Anthropic’s answer to enterprise complaints about the cost of frontier AI arrived on 1 September, and it was not in the price list. Claude Fable 5.1 shipped at the same $10 per million input tokens and $50 per million output tokens as Fable 5, which opened the Mythos class to general use in June. What moved was the price of a cache read, the charge for processing context the model has already stored, which fell 75% to $0.25 per million tokens.

That single line lands on a specific shape of work. Cache reads accumulate when a system sends a model the same material over and over: tone-of-voice documents, product catalogues, a quarter of campaign briefs, a fixed research corpus. Marketing teams running content or research agents against a stable body of reference material generate exactly that pattern, and they have been paying a tenth of the base input rate every time the model reads it again. By Anthropic’s estimate, the change lowers typical running costs by around 25%, and by as much as 45% for heavily agentic work.

Why does a cache-read cut matter more than the sticker price?

Anthropic priced Fable 5.1’s cache hits at 0.025 times the base input rate, against the 0.1 multiplier every other current Claude model uses. On a $10 base input price, that is the difference between $1.00 and $0.25 per million tokens. Cache writes did not move: five-minute writes remain $12.50 per million tokens and one-hour writes $20, according to Anthropic’s platform pricing documentation.

The asymmetry decides who benefits. A workflow that writes a brief to cache and reads it once gains almost nothing. A workflow that holds the same 100,000-token brand and product context across a session lasting hours, re-reading it on every turn, sees most of its bill fall. On OpenRouter figures reported last month, cached prompts already account for more than 85% of the tokens AI agents consume, which puts the cut on the bulk of agentic spending rather than its margins.

The 25% and 45% figures are Anthropic’s own, and the company was specific about where they came from. They are drawn from four weeks of actual usage in August 2026 across Claude Enterprise, Claude Code and the API, measured at default effort. They have not been independently verified. Teams with their own token logs can check the claim directly rather than take it on trust, by working out what share of their input spend is cache reads. Where that share is small, the saving will be too. Anthropic lists and bills these prices in US dollars.

What did Fable 5.1 actually get better at?

Anthropic reported its largest gain on Terminal-Bench-Science 0.1, an agentic scientific research test, where Fable 5.1 scored 52.6% against Fable 5’s 24.7% and Claude Opus 5’s 29.0%. On Terminal-Bench 4.0, an agentic coding benchmark, it scored 55.8% against Fable 5’s 42.0% and GPT-5.6 Sol’s 37.3%.

One disclosure in that table is more revealing than the scores. Claude Mythos 5.1, which Anthropic describes as the same underlying model with looser safeguards, scored 60.9% on the coding test against Fable 5.1’s 55.8%. Anthropic attributed the five-point gap to tasks where the earlier, less precise cybersecurity safeguards intervened, and said it expects the difference to narrow. Vendors rarely publish the cost of their own guardrails. Anthropic also gave a standard error of 3.5 to 4.5 points per model on the science benchmark, which is more arithmetic than most launch tables carry.

None of these benchmarks measure marketing work. The nearest signals come from early-access partners, and they are those partners’ own claims rather than tested results. Canva’s head of AI, Danny Wu, said “the standout in Claude Fable 5.1 is the writing”, reporting that it followed the company’s writing guidance more closely in blind tests against Fable 5. Glean said its judges preferred Fable 5.1’s answers roughly two to one over Fable 5 on knowledge questions, research and drafting. Neither company published its methodology.

What changes for data retention, safeguards and disclosure?

Anthropic introduced Enterprise Frontier Safeguards, which keeps customer data on infrastructure the customer controls rather than on Anthropic’s systems, with human review carried out by the customer by default. The company said it built the system with more than 100 customers and with Amazon Web Services, Google Cloud and Microsoft Azure. It rolls out in phases starting this autumn; until it reaches an eligible customer, that customer can run Fable 5.1 with zero data retention. The standing default outside those arrangements is 30-day retention for safety monitoring.

The safeguards themselves became less blunt. Anthropic said Claude Code users should see around 60% fewer interventions per session from its cybersecurity safeguards. Its biology safeguards, the company said, now fire 85% less often on benign elementary biology and medical questions than the versions that shipped with Fable 5. Fable 5.1 can now be used to identify software vulnerabilities, though not to build exploits for them; penetration testing, exploit generation and binary-based vulnerability scanning still route to Anthropic’s Opus models.

The watermark detector is the quieter change, and the one that travels furthest. Fable 5.1 carries the invisible text mark Anthropic has applied to models released since 2 August 2026 under the EU AI Act’s Code of Practice on Transparency of AI-Generated Content. When the Helm reported on how that mark works on 17 August, no detector existed outside Anthropic and the company had given no date for one. It now says a detection API is in private preview, available to regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises with their own compliance duties, with access widening over time. Copy carrying the mark can therefore be tested by parties holding the key, which moves disclosure from a matter of internal judgement towards a matter of record.

Enterprise Frontier Safeguards begins its phased rollout this autumn, though Anthropic has not named a date. Claude Mythos 5.1 remains available only to vetted organisations in the United States.