Anthropic released Claude Fable 5.1 and Mythos 5.1 on 1 September 2026 without touching the headline $10 / $50 per-million rate. What moved was the cache read, cut from $1.00 to $0.25 per million tokens. The resulting price shape is strange enough to be deliberate: Fable 5.1's cached input is half the price of Opus 5's despite a base input price twice as high, and only 25% above Sonnet 5's despite a five-fold gap on uncached tokens. Ramp transaction data reported by the Financial Times suggests why — Fable 5 had captured only about 11% of Anthropic model spending more than two months after launch. The discount is engineered to arrive for the workload Anthropic wants to win, and not for anyone else.
On 1 September 2026 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, and did not change the price. Fable 5.1 bills at $10 per million input tokens and $50 per million output tokens, the same rates Fable 5 carried. Yet Anthropic advertises an effective cost reduction of roughly 25% for typical workloads and as much as 45% for heavily agentic ones. Both statements are true. The reconciliation is a single line item.
A cache read on Fable 5.1 costs $0.25 per million tokens, down from $1.00 on Fable 5. Cache writes did not move: $12.50 per million for the five-minute cache, $20 per million for the one-hour cache. Only the price of reading back something you already paid to store went down, and it went down by three quarters.
Anthropic's published pricing table makes the mechanism explicit. Across the Claude line a cache hit is billed at 0.1x the base input price: Opus 5 at $5 input and $0.50 read, Sonnet 5 at $2 and $0.20, Haiku 4.5 at $1 and $0.10. Fable 5.1 and Mythos 5.1 are the exceptions. Their cache reads are priced at 0.025x input, a fact Anthropic footnotes directly in its documentation, noting that all other models use the standard 0.1x multiplier.
That produces an inversion worth stating plainly, because none of the launch coverage set the numbers side by side this way. Fable 5.1's uncached input costs twice what Opus 5's does. Its cached input costs half what Opus 5's does. Against Sonnet 5 the base input gap is five to one, $10 against $2, but the cache read gap is 25%: $0.25 against $0.20. On repeated context, the most expensive model in Anthropic's catalogue is priced within a rounding error of the cheapest.
Fable 5.1 is the only model in Anthropic's lineup where reading cached context costs less than it does on a model with half the base input price. On uncached tokens it remains among the most expensive commercially available models anywhere.
Prompt caching stores a processed prefix, such as a system prompt, a set of tool definitions, a codebase, a long document or an accumulated conversation, so that subsequent requests do not reprocess it from scratch. A chatbot answering one question benefits marginally. An agent benefits enormously, because an agent's defining behaviour is re-reading. Every turn of a coding agent re-sends the same repository context, the same tool schemas and the same instructions, plus whatever it has learned since.
Work the arithmetic on a concrete shape. Take an agent holding 200,000 tokens of cached context across 50 turns in an hour. Under Fable 5's $1.00 read those replays cost 50 x 0.2M x $1.00, or $10.00. Under Fable 5.1's $0.25 the same replays cost $2.50. Nothing about the model's list price, output price or write price changed. The recurring cost of the agent remembering fell by $7.50 an hour. Scale that across a fleet of long-running agents and the 25-to-45% claim stops looking like marketing and starts looking like a description of one specific workload.
The break-even maths barely moved, which is the point. A five-minute cache write costs 1.25x input, so writing carries a $2.50-per-million premium over simply sending the tokens uncached; at either read price, one read recovers it. The one-hour write at 2x input needs a little over two reads. What changed is not whether caching pays, because it already did, but the floor on how cheap sustained reuse can get. Anthropic did not make caching worthwhile. It made unbounded caching cheap.
There is a countervailing detail in the same documentation that no coverage of the launch mentioned, and it cuts the other way. Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text than the one Sonnet 4.6 and earlier used. Per-token comparisons across model generations therefore understate the cost of a given document. Anyone building a cost model from list prices alone, without normalising for tokenizer, will be wrong in a direction that flatters the newer models.
The shape of the discount makes more sense against a number Anthropic did not put in its announcement. The Financial Times, analysing transaction data from expense management provider Ramp covering roughly 70,000 companies, found that more than two months after launch Fable 5 accounted for only about 11% of Anthropic model spending, while the cheaper Opus 5 and Opus 4.8 gained share. The Information separately reported enterprise anxiety about unpredictable AI bills, including ServiceNow monitoring employee usage after moving through its annual Anthropic budget faster than expected.
Read together, those reports describe a specific commercial failure. Capability was not the objection. Fable 5 was Anthropic's most capable model and customers still routed the bulk of their spend to something cheaper. The obstacle was the bill, and particularly the unpredictability of the bill for workloads that run without a human deciding when to stop.
A general price cut would have addressed that by giving up margin on every token, including the one-shot generations where Anthropic faces the fiercest price competition. Gemini 3.7 Flash lists at $0.75 per million input through the end of 2026; several Chinese models sit below a dollar. Fable cannot win that comparison and Anthropic has not tried to. Instead it cut the price of exactly the traffic that only appears when a customer commits to leaving an agent running, traffic that a rival's one-shot pricing does not address at all. It is price discrimination, executed cleanly, and it is aimed at converting the 11%.
