Anthropic launches Claude Opus 4.8, with honesty as a flag when it makes mistakes, and announces the arrival of Mythos

Anthropic has released Claude Opus 4.8, with improvements that make you a more effective collaborator and they make him more honest in his answers, which is now available at the same price, and has announced the availability of Mythos for the coming weeks.

Claude Opus 4.8 is available this Thursday with improvements that allow you to be more reliable in the performance of agents and more honest in their responses. “We train all our models under this premise, so that they, at least, try to avoid claims that cannot be supported,” says Anthropic from the announcement on its website.

This data is supported by Anthropic stating that in the tests carried out by the first testers, Claude Opus 4.8 “is more prone to point out the uncertainties about its operation and less likely to make unsubstantiated claims.

As for the positive qualities, Opus 4.8 also shows progress in what the company calls “prosocial traits”, that is, supporting user autonomy and updating in the user’s best interest.

The evaluation also showed that Opus 4.8 has rates of misaligned behavior that are similar to those of your best-aligned modelClaude Mythos PreviewAnthropic has pointed out about alignment, understood as the process of designing and training an AI so that its objectives, behaviors and decisions coincide with human values, ethics and intentions.

On the other hand, Claude Opus 4.8 introduces dynamic workflows, which, being available in the research preview version, allows AI to plan work and run hundreds of parallel subagents in a single session to tackle more complex tasks in Claude Code, in addition to verifying the results.

Another novelty is effort control, which allows users to choose the amount of effort Claude puts into an answer. It is a measure that helps mitigate the increased consumption of tokens that is already being attributed to Claude Opus 4.8 and that is being reported by the user community on Reddit.

The AI ​​laboratory released Claude Opus 4.7 in April of this year, which has been characterized by improved high-quality image processing and greater precision in following instructions. In the coming weeks, Anthropic expects to launch Mythos with more robust cybersecurity measures.

“We are making rapid progress in developing those safeguards and look forward to bringing Mythos-class models to all of our customers in the coming weeks,” he said.

The price of regular use of Claude Opus 4.8 has not changed compared to Opus 4.7and is still $5 per million tokens in and $25 per million tokens out. For the fast mode it remains at 10 dollars per million input tokens and 50 dollars per million output tokens. claude-opus-4-8 is now available through the Claude API for developers.