Anthropic introduced Claude Opus 4.8 as an updated top-tier model, pairing claimed improvements in coding and autonomous task performance with new controls over how much computation the system applies. The company made the model available through its consumer service, Claude Code and the Claude API under the identifier claude-opus-4-8.
For every million tokens, regular input cost $5 and the corresponding output charge was $25, matching Opus 4.7. Fast mode operated at 2.5 times the speed. Under that option, a million input tokens cost $10 and a million output tokens cost $50. Anthropic said that option was three times cheaper than comparable fast operation for previous models.
Users of claude.ai gained an effort setting. Opus 4.8 defaulted to high effort, which Anthropic said used a similar token count on coding tasks as the prior model’s default while improving results. Extra and max levels allowed greater token consumption for difficult assignments and long asynchronous work. Claude Code also received higher rate limits and a “dynamic workflows” feature for unusually large problems.
The company emphasized reliability as well as benchmark performance. According to Anthropic’s own evaluations and early-tester reports, Opus 4.8 more often identified uncertainty and avoided unsupported conclusions. It was about four times less likely than Opus 4.7 to leave flaws in its generated code unmentioned, the company said. Those findings were vendor claims; the supplied evidence did not include an independent assessment.
Anthropic’s alignment team reported higher measures of behavior supporting user autonomy and lower rates of deception or cooperation with misuse than Opus 4.7. It said the new model’s misalignment levels were similar to Claude Mythos Preview, a more capable system being tested by a small number of organizations for cybersecurity under Project Glasswing.
The release also signaled future product tiers. Anthropic described Opus 4.8 as a modest improvement and said it was working to deliver similar capabilities at lower cost. It expected stronger safeguards would allow Mythos-class systems to reach broader availability in the following weeks.
Alongside the model, the Model Hardware Standard entered research preview. Intended for agents operating physical devices, it opened to selected laboratories and advanced manufacturers. It offered as many as 10,000 scientists free Claude access for up to a year through eligible research teams, with Premium seats costing $15 monthly. Anthropic also announced a $5 million grant program for independent study of AI’s effects on wellbeing.
The announcement did not provide independent benchmark replication or production results from customers using the model. The launch combined a model update with workflow, research and safety initiatives, while its comparative performance claims remained those of the developer.
