benchr
On paper this is a point release. Opus 4.7 to 4.8, same price, same tier. That framing undersells what Anthropic chose to spend the upgrade on: making the model better at running long, autonomous coding work, and better at telling you when it is unsure. Both are about trust in agentic settings, where you are not watching every step.
What Anthropic actually announced
- Pricing held flat. Regular usage stays at $5 input / $25 output per million tokens, identical to 4.7. A faster-serving "fast mode" is priced at $10 input / $50 output. For where that sits against everyone else, see our AI model pricing comparison.
- A concrete honesty improvement. Anthropic states Opus 4.8 is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked." In plain terms: it is more willing to flag its own mistakes instead of presenting broken code confidently.
- Dynamic workflows. A new capability in Claude Code where the model can plan a task and then "run hundreds of parallel subagents in a single session," including "codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge." Anthropic lists it as available in Claude Code for Enterprise, Team, and Max plans.
- Alignment framed against Mythos. Anthropic reports rates of misaligned behavior "substantially lower than Opus 4.7," and positions the model as a step toward the safety profile of its gated frontier work.
Why the agentic framing matters
A model you supervise turn-by-turn does not need to flag its own uncertainty very loudly — you are right there. A model running hundreds of subagents through a migration for an hour absolutely does, because nobody is reading each step. The honesty gain and the dynamic-workflows feature are the same bet from two directions: Opus 4.8 is built to be left alone with bigger tasks.
That is the practical difference from the version number. If your use of Claude is interactive chat, 4.8 is a modest, free-of-extra-cost upgrade. If your use is autonomous coding agents, the combination of longer-running agents and a model that interrupts itself when something looks wrong is the part that changes how much you can safely delegate.
The Mythos pressure
Analysis: The launch post does not end with Opus 4.8. Anthropic says it plans "to release a new class of model with even higher intelligence than Opus" and expects "to bring Mythos-class models to all our customers in the coming weeks." That reframes 4.8 as a way station rather than a destination — and it lines up with the company's Project Glasswing expansion, where the unreleased Mythos Preview is already doing security work at scale.
The competitive read is straightforward. Anthropic is signaling that its strongest model is close to general availability, while shipping a public model tuned for exactly the autonomous workloads that model would dominate. The pressure runs in two directions at once — on rivals, and on Anthropic's own current flagship.
Practical takeaways
- No reason to delay upgrading. Same price, strictly better behavior on self-checking. If you are on 4.7, moving to 4.8 is low-risk.
- Re-test your agent guardrails. If you cap how long an agent may run or how many subagents it spawns, dynamic workflows changes the cost and blast radius of those limits. Revisit them deliberately.
- Compare before committing budget. The interesting head-to-head is still against GPT-5.5, which charges the same $5 input but diverges fast after that — see Opus 4.8 vs GPT-5.5, or browse the full model rankings to place it in context.
- Treat "coming weeks" as a plan, not a date. A Mythos-class general release is announced, not shipped. Build for what is available today.
One discipline note, since this is a launch story: the only benchmark numbers worth quoting are the ones in front of a methodology you can read. We keep the scored testing in the review rather than repeating marketing figures here.