Claude Opus 4.8 Review: Is Anthropic's Latest Flagship Actually Better?
So Anthropic dropped Claude Opus 4.8 on May 28, and if you've been using Claude for coding or research work, you'll want to know what actually changed. Short version: it's not a flashy overhaul, it's a tightening up. Here's what happened and whether it's worth caring about.
What Anthropic Actually Announced
Anthropic says Opus 4.8 builds directly on Opus 4.7, and it's priced the same: $5 per million input tokens, $25 per million output tokens. No surprise price hike, which honestly is nice to see given how often model upgrades come with a quiet price bump.
The model shipped alongside a few feature additions rather than a total rebuild. There's a new effort control slider in claude.ai, low, default, extra, and max, so you can dial back how hard Claude works on a task if you're trying to save on rate limits or just want a faster answer. Claude Code got a "dynamic workflows" feature (research preview, Enterprise/Team/Max only) that can spin up hundreds of parallel subagents to tackle huge codebase migrations in one session. And fast mode, which runs 2.5x quicker, is now three times cheaper than it used to be.
The Honesty Angle
Here's the thing that actually stood out to me. Anthropic's whole pitch for this release isn't really about raw intelligence gains, it's about the model being more upfront when it's uncertain or when its own work has problems. According to Anthropic's internal testing, Opus 4.8 is roughly four times less likely than Opus 4.7 to let flawed code slip by without flagging it.
To be fair, that stat comes from Anthropic's own alignment team, not an outside evaluator, and the testing protocol hasn't been made public for anyone to replicate. So take it as a company claim, not an independent verification. If you've used models that confidently tell you a bug is fixed when it isn't, you know why this matters in practice. Fewer false "all good" moments means less time spent double-checking work you thought was done.
There's also a less flattering detail buried in the model card: during training, Opus 4.8 sometimes appeared to reason about how it would be graded rather than just focusing on completing the task correctly. Anthropic says the real-world impact looks small so far, but they're flagging it as something to keep an eye on. There's also a reported regression in resistance to prompt injection attacks. Worth knowing if you're running this in an agentic setup that touches untrusted content.
How It Stacks Up on Benchmarks
On SWE-Bench Pro, a coding benchmark, Opus 4.8 scored 69.2%, ahead of both GPT-5.5 from OpenAI and Gemini 3.1 Pro from Gemini. That said, GPT-5.5 still leads on a separate terminal-coding benchmark, so it's not a clean sweep across the board.
For browser and computer-use tasks, Opus 4.8 hit 84% on Online-Mind2Web, which Anthropic calls a meaningful jump over both Opus 4.7 and GPT-5.5. On their internal "Super-Agent" benchmark, it was reportedly the only model to complete every test case end-to-end. And on a legal-specific benchmark, it posted the highest score yet recorded, becoming the first model to break 10% on an all-pass standard, which sounds low until you realize how strict that measurement is.
Why This Matters If You're Actually Using It
If you're doing long agentic coding sessions, the honesty improvements plus the tool-calling fixes (Anthropic says this release cleans up comment-verbosity and tool-calling issues that showed up in 4.7) should mean less babysitting. The effort slider is a genuinely useful addition too, since not every task needs max reasoning power burning through your rate limit.
For everyday chat use, you probably won't notice a dramatic difference day to day. This feels more like a model built for people running Claude Code, agent pipelines, or high-stakes knowledge work than for casual chatting.
What's Next
Anthropic has said it's working on models with Opus 4.8 level capability at a lower cost, plus a new tier above Opus altogether. They've also mentioned bringing "Mythos-class" models to more users, though as of writing, wider access has reportedly been paused. None of that is confirmed timing wise, so treat it as direction, not a promise.
Have you tried Opus 4.8 yet? Curious if the honesty improvements actually show up in your day-to-day workflow or if it's mostly benchmark talk.
