There’s also a tiny regression on the near-saturated GPQA Diamond benchmark (93.6% vs 94.2%), which is within expected top-end variance.
What’s coming next
Anthropic pointed to two directions ahead. First, cheaper models that deliver much of Opus’s capability at a lower price point. Second, a more powerful Mythos class of model — currently limited to a small group through Project Glasswing for cybersecurity work — which the company expects to open up more widely “in the coming weeks,” once additional safeguards are in place.
The bottom line
Claude Opus 4.8 won’t reinvent your workflow overnight — and it doesn’t need to. It’s faster, more careful, and more honest about its own limits, which are exactly the qualities that matter when you’re handing a model real, long-running work. If you were already relying on Opus 4.7, the upgrade is essentially free and well worth switching to today.
Frequently asked questions
When was Claude Opus 4.8 released?
May 28, 2026 — just 41 days after Opus 4.7, an unusually fast cycle for Anthropic.
How much does Opus 4.8 cost?
Standard usage is $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.7. Fast mode runs about 2.5× faster at $10 / $50 per million, now three times cheaper than fast mode on previous models.
What is the “4× fewer flaws” claim?
Anthropic’s internal evaluation found Opus 4.8 is roughly four times less likely than Opus 4.7 to let defects in its own generated code pass without comment. It’s an internal claim, so independent studies are still worth watching.
What is effort control?
A per-response dial in Claude.ai and Cowork that sets how much thinking Claude does. It defaults to “high,” with “extra” (xhigh) and “max” for harder tasks. Higher effort improves quality; lower effort is faster and lighter on your limits. Available on every plan.
What are dynamic workflows, and who gets them?
A research-preview feature in Claude Code that lets Claude plan a big task, run hundreds of parallel subagents (up to 1,000), verify outputs, and report back — enough to handle repo-wide migrations. It’s on Enterprise, Team, and Max plans; Pro gets effort control but not dynamic workflows.
Should I upgrade from Opus 4.7?
For most coding and agent workloads, yes. It’s the same standard price with better reliability and judgment. Just review sandboxing if you run agents on untrusted input, given the small prompt-injection regression.
Stay on top of every Claude release
Model updates, pricing shifts, and new features are landing fast. Bookmark AISetApp for clear, practical breakdowns as they ship.
Explore more on AISetApp →