One comparison makes the targeting obvious. OpenAI's promotional pricing for GPT-5.6 Sol runs at $5 per million input with cached input at $0.40, through at least 21 November. Fable 5.1 costs twice as much as that on uncached input and 38% less on cached. Whichever model is cheaper for you is now a question about your context reuse ratio, not about the models.
For buyers this reframes model selection. The useful comparison is cost per completed task, including retries, replayed context, tool calls and the tokens consumed before a usable answer arrives. A code review agent sitting on an enormous reusable repository context and a report generator dominated by fresh output tokens now have materially different economics on the same model. Anthropic's other levers compound the effect: the Batch API halves Fable 5.1 to $5 and $25 per million for asynchronous work, and the caching multipliers stack on top of it, while US-only inference adds a 1.1x multiplier across every category.
The second half of the release is architectural rather than commercial. Fable 5.1 and Mythos 5.1 are the same underlying model. Fable is generally available with Anthropic's production safeguards applied. Mythos exposes more permissive safeguards and is restricted to vetted cybersecurity and life sciences organisations through access programmes. The difference is a setting, not a set of weights.
That relocates the access question. Historically, whether an organisation could do a particular thing with a frontier model depended on what the model could do. Under this split it depends on whether the organisation can prove who it is. Anthropic reports that Fable 5.1's cyber protections produce around 60% fewer interventions per Claude Code session than Fable 5's, and that vulnerability discovery for defensive purposes is permitted while exploit generation, penetration testing and some binary-based scanning remain restricted or redirected. Precision, rather than the mere existence of a filter, is what makes a security workflow usable in production.
Every performance figure attached to this launch is vendor-reported, generated in Anthropic's own evaluation setup, and should be read as such. Anthropic reports Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1, against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol; and 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5 and 52.3% for Opus 5, with Mythos 5.1 reaching 60.9% under its more permissive safeguards. The science benchmark is at version 0.1. Anthropic itself notes that production safeguards can affect scores.
The customer results circulating with the launch are testimonials supplied by Anthropic, not reproduced experiments. Millennium says the model traced a rare software crash to a bug inside an external vendor library after four to five years of the problem resisting explanation. Ramp describes a 38-hour unattended machine learning run in which the model re-evaluated a prior result and launched six experiments. Browserbase reports 82% completion on its hardest browser agent benchmark against 74% for Opus 5 and 57% for Fable 5. These are useful as descriptions of the kind of work being attempted. They are not evidence of a general capability margin.
One counter-argument deserves stating, because it runs opposite to everything above. On 29 August 2026 Anthropic announced changes to Claude Code's subscription limits. A temporary 50% increase in weekly limits remains in place through 13 September; from 14 September, standard weekly limits rise permanently by 25% for Pro, Max, Team and seat-based Enterprise plans. Measured against the pre-promotion baseline that is more. Measured against what subscribers have today it is less, and Anthropic said so itself in a clarification posted after it deleted the original announcement, describing the change as a 17% reduction in weekly limits on Claude Code compared to today.
So the metered API got cheaper for the workload Anthropic wants to grow, and the flat-rate subscription got tighter for the workload it apparently does not. Both moves point the same way. Consumption where the customer sees and controls the meter is the business being optimised for; flat-rate consumption of the same underlying compute is the one being trimmed. Anyone reading the launch as a straightforward price cut should check which of those two products they are actually buying.
The honest summary is narrow. Anthropic did not make Fable cheaper. It made remembering cheaper, and remembering is what agents spend their money on. Whether that lands depends entirely on whether your workload reuses context, and on Anthropic's ability to hold a $10 base input price while much of the market prices frontier-adjacent inference in cents.
Only for some. The list price is identical at $10 per million input tokens and $50 per million output, and cache write prices are unchanged. The only reduction is the cache read, from $1.00 to $0.25 per million. If your workload sends mostly fresh input and produces long outputs, your bill will not change at all. If it repeatedly replays a large cached prefix, the saving can be substantial. Anthropic's own figures of roughly 25% for typical workloads and up to 45% for highly agentic ones describe that spread.
We track model pricing, inference economics and the gap between list prices and what enterprises really pay.
Read our AI coverageAnthropic was not breached. The malware was already on the customer's machine, and it took a session cookie rather than a password — which is why two-factor authentication did nothing and why server-side revocation is the only lever the vendor has.
OpenAI's incident report on the Hugging Face breach leads on a security failure. The remarkable part is a verbatim chain-of-thought in which a model identifies its action as unauthorised and proceeds anyway — and an escape route that was a package manager, not a superintelligence.
Three separate events happened this week and coverage is treating them as one. Ninety days of notice for the chat interface, six months for the API snapshots, twelve months for the Assistants API — and no automated migration anywhere.