Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, describing them as "the world’s most advanced models for coding and knowledge work." The headline traveling with it is a cost reduction of around 25%. That number is real and it is not a price cut: input and output pricing is identical to Fable 5, at $10 and $50 per million tokens. What changed is the cache, and separately, something we flagged as precedent-setting in June has quietly moved. This piece covers what actually shipped, why the list price matters more than the percentage, where the savings genuinely come from, the retention reversal, what it does to the Opus 5 question, how Mythos 5.1 access works, and what to do about any of it.
What Claude Fable 5.1 actually is
Two models with one relationship, stated plainly by Anthropic: "Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards."
Fable 5.1 is generally available, with the API string claude-fable-5-1. Mythos 5.1 is restricted, carrying "safeguards specifically designed to support work in cybersecurity and the life sciences," and reached through verification programs rather than bought.
That structure is unchanged from the original release, which we covered in Claude Fable 5: one model, two safeguard configurations, one of them public. If you understood the Fable and Mythos split in June, you understand it now.
The announcement does not mention deprecating Fable 5, and no retirement date was published.
The list price did not change
This is the part to be careful about, because the framing everywhere is "25% cheaper" and the sticker did not move.
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens. Those are exactly the numbers Fable 5 launched with in June and carried through the metering changes in July.
So a team that sends a request today and an identical request tomorrow, with no caching involved, pays the same. The reduction Anthropic describes is a claim about typical workloads, not about the rate card, and the two are only the same thing if your workload matches the assumptions behind the word "typical."
Anthropic’s own wording is careful and worth quoting exactly: "For typical workloads, costs are reduced by around 25% relative to Fable 5. For complex coding and highly agentic tasks, the savings could be up to around 45%." Note "around," note "could be," and note that both figures are vendor estimates with no published methodology.
Where the savings actually come from
The mechanism is stated in one line: "Cache reads now cost 75% less, or $0.25 per million tokens."
That is a large cut on a specific line item, and it explains the shape of the claim. Prompt caching benefits workloads that repeatedly send the same large context: long system prompts, big document sets, agent loops that re-read the same state on every turn. The more of your token spend sits in cache reads, the closer you get to the 45% end. The less it does, the closer you get to zero.
Which means the honest way to evaluate this is not to accept 25% and adjust your budget. It is to look at your own billing and find out what share of your token spend is currently cache reads. If you do not cache at all, this release does not reduce your costs. If you run long-context agents, it may reduce them more than the headline suggests.
That is a much better position for Anthropic than a straight price cut, incidentally, because it rewards the usage pattern that agentic products generate, and it is the same pattern that made Fable 5’s metered pricing painful for Pro subscribers, where output at $50 per million was what consumed the credit.
The retention policy moved, and that is the real news
When Fable 5 launched, we called the data-retention policy the most operationally significant detail in the release, and described it as setting a possible industry precedent. Anthropic required a 30-day retention period on all traffic to the model regardless of whether a customer held a zero-retention agreement, overriding those agreements for this model class. We wrote at the time that for organizations with strict regulatory constraints, it could be a deal-breaker that locked them out of Fable entirely.
That has changed. Anthropic now describes Enterprise Frontier Safeguards, or EFS, as giving "customers complete privacy" while "still being state-of-the-art at preventing adversarial use," and says that until EFS is ready, "eligible customers will be able to use Fable 5.1 with zero data retention."
Read that against June and it is a reversal of the specific constraint that made Fable unusable for a category of buyer. The trade-off we described, maximum capability with mandatory retention or zero retention with a lesser model, is at least partially withdrawn.
Two cautions before anyone in a regulated industry acts on it. "Eligible customers" is undefined in the announcement, and eligibility is doing real work in that sentence. And EFS itself is described as not yet ready, so the zero-retention arrangement is explicitly an interim state, with no published detail on what replaces it or when. If your compliance position depends on this, get the eligibility criteria and the EFS transition terms in writing rather than from an announcement page.
What this does to the Opus 5 question
In July we published Opus 5 vs Fable 5 and the recommendation was clear: route most work to Opus 5, at $5 and $25 per million, because Anthropic’s own evaluations had it matching or beating Fable 5 on a lot of coding and knowledge work at half the price. Fable was the specialist for the hardest problems.
Fable 5.1 does not resolve that by lowering its price, because it did not lower its price. It reopens it in a narrower way: if your workload is cache-heavy, the effective gap between Opus 5 and Fable 5.1 is smaller than the rate card suggests, and the new positioning as most advanced for coding and knowledge work is aimed squarely at the ground Opus 5 was said to hold.
The comparison that would settle it does not exist yet. Anthropic has not published head-to-head evaluations of Fable 5.1 against Opus 5, and no independent benchmarks have landed on a model released hours ago. Treat the capability framing as vendor positioning until someone tests it, which is the same caution that post applied to the original comparison.
Mythos 5.1 and how access works
Mythos 5.1 is not purchasable. It is reached through two verification programs: a Cyber Verification Program for defensive security work, and a Life Sciences Verification Program for life sciences professionals.
One availability detail matters for planning: the Life Sciences program is described as currently available only to a set of US organizations. That is a real limit rather than a rollout note, and non-US teams evaluating Mythos for life sciences work should not assume access is coming on any particular timeline.
The structure is consistent with how Anthropic has handled this model class since April, including through the suspension and restoration we covered in Claude Fable 5 is back. Public access goes to the safeguarded variant; the unrestricted one goes through a gate with a named program behind it.
What to actually do
Pull your caching numbers first. Everything about whether this release saves you anything depends on what fraction of your token spend is cache reads, and that is a number you already have.
Do not rewrite budgets from the 25% figure. It is a vendor estimate about typical workloads with no methodology attached, and your workload is either more cache-heavy than typical or less.
If you passed on Fable 5 over the retention policy, look again, and get eligibility and EFS terms confirmed in writing before you commit. This is the single change most likely to move a real decision, and it is the one the announcement says least about.
If you are already on Opus 5 for general work, stay there for now. Nothing published today gives you a reason to move, and the head-to-head evidence that would justify it does not exist.
And leave claude-fable-5 alone until you have tested claude-fable-5-1 on your own evaluations. Anthropic has not announced a retirement date for Fable 5, so there is no clock forcing a migration.
Frequently Asked Questions
What is Claude Fable 5.1?
It is the generally available Mythos-class model Anthropic released on September 1, 2026, alongside the restricted Claude Mythos 5.1. Anthropic states the two are “the same model, but with different levels of safeguards,” and positions them as “the world’s most advanced models for coding and knowledge work.” The API string is `claude-fable-5-1`.
Is Fable 5.1 cheaper than Fable 5?
Not on list price, which is unchanged at $10 per million input tokens and $50 per million output tokens. The savings come from cache reads, which now cost 75% less at $0.25 per million. Anthropic estimates around 25% lower costs for typical workloads and up to around 45% for complex coding and highly agentic tasks, both vendor figures without published methodology.
Will I actually see the 25% reduction?
Only in proportion to how much of your token spend is cache reads. Workloads that repeatedly send large shared context, such as long system prompts or agent loops re-reading state, benefit most. A workload that does not use prompt caching sees no reduction at all, because nothing else about the pricing changed.
What changed about data retention?
Fable 5 required a 30-day retention period on all traffic, overriding existing zero-retention enterprise agreements. Anthropic now says that until Enterprise Frontier Safeguards is ready, “eligible customers will be able to use Fable 5.1 with zero data retention.” The announcement does not define eligibility or give a date for EFS, so treat it as an interim arrangement to confirm in writing.
How do I get Mythos 5.1?
Through verification rather than purchase. There is a Cyber Verification Program for defensive security work and a Life Sciences Verification Program for life sciences professionals. The Life Sciences program is described as currently available only to a set of US organizations.
Should I switch from Opus 5 to Fable 5.1?
Not on the evidence available. Opus 5 remains half the list price at $5 and $25 per million, and the July comparison had it matching or beating Fable 5 on much coding and knowledge work. Fable 5.1’s capability claims are Anthropic’s own, no head-to-head against Opus 5 has been published, and no independent benchmarks exist yet.
Is Fable 5 being retired?
No retirement was announced and no date was published. The `claude-fable-5` string is not described as deprecated, so there is no forced migration timeline at present.
How much of this is verified versus vendor-stated?
The pricing, the cache read reduction, the model relationship, the API string, the access programs and the retention language all come from Anthropic’s own announcement, so they are the company’s account of its own product. The cost-reduction percentages are explicitly estimates. All capability and benchmark framing is vendor positioning, with no third-party testing published at time of writing.
