Anthropic has released Claude Opus 4.8, its newest flagship model, and is keeping the price unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens. The company is pitching the upgrade as more reliable, less prone to confident nonsense, and better at the kind of coding work that now defines the AI model arms race.
The Claude Opus 4.8 update also adds new controls for effort and workflow handling, while Anthropic teases its still-unreleased Mythos-class models. In short, this is a release about better coding performance, steadier behavior, and no extra bill.
Claude Opus 4.8 keeps the same price
Anthropic says Claude Opus 4.8 posts leading results on agentic coding and agentic computer use, which is exactly where these frontier models are being judged now. More interesting is the company’s emphasis on ”honesty”: it says the model is better at spotting its own mistakes, flagging uncertainty, and avoiding unsupported claims.
That pitch lands in a market where trust is becoming a product feature, not a vague virtue signal. If a model can warn users that its analysis is shaky before a finance team or developer ships something dumb, that is a real advantage over rivals that still sound certain while being wrong.
- Price: $5 per million input tokens
- Price: $25 per million output tokens
Anthropic also says the model carries a ”substantially lower” risk of misaligned and dangerous behavior, including harmful sexual content and ”undermining liberal democracy,” according to its system card. Those are not exactly subtle failure modes, which is why every serious AI lab now has to spend as much time talking about safeguards as it does about benchmarks.
Dynamic workflows and effort control arrive together
Alongside the model, Anthropic is introducing ”dynamic workflows” as a research preview. The feature lets Claude break complex coding jobs into hundreds of subagents working in parallel, a sensible move for heavier tasks where brute-force coordination can beat one oversized guesser pretending to be an engineer.
There is also a new ”effort control” panel in the model selector dropdown. Users can set effort to Low, Medium, High, or Max, or switch on adaptive thinking mode. That is a quiet admission that one of the biggest complaints around Opus 4.7 was not raw intelligence, but judgment about how much effort to spend on a task in the first place.
- Effort settings: Low, Medium, High, Max
- Adaptive thinking mode: available as a toggle
Mythos-class models are still under wraps
Anthropic also teased ”Mythos-class models,” which it says are based on a new class of model with capabilities on par with Mythos. The company has not released Mythos publicly, citing unusual power and cybersecurity risks, and says it is still testing safeguards before a wider rollout in the coming weeks.
That kind of language is catnip for the AI industry and a headache for everyone else. The pattern is familiar: the more dramatic the warning, the more hype builds around whatever finally appears. Whether Mythos-class models arrive as a genuine leap or a heavily managed downgrade will decide whether Anthropic just updated its flagship or opened a new chapter in the race to build models that can be trusted with more than a chat window.