Anthropic's Opus 5 Update: Even the AI Remembers Less When It's Cheaper

Earlier today, Anthropic's flagship Opus 5 model launched at roughly half the cost of the top-tier Fable 5, with near-equivalent performance on key benchmarks and a privacy carve-out that has drawn pushback from the same privacy-conscious users who objected to the original 30-day data retention window on the company's heavier models.

Underneath those reader-facing changes sits a deliberate design choice that runs through Anthropic's lineup: capability governs friction. The cheaper, less capable model is the one that forgets you fastest, gets watched the least, and answers questions the bigger model won't.

For most use cases, that makes Opus 5 preferable to Fable 5, despite sitting a tier below it. On CursorBench 3.2 at maximum effort, it scores within 0.5% of Fable 5's peak while costing half as much per task. On OSWorld 2.0, a computer-use benchmark, it beats Fable 5 at just over a third of the cost.

The story is the second-order design: in Anthropic's tiering, refusal is becoming a routing problem rather than a hard stop.

The 30-day carve-out is the most concrete user-facing benefit. When Anthropic introduced the policy covering Fable and Mythos, privacy-conscious users objected that their prompts and outputs were being held for a month. Opus 5, like its predecessor Opus 4.8, is not subject to that window. The cheaper, less capable model is the one Anthropic is willing to forget faster.

The classifier number cuts the same direction. Anthropic expects its safety classifiers to engage about 85% less often on Opus 5 than on Fable 5, a reflection of the lighter touch given to the less capable model. Where the classifiers do engage, the company is rolling out an opt-in beta called Automatic Fallbacks, which routes the refused request to a less powerful model so API users get a functional response instead of an error. The refused prompt is answered; it is just answered by a smaller Claude.

Opus 5 still refuses to scan compiled software binaries for vulnerabilities, but it is allowed to search source code, where the defensive-use case is more likely. Lighter friction is not the same as no friction, and the binary-versus-source-code distinction is exactly the kind of detail a sloppy paraphrase will get wrong.

Opus 4.8 shipped on May 28, 2026. Mythos 5, Fable 5, and Sonnet 5 all launched in June, leaving only the lightweight Haiku model still on a pre-5-series tier. Opus 5 arrived two months after its predecessor, and the rest of the family followed in a single month. The new model is not a standalone event; it is the slot below the flagship being filled.

For a buyer, the practical question is which model to route which workload through, and at what price. At half the cost per task with near-flagship scores on coding and knowledge work, the default answer is Opus 5 for most production traffic, with Fable 5 reserved for the hardest agentic and cybersecurity-sensitive work where the heavier safety net is the point.