Anthropic Launches Claude Opus 4.8 With Smarter Coding, More Honest Behavior
Anthropic has just unveiled its latest and most powerful publicly available AI model: Claude Opus 4.8. The update follows just 41 days after its predecessor Opus 4.7 and comes with improved capabilities in coding, reasoning, and agentic tasks. The even more powerful model Claude Mythos remains reserved for selected partners for the time being.
The unusually short development cycle of just six weeks has several causes. Opus 4.7 was perceived as disappointing by part of the user community, and at the same time competitors such as OpenAI with GPT-5.5 and Google with Gemini 3.5 Flash have recently released significant new models. Anthropic responds with an upgrade that achieves new best scores in benchmarks for coding, agentic tasks, reasoning, and knowledge work.
The most important new feature: Honesty and judgment
A central theme at launch is the improved reliability and honesty of the model. According to Anthropic, Opus 4.8 is around four times less likely than its predecessor to leave errors in its own code without comment. Early testers report that the model proactively flags uncertainties and refrains from making unsubstantiated claims.
Anthropic’s internal alignment team concludes that Opus 4.8 reaches new peak values in prosocial traits such as supporting user autonomy and shows significantly less deceptive or abuse-enabling behavior than Opus 4.7.
New features at a glance
Dynamic Workflows
Probably the biggest new feature for developers and businesses is Dynamic Workflows, available in Research Preview. It allows Claude Code to distribute complex tasks across hundreds of parallel subagents, verify the results, and then report back. This makes codebase-wide migrations across hundreds of thousands of lines of code possible, from planning to the finished merge. The feature is available for Enterprise, Team, and Max plans.
Effort Control
Users on claude.ai can now control how much effort the model invests in a response. Higher effort levels deliver better results but consume more tokens and rate limits. Lower levels respond faster and conserve the quota. The setting is available on all plans.
API Update: System entries in the messages array
For developers, there is a practical new addition to the Messages API: system entries can now be passed directly in the messages array. This makes it possible to update instructions in the middle of an ongoing agent session without interrupting the prompt cache or routing the update through a user turn.
Pricing: Same as the predecessor
Anthropic keeps prices for Opus 4.8 stable. A direct comparison of pricing models:
| Mode | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Standard | 5 USD | 25 USD |
| Fast Mode (2.5x speed) | 10 USD | 50 USD |
Particularly noteworthy: The Fast Mode is three times cheaper compared to previous models. The model is available via the Claude API under the identifier claude-opus-4-8 and is available worldwide.
Mythos remains locked for now
The even more powerful model Claude Mythos is still only accessible to a small group of selected partners, currently primarily for cybersecurity applications as part of Project Glasswing. The background is security concerns that emerged during an initial preview in April 2026. However, Anthropic signals that the necessary safeguards are being developed swiftly.
“We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks.”, the statement reads.
Independent benchmarks are still pending
Claude Opus 4.8 is a solid, if not revolutionary, update. The improvements in honesty, judgment, and agentic tasks are noticeable and address concrete points of criticism of its predecessor. With Dynamic Workflows and Effort Control, Opus 4.8 also gains practical new tools that are particularly relevant for developers and enterprise customers. Those waiting for the truly big leap will have to be a little more patient: Mythos is coming, but not yet for everyone.
Independent benchmarks for Claude Opus 4.8, such as those at Arena.ai or Artificial Analysis, are not yet available. At the latter, it will be particularly interesting to see whether Anthropic’s new model can beat GPT-5.5 from OpenAI. At Arena.ai, Anthropic models are already leading; it can be assumed that Opus 4.8 will move to the top.
