On Thursday, Anthropic launched Claude Opus 4.8, the latest and most advanced version of its flagship AI model. It’s available everywhere at the same price as its predecessor, Opus 4.7 ($5 per million input tokens and $25 per million output tokens).
Opus 4.8 boasts industry-leading scores on tasks like agentic coding and agentic computer use, which is par for the course for a new Anthropic model. The key differentiator that’s being underscored by the company is the model’s “honesty”—and by extension, its overall reliability.
According to a company blog post, Opus 4.8 specializes in catching its own mistakes and flagging them to users: “a general problem with AI models is that they sometimes jump to conclusions, confidently claiming to have made progress in their work despite the evidence being thin,” the company wrote. “Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims.”
For example, Michael Ran, a senior investment associate at asset management firm Bridgewater, was quoted in Anthropic’s blog post saying that Opus 4.8 was able to “proactively flag issues with the inputs and outputs of an analysis, something other models routinely missed and left to the users to catch.”
Opus 4.8 also presents a “substantially lower” risk of misaligned and dangerous behaviors, including the generation of harmful sexual content and “undermining liberal democracy,” according to the model’s system card.
Dynamic workflows and effort control
In addition to the new model, Anthropic also announced the launch of “dynamic workflows,” a new feature now available as a research preview, which allows Claude to handle more complex coding tasks by deploying hundreds of subagents that can work in parallel to one another.
Users can expect a noticeable improvement from Opus 4.8, especially for bigger coding tasks, but it’s not a game-changer. Anthropic even tried to hedge expectations, writing in its new blog post that Opus 4.8 is “a modest but tangible improvement on its predecessor,” Opus 4.7. That model debuted a little over a month ago and received a tepid early response from users, some of whom complained that its “adaptive thinking” feature sometimes caused it to spend too much time on tasks that should’ve been quick and easy, and not enough time on tasks that deserve more effort.
Perhaps in direct response to that complaint, Anthropic also announced on Thursday the launch of a new “effort control” panel (found in the model selector dropdown menu) for Claude, which lets you manually choose the amount of effort—and tokens—you want it to spend on a given task. It’s set to “Low” by default, and you can switch it to “Medium,” “High,” and “Max,” or toggle on adaptive thinking mode.
“Mythos-class models”
Anthropic also teased the upcoming debut of “a new class of model” with capabilities that are allegedly on par with those of Mythos, the mysterious model that’s been sending cold shivers up Silicon Valley’s spine. The company has yet to release the model publicly, citing the model’s unprecedented power and its cybersecurity risks.
According to its new blog post, Anthropic is currently working on testing safeguards for Mythos and expects to release “Mythos-class models to all our customers in the coming weeks.”
That’s obviously extremely vague, probably intentionally so. Time will tell whether these new models live up to the paradigm-shattering early rumors that have been circulating around Mythos, or if (more likely than not) the new category of models is a substantially watered-down version of the original behemoth. AI developers, after all, tend to hype up their own models’ abilities and dangers before they’re released, and in most cases, the reality doesn’t quite meet the expectations. (Remember all the excitement about GPT-5 being AGI?) Then again, maybe Anthropic is actually ready to unleash world-shattering models that were deemed an existential threat to global security just a couple of months ago. Time will tell, and we’ll report back as soon as we know more.
