Anthropic released Claude Opus 4.8 today, an upgrade to its flagship model that ships at the same price as its predecessor and arrives alongside a handful of new product features. The benchmark gains are described as modest but real. The more distinctive claim, and the one worth focusing on, is about honesty.
A persistent problem with large language models is the confident wrong answer. Models sometimes claim to have made progress when the evidence is thin, declare a task complete when it isn’t, or wave through their own mistakes without flagging them. Anthropic says Opus 4.8 is meaningfully better at not doing that. Internal evaluations report the new model is roughly four times less likely than Opus 4.7 to let flaws in code it has written pass unremarked, and early testers describe it as more willing to surface uncertainty about its own work.
This is the kind of improvement that does not show up cleanly on a single benchmark line but compounds heavily in real use. An assistant that volunteers “this part of my work might be wrong” is dramatically more useful in agentic and code-review settings than one that polishes its own output uncritically.
What else launched
Three new product features ship alongside the model.
Effort control arrives in claude.ai and Cowork as a new selector next to the model picker. Users can now choose how hard Claude thinks about a given response. Higher settings mean deeper reasoning and better answers at the cost of more tokens consumed. Lower settings deliver faster responses and slower depletion of rate limits. The control is available on all plans, which is a notable concession for anyone who has hit a daily usage limit before lunch. Opus 4.8 defaults to “high,” with “extra” (called “xhigh” in Claude Code) and “max” available for harder work and long-running asynchronous workflows.
Dynamic workflows lands in Claude Code as a research preview. The feature lets Claude plan a complex task and then run hundreds of parallel subagents within a single session, with Opus 4.8 capable of letting agents run longer than before. The system verifies outputs before reporting back. Anthropic’s headline example is unusually ambitious: a codebase-scale migration across hundreds of thousands of lines of code, taken from kickoff through merge using the existing test suite as a quality bar. The feature is available in Claude Code for Enterprise, Team, and Max plans.
Messages API change. Developers building agentic systems get a useful primitive: the Messages API now accepts system entries inside the messages array. Practically, this lets a harness update Claude’s instructions mid-task, including permissions, token budgets, and environment context, without breaking the prompt cache or routing the change through a user turn. It is a small change with real implications for anyone building long-running agents.
Pricing and fast mode
Regular pricing for Opus 4.8 is unchanged: $5 per million input tokens and $25 per million output tokens, identical to Opus 4.7. Anthropic API users can access the new model via the identifier claude-opus-4-8.
The fast mode pricing is the more interesting shift. Opus 4.8’s fast mode, which Anthropic describes as running at 2.5x the speed of standard mode, costs $10 per million input tokens and $50 per million output tokens. That is three times cheaper than fast mode was for previous models. For latency-sensitive deployments, this materially reshapes the cost calculation.
On alignment
Model releases now routinely come with alignment statements, and Anthropic’s pre-deployment assessment for Opus 4.8 follows the format that has become standard practice. The company’s Alignment team reports that the model “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest,” with rates of misaligned behavior such as deception or cooperation with misuse substantially lower than Opus 4.7. The fuller picture sits in the Opus 4.8 System Card.
The alignment notes also reference a model that has not yet shipped to the general public, which is the more interesting thread to pull on.
What’s next: the Mythos teaser
Anthropic used the launch to disclose two forward-looking items.
The first is a model class beyond Opus. Anthropic says it plans to release a tier with higher intelligence than Opus, and confirms that a small number of organizations are already using a model called Claude Mythos Preview for cybersecurity work as part of something called Project Glasswing. The company says models at this capability level require stronger cyber safeguards before general release. It expects to bring Mythos-class models to all customers “in the coming weeks.”
The second is the more mundane but commercially important note: Anthropic is also working on smaller, cheaper models that deliver Opus-class capability at lower cost. That is the same compression pressure every frontier lab is now navigating.
The Mythos teaser is the part of this announcement worth tracking. Anthropic is signaling a new capability tier and gating it behind cyber safeguards rather than the usual staged rollout. Whether those safeguards become a template for releasing frontier models, or remain Anthropic-specific, is one of the more interesting questions of the next quarter.
The takeaway
Opus 4.8 is a calibration release more than a leap. The benchmark gains are real but the company itself describes them as modest. The features that matter most for daily users (the effort control, the dynamic workflows in Claude Code, the cheaper fast mode) are arguably more consequential than the model number ticking up.
The honesty improvement is the one worth keeping an eye on in production. If the four-times-less-likely-to-let-code-flaws-pass claim holds up at scale, the agentic use cases that depend on a model knowing when to stop and ask, rather than confidently barreling forward, become meaningfully more reliable. That is the kind of change that does not need a new model number to matter.
Frequently Asked Questions
How is Claude Opus 4.8 different from Opus 4.7?
The headline differences are calibration and honesty. Anthropic reports that Opus 4.8 is around four times less likely than its predecessor to let flaws in code pass unremarked, and is more likely to flag uncertainty about its own work. It also posts benchmark improvements across coding, agentic, reasoning, and knowledge tasks, and ships with new product features including effort control and dynamic workflows.
Does Opus 4.8 cost more than Opus 4.7?
No. Regular pricing is unchanged at $5 per million input tokens and $25 per million output tokens. Fast mode is three times cheaper than fast mode for previous models, at $10 per million input and $50 per million output tokens.
What is the new effort control?
A selector next to the model picker in claude.ai and Cowork that lets users choose how deeply Claude thinks about a response. Higher settings produce better answers and consume more tokens. Lower settings respond faster and use rate limits more slowly. The default is “high,” with “extra” and “max” available for difficult work. It is available on all plans.
What are dynamic workflows in Claude Code?
A research-preview feature that lets Claude plan a complex task and run hundreds of parallel subagents in a single session, then verify outputs before reporting back. Anthropic’s headline example is a codebase-scale migration across hundreds of thousands of lines of code, run end-to-end. It is available on Claude Code Enterprise, Team, and Max plans.
What is Project Glasswing and Claude Mythos?
Project Glasswing is the name Anthropic gave to a limited deployment of a higher-capability model class called Claude Mythos Preview, currently in use by a small number of organizations for cybersecurity work. Anthropic says it plans to release Mythos-class models to all customers in the coming weeks, pending stronger cyber safeguards.
