Anthropic published Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. The listed price does not change — around $10 per million input tokens, $50 per million output — but the cost of cache reads drops from around $1 to $0.25, a 75% decrease. According to the company, this cuts real costs by around 25% for typical usage and up to 45% for heavily agentic workloads. Scores on long-running autonomous tasks improve far more than general reasoning.
Another model release, it would seem. Except the real story is not where you would usually look, and this announcement is more interesting than most of what we cover.
The real issue: the cut is where nobody is looking
This is the exact opposite of what we described two days ago about the Sonnet 5 price increase, and the mechanism is the same: what matters is not the listed price.
Here, the price per million tokens does not move. What drops is cache reads — that is, what you pay when the model reuses already-processed context rather than processing it again.
We explained this mechanism in our article on prompt caching. For typical conversational use, this line item is marginal. For an agent, it is dominant: an agent constantly re-reads the same codebase, the same system instructions, the same tool definitions, the same accumulated history.
Hence the gap between 25% and 45% savings depending on usage profile. Anthropic says it measured these figures over four weeks of real usage in August, which is more honest than a theoretical projection.
Deliberately lopsided gains
The second point is the most telling, and it deserves a close read.
On general reasoning, the improvement is a few points. On autonomous, long-running, tool-using tasks, it is far stronger: the score on a terminal scientific-work benchmark would have more than doubled, and the one covering business workflows would rise from around 17% to 31%.
This imbalance is no accident. It is a statement of intent: this model is not optimised to answer your questions better, it is optimised to run agents.
Combined with the cache price cut, the whole forms a coherent product aimed at a specific segment. This confirms what we have been documenting for weeks: agents have become the battleground where the competition plays out, as inference-dedicated chips and the standardisation of agent protocols already showed.
These results are reported by the company and have not been independently audited. To its credit, Anthropic itself raises several caveats: production guardrails can affect scores, and some evaluations are not directly comparable to previously published results. That is a transparency you do not always find elsewhere, and it does not replace external measurement. As always, the only evaluation that matters is the one you run on your own tasks.
The two-tier structure, which is a governance novelty
Here is the most interesting aspect, and it goes beyond the technical.
Fable 5.1 and Mythos 5.1 are the same underlying model, with different guardrails. Fable is the generally available version, with production protections. Mythos is accessible through restricted-access programmes, reserved for verified organisations in cybersecurity and life sciences, and limited to US organisations.
This is a concrete answer to a dilemma we regularly describe. Certain capabilities — finding vulnerabilities, working on biological topics — are valuable for defence and dangerous in open access. We raised this about GLM-5.3 and its thousand found flaws: the asymmetry favours the attacker, who only needs one flaw.
The answer chosen here is to separate not the models but the permissions, with vetting up front. It is the same principle we described for agents authorised to pay: you do not restrict the capability, you vet who gets access to it.
The June context, which explains a lot
It is worth recalling what happened, because it sheds light on these choices.
Fable 5 and Mythos 5 were released on 9 June. They were briefly pulled from the market following an export-control directive, before being reinstated. We covered that episode as proof that a model delivered as a service can be cut off overnight.
The guardrail changes introduced in this version are a direct response to that episode. In other words, a regulatory constraint produced a product architecture. That is a clear illustration of what we have written: what gets decided in a ministry now weighs as much as what gets decided in a lab.
What it says about the company's trajectory
This release comes with a figure worth noting: an annualised revenue run rate above $30 billion, versus around $9 billion at the end of 2025.
That matches what we reported about the preparation for an initial public offering. And it explains the compute race we have been documenting, with the company simultaneously securing accelerators from multiple suppliers.
Twelve weeks between two generations is also a notable pace, and it contrasts with the four successive delays of Gemini 3.5 Pro. The ability to ship on a regular cadence has become a competitive advantage in itself.
What to take away
Three things, if you only keep three.
If you run agents, this release concerns you directly, and the calculation to redo is about the share of cached context in your requests. That is where the savings are, not in the listed price.
If you use a conversational assistant, the change will be modest. The gains are not there, and that is deliberate.
If you follow the governance of these technologies, the two-tier structure is the real story. This is the first time a major lab has so explicitly split the same model into two access regimes based on user vetting. If the setup works, it will be copied. If it leaks, it will serve as ammunition for those who argue these capabilities should not exist in this form.