Opus 4.8 Released. Cheaper, Calmer, and More Honest.
Ronni Holmvig Strøm · 2026-05-28
Anthropic released Claude Opus 4.8 today, and the company is admirably blunt about what it is: a modest but tangible improvement on Opus 4.7.
Anthropic released Claude Opus 4.8 today, and the company is admirably blunt about what it is: a modest but tangible improvement on Opus 4.7. There's no capability moonshot here, and the price holds steady at $5 per million input tokens and $25 per million output, the same as the model it replaces.
That restraint is the reason to pay attention.
The meaningful gains in 4.8 live below the benchmark line. They're the ones that make the model safe to leave running on its own, and for the people building production agentic systems, that reliability has been the bottleneck for years.
The headline feature is honesty
The single most consequential change in Opus 4.8 is one Anthropic puts front and center: the model is more honest about its own work. It is roughly four times less likely than Opus 4.7 to let a flaw in code it wrote slip by unremarked. It flags its own uncertainty more readily, and it makes fewer confident claims it can't support. Every practitioner knows the failure mode this targets: the model that announces "Done!" over a pile of broken output.
If you've ever left an agent running for an hour and come back to a tidy summary describing work that never happened, you understand why this beats another two points on a coding eval. A model that knows the limits of its own work and says so is what separates an agent you can deploy from an agent you have to babysit. Anthropic's alignment team puts it in their own terms, reporting that 4.8 reaches new highs on prosocial traits like supporting user autonomy and acting in the user's best interest, with measured rates of misaligned behavior well below 4.7 and close to the company's best-aligned system, the Mythos preview.
Reliability is the unglamorous bottleneck of the whole agentic era, and 4.8 goes straight at it.
The economics quietly got better, again
Here's the trend line that should make TheoryLab readers optimistic regardless of how they feel about any single release: the frontier keeps getting cheaper to run, even as it gets more capable.
Fast mode, where Opus runs at 2.5x speed, is now three times cheaper than it was on previous models. Anthropic also shipped an effort control today, available on every plan, that gives users a direct dial between maximum reasoning and maximum speed. Turn it up and Claude thinks harder and longer for a better answer. Turn it down and it responds faster while using up rate limits more slowly. That's a useful admission that most queries don't need the model's full reasoning budget, and that users should pay for depth only when depth earns its keep.
The partner numbers back it up. Databricks reports that 4.8 reasons over PDFs, diagrams, and other unstructured content at 61% lower token cost than 4.7 inside its Genie agent. That's a line item, and it changes whether a multimodal workflow pencils out at scale.