Anthropic did not put out one model today. It put out one set of weights, twice, under two names, at two different levels of restraint. The framing Anthropic used for both was consistent and specific: a model that “sets a new standard on coding, knowledge work, and long-running problem-solving tasks,” language repeated closely enough across Anthropic’s own page and the coverage that followed it this morning that it reads as the deliberate headline of the release rather than incidental copy [1] [8]. Claude Fable 5.1 went into general release on the Claude API, Amazon Web Services, Google Cloud, and Microsoft Azure — the same four surfaces Anthropic’s own model-comparison table lists an ID for, claude-fable-5-1, right down the row [3]. Claude Mythos 5.1 went out the same day to a far smaller list: vetted cybersecurity professionals in Anthropic’s Cyber Verification Program and, through a newly opened Life Sciences Verification Program, vetted biomedical researchers — for now, US organizations only, with international expansion described as coordinated with government partners [1]. Anthropic’s own pricing table confirms the two are billed identically and, more tellingly, priced as a matched pair distinct from every other model on the list: Claude Fable 5.1 and Claude Mythos 5.1 are the only two rows in the current lineup [4]. That is the shape of the release, and everything else this article covers sits downstream of it.
One model, two safeguard settings
The headline framing, repeated across Anthropic’s announcement and in the trade coverage that followed it this morning, is architectural rather than marketing: Fable 5.1 and Mythos 5.1 are the same underlying model, differing only in which protective classifiers are switched on [1] [7]. Fable 5.1 ships with the standard safeguards Anthropic runs across its public models — cybersecurity, biology and chemistry, and anti-distillation protections all active. Mythos 5.1 has those specific restrictions lifted for its narrow, vetted user base, on the premise that a defensive security researcher or a credentialed life scientist needs the model to discuss and act on material the general public should not be routed toward without a filter in the way. Android Headlines’ description of the divergence tracks Anthropic’s own: “Fable 5.1 is generally available, while Mythos 5.1 is limited to its trusted-access programs for work in cybersecurity and the life sciences” [7].
This two-tier structure is not new with the .1 release — it is the second outing of a design Anthropic introduced in June with Fable 5 and Mythos 5, the “Mythos-class” tier Anthropic has positioned above the Opus class in its lineup, with Fable and Mythos taking their names from the Latin fabula and the Greek mythos, both meaning roughly “that which is told” [2]. At that original launch, Mythos 5 was restricted to Project Glasswing cybersecurity partners and a smaller trusted-access group of biomedical researchers [2]. The .1 release keeps that structure intact and adds the formal Life Sciences Verification Program as its own named on-ramp, with open enrollment for scientists described as coming soon [1]. The full mechanics of how that governance actually works — what a verification program checks, what a lifted safeguard changes about what the model will do, how Anthropic decided cybersecurity and life sciences specifically warranted this treatment — is its own subject, and this series covers it in the article that follows this one. What matters here is simpler: as of today, one of these two names is something almost any developer can call by API key, and the other is something almost none of them can.
What actually changed in the price
Anthropic did not touch the sticker price. Base input tokens on Fable 5.1 remain ten dollars per million, output tokens remain fifty dollars per million — the identical rate Fable 5 launched at in June, and the identical rate Mythos 5.1 carries alongside it on Anthropic’s current pricing table [4] [2]. What changed sits one column over. Anthropic’s pricing documentation specifies that a cache-read hit — pulling previously processed prompt content back out of cache rather than reprocessing it — costs 0.1x the base input price on every model in the current lineup except two: Fable 5.1 and Mythos 5.1, where the multiplier drops to 0.025x, working out to $0.25 per million tokens against the $1 per million tokens Fable 5 charged for the same cache hit [4]. That is a 75% cut to one specific, narrow line item, not a broad token-price reduction, and it is worth being precise about that distinction before repeating the headline savings figures Anthropic attaches to it: “reduce token-billed costs by about 25% for typical workloads and by as much as roughly 45% for highly agentic work,” a framing multiple outlets covering the launch repeated back with the same two numbers this morning [1] [5] [8].
Why cache reads specifically, and why does a discount on them move the needle by that much? Anthropic’s own pricing notes point at the mechanism: a workload that leans on prompt caching — a long system prompt reused across many turns, a large document held in context across a multi-step task, an agent replaying accumulated history on every step of a long-running job — spends a large share of its input tokens on cache hits rather than fresh tokens, precisely because caching exists to avoid reprocessing the same content twice [4]. An agentic pipeline that re-reads a growing conversation history dozens of times over the course of one task is, by construction, a cache-read-heavy workload; a single one-shot question is not. That is the shape of Anthropic’s own 25%-versus-45% split: ordinary use gets the smaller number, because a smaller share of its tokens were ever cache reads to begin with, and agentic use — the kind of long-running, tool-calling, context-reloading work Anthropic is explicitly positioning this release around — gets close to the larger one [1]. CryptoBriefing’s coverage of the launch landed on the same read: “Price reductions apply specifically to cached reads… input/output pricing remains unchanged from Fable 5” [9]. The economics of that cache-pricing mechanism in more depth, and how a developer should actually structure a prompt to capture it, is the API practitioner’s question this series answers separately; the launch-day fact is simply that the discount is real, it is specific, and it is the entire pricing story of this release.
Fable 5.1 also carries the full million-token context window and a 128,000-token maximum output that Anthropic’s model-comparison table lists for the current model generation generally — Fable 5.1, Opus 5, and Sonnet 5 all share that same 1M context and 128K output ceiling on the synchronous Messages API [3]. That figure has circulated in secondary coverage of this launch specifically as well; Anthropic’s own documentation table is the source used here rather than any outlet’s restatement of it.
The headline benchmark numbers
Anthropic led its announcement with a small set of benchmark results rather than a full scorecard, and this article follows that same restraint — the complete set of evaluations Anthropic published, and what each one is actually measuring, belongs to the deeper benchmark piece elsewhere in this series. Three numbers carried the most weight in Anthropic’s own framing and in this morning’s coverage. On Terminal-Bench 4.0, a coding-agent evaluation, Fable 5.1 scored 55.8%, against 42.0% for Fable 5 at its June launch — a jump of nearly fourteen points on the same test [1]. On Humanity’s Last Exam, a difficult multidisciplinary reasoning benchmark, Fable 5.1 reached 60.9% without external tools and 65.0% with tool use enabled [1]. And on GDPval-AA v2, Anthropic’s internal measure of knowledge-work quality, Fable 5.1 posted a score of 1853 [1]. MacRumors’ framing of Anthropic’s own claim captures how the company wants those numbers read together: Fable 5.1 “outperforms Fable 5, Opus 5, and OpenAI’s GPT–5.6 Sol across multiple benchmarks,” and specifically “achieves similar or better results than Fable 5 at low or medium effort, and has much higher performance at higher effort tiers” [5]. That second clause is doing real work: it is a claim that the model is not simply better when pushed to its maximum reasoning budget, but that it holds its ground against the previous generation even when a developer dials effort down to save on compute — a claim this series’ benchmark article is better placed to actually test than to simply repeat.
What customers said on day one
Anthropic’s launch materials leaned on named customer reaction more than most model releases do, and three quotes recurred across today’s coverage. Jane Street’s assessment, quoted directly by Anthropic, was pointed at both the previous flagship and the mid-tier model in one sentence: Fable 5.1 “solves more of our coding problems than Fable 5 or Opus 5, and achieves state of the art on trading intuition” [1]. Millennium offered a concrete anecdote rather than a superlative: the model reportedly found the root cause of a rare software crash that, per Anthropic’s account of the customer’s own words, “nobody on our team had explained in four to five years” — a claim about debugging depth rather than raw benchmark speed, and one MacRumors’ coverage repeated in similar terms [1] [5]. Every, the AI-native media company, characterized the release in a line built for repetition: “Fable-level intelligence, Opus-level price, Sonnet-speed,” with performance described as roughly twice as fast as Opus 5 [1]. Whether that price-and-speed framing survives contact with a real production workload, and what “Opus-level price” actually resolves to once cache pricing and effort settings are accounted for, is a question for the API practitioner article later in this series rather than for a launch-day account of what a customer said Anthropic could quote.
Where this leaves the Claude 5 family
Fable 5.1 does not replace anything below it in Anthropic’s current lineup so much as it re-anchors the top of it. Anthropic’s own model-comparison table lays the current family out side by side: Fable 5.1 at ten dollars per million input tokens and fifty per million output, Opus 5 at five and twenty-five, Sonnet 5 at two and ten, and Haiku 4.5 at one and five — a clean halving at each step down, with Fable 5.1 sitting alone at the top of that ladder exactly where Fable 5 sat before it [3]. The table’s own one-line guidance to developers choosing between them is blunt about when the top tier is actually warranted: reach for Opus 5 “for most workloads,” and reach for Fable 5.1 specifically “for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short” [3]. That is a narrower recommendation than “the best model, use it for everything,” and it matches the shape of Cognition’s reported reaction to the launch — the coding-agent company is described as having moved its own Opus 5 traffic inside Devin over to Fable 5.1 “on launch day,” specifically because the economics, not just the raw capability, had shifