Claude Opus 4.8 is excellent for coding agents, but expensive to live in

Claude Opus 4.8: better judgment is the headline

Anthropic released Claude Opus 4.8 on May 28, 2026, positioning it as an incremental upgrade over Opus 4.7 rather than a new model class. That framing matters. Anthropic itself calls it a “modest but tangible improvement,” which is more useful than pretending every release resets the field. (anthropic.com)

The most interesting claim is not raw benchmark movement. It is honesty. Anthropic says Opus 4.8 is around four times less likely than Opus 4.7 to let flaws in code it has written pass without comment. Early testers also reported that the model is more likely to flag uncertainty and less likely to make unsupported claims. For coding-agent work, that is a meaningful kind of improvement: the model does not just need to write code; it needs to notice when its own plan is weak. (anthropic.com)

That matches my own impression. Opus 4.8 is very good. For serious coding-agent work, it feels like the model you reach for when you need judgment, patience, and architectural sense. Compared with GPT-5.5, the best way I can put it is: Opus 4.8 feels like the older, wiser brother. It is more reflective, more careful, and often better at pushing back. But that wiser brother is only around for a couple of hours a day, while GPT-5.5 is available for longer and with fewer practical limits.

The product update around the model also matters. Anthropic added Dynamic Workflows in Claude Code, a research-preview feature for breaking larger engineering tasks into parallel subagents. Anthropic says Claude Code with Opus 4.8 can handle codebase-scale migrations across hundreds of thousands of lines, using the existing test suite as the bar. That is exactly the sort of work where Opus’ stronger judgment could be valuable — not autocomplete, but multi-step software maintenance. (anthropic.com)

The complication is cost. Regular Opus 4.8 pricing is unchanged from Opus 4.7 at $5 per million input tokens and $25 per million output tokens. Fast mode is $10 per million input tokens and $50 per million output tokens. That is workable for high-value coding tasks, but not cheap enough to leave running casually. Even outside the API, the same practical issue shows up as usage limits: Opus can feel wonderful until you hit the wall. (anthropic.com)

So the verdict is: worth trying for serious coding-agent work, especially if you already use Claude Code or need help with complex refactors, migrations, debugging, or architecture. I would not treat it as a cheap daily-driver model for every task. Use it where better judgment saves time, and keep a lower-cost or less-limited model nearby for routine work.

What still needs independent verification is the size of the real-world improvement. Anthropic’s honesty and coding claims are promising, and some external coverage has noted the unusual emphasis on uncertainty handling. But we still need more independent evals across live repositories, long-running agent tasks, and failure modes like package hallucinations, over-editing, and test-suite gaming. For now, Opus 4.8 looks less like a revolution and more like a very capable specialist that you should spend carefully.