What Claude Opus 4.8 Is—and How It Fits with Mythos
Claude Opus 4.8 is Anthropic’s latest flagship Claude Opus 4.8 model, a bridge upgrade that improves reasoning, coding, and safety while remaining less capable than the Mythos AI model preview that still sits at the top of Anthropic’s internal capability ladder. Anthropic describes Opus 4.8 as a “modest but tangible improvement” over Opus 4.7 across internal benchmarks, covering software engineering, agentic tasks, multimodal input, and reasoning. At the same time, the company is explicit that Opus 4.8 does not surpass Claude Mythos Preview, which remains restricted under Project Glasswing to a select set of partners and cybersecurity experts. This puts Opus 4.8 in an interesting position: it is the Anthropic new release most users can run today, yet it is framed as a safe, incremental step that deliberately stops short of the capability frontier Mythos currently represents.
Incremental Gains: What Opus 4.8 Improves Over Opus 4.7
For users of Claude Opus 4.8, the upgrade over Opus 4.7 is about sharper behavior rather than a dramatic power jump. Anthropic highlights better honesty: the model is more likely to say when it lacks enough information and less likely to make unsupported claims. On coding tasks, evaluations “showed Opus 4.8 was around four times less likely than its predecessor to allow flaws in code it generated to go unremarked,” which matters for production use. Internal alignment testing shows lower deceptive behavior and a profile close to Claude Mythos Preview when it comes to following user interests and instructions. In cybersecurity, Opus 4.8 without safeguards is modestly more capable than Opus 4.7, but with guardrails enabled the two are roughly equal. Overall, Opus 4.8 refines reliability, safety, and agentic execution without shifting Anthropic into a new risk tier.
Why Mythos Still Defines the Capability Frontier
Even as Opus 4.8 improves, Anthropic is clear that the Mythos AI model preview remains its most capable system. Internal safety evaluations around biological and virology risk show why. On a DNA Synthesis Screening Evasion test, Opus 4.8 scores 0.30 on one criterion versus Mythos Preview’s 0.842, where lower is safer because it reflects less ability to evade screening. On the Virology Capabilities Test, Opus 4.8 scores 0.470 compared to 0.574 for Mythos Preview. Anthropic concludes that “Opus 4.8 does not advance the capability frontier beyond our most capable model.” In cybersecurity, Mythos Preview is substantially stronger than both Opus 4.7 and 4.8, with reports that it has found more than 200 Firefox flaws and that its autonomous exploit discovery helped prompt rivals to accelerate their own models. This capability gap keeps Mythos under tighter controls for now.
Mythos-Class Models for Everyone: Anthropic’s Timeline
Today, Mythos Preview is confined to Project Glasswing partners and security professionals, but that is set to change. Anthropic says it is making “swift progress” on cyber safeguards strong enough to support general availability, and it “expects to be able to bring Mythos-class models to all our customers in the coming weeks.” Security experts note that this staged Anthropic new release strategy has tradeoffs: it gives defenders lead time, but the window between release and broad deployment of defenses stays risky. Mythos is also more expensive to run than earlier Opus models, which could limit some attackers but not determined or well-funded ones. When released with stronger guardrails, Mythos-class systems will represent a clear step beyond Opus 4.8 on AI model comparison metrics, especially in cyber and advanced reasoning, but they will arrive with heavier safety layers and likely narrower access controls.
How Users Should Approach Opus 4.8 While Waiting for Mythos
For most customers, the practical choice today is to use Claude Opus 4.8 as the default workhorse while preparing workflows that can later tap Mythos-class power. Opus 4.8 keeps Opus 4.7 pricing, so existing users can upgrade without cost changes. Early testers report that it is “more reliable and sharper in its judgement when it’s performing agentic tasks,” especially when combined with Dynamic Workflows in Claude Code, which can orchestrate hundreds of parallel subagents and verify outputs in a single session. Anthropic also adds controls to tune how much effort the model spends on each task, letting teams balance speed, depth, and cost. In short, Opus 4.8 offers a safer, more dependable daily model now, while Anthropic positions Mythos-class systems as a coming high-end tier for the most demanding and sensitive workloads.