Claude Opus 4.8 has arrived: Anthropic bets everything on model honesty
In the past few hours, Anthropic has released Claude Opus 4.8, the update of its flagship model, available immediately on claude.ai, API, Claude Code, and Cowork with the string claude-opus-4-8. The price remains the same as the predecessor: $5 per million tokens for input and $25 for output. The model arrives a little less than six weeks after Opus 4.7, the predecessor which in April had raised the bar on coding and high-resolution images: an unusually close release cadence for the company. Regarding the quality of the new model, Anthropic emphasizes more than anything else: honesty.
The term should be understood in an operational sense. Opus 4.8 signals its uncertainties more often and makes fewer unsupported claims about the work it is doing. On the company's internal evaluations, it is about four times less likely than its predecessor to let defects in the code it has written pass without comment, and Anthropic presents this as the data that best summarizes the leap from Opus 4.7. This is a concrete problem for those working with these models: a helper tends to declare itself confident that it has completed a task even when the evidence supporting it is weak, and a model that acknowledges its own limits reduces downstream verification work.
The Numbers, with One Exception
On declared tests, Opus 4.8 improves almost across the board. On SWE-Bench Pro, an agent-based coding test, it rises to 69.2% from 64.3% of Opus 4.7, ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). Anthropic reports progress also on Humanity's Last Exam (49.8%), on Finance Agent v2 (53.9%), and on computer use in OSWorld-Verified (83.4%). However, the lead is not absolute: on Terminal-Bench 2.1, which measures terminal coding, GPT-5.5 remains ahead with 78.2% against 74.6% of Opus 4.8. On the cost front, the fast mode runs at 2.5 times the standard and costs three times less than the previous Opus models, at a rate of $10 for input and $50 for output per million tokens.
Dynamic Workflows and Commitment Control
Along with the model, three product innovations arrive. The first, Dynamic Workflows in research preview, extends Claude Code to large-scale tasks: the agent plans the work, launches hundreds of subagents in parallel in the same session, and verifies the results before reporting them. According to Anthropic, in this configuration Claude can complete migrations on codebases of hundreds of thousands of lines, from initial planning to merge, using the existing test suite as a benchmark. This feature is reserved for the Enterprise, Team, and Max plans of Claude Code.
The second is commitment control, or effort, now available in claude.ai and in Cowork via a selector next to the model's one. It ranges from a low level, which responds faster and consumes usage limits more slowly, to the maximum; the default value is high, which Anthropic indicates as the best compromise between quality and user experience. The third concerns developers: the Messages API now accepts system voices within the message array, allowing updates to the instructions mid-task without breaking the prompt's cache.
A Note in the System Card
On the theme of honesty, the system card contains a note that Anthropic itself indicates as the most sensitive issue that emerged from training. Opus 4.8 shows an increasing tendency to reason explicitly about how its responses will be evaluated, even in contexts where it is not told that it is under review: it seems to steer output toward what would get a good grade, rather than what it would produce in the absence of evaluation. Preliminary interpretability work has identified reasoning of this type, not verbalized, in about 5% of training episodes. The company clarifies that this phenomenon has not resulted in observable worse behavior; indeed, the model makes fewer misleading claims about successful task completion, but describes it as a trend that could complicate training in the future.
Opus 4.8 is available from today on all Claude products, direct APIs, and cloud platforms, at the same price as Opus 4.7. Anthropic also states that it is close to making Mythos class models, previously reserved for a few customers, available to all its users in the coming weeks, once cyber safeguards are completed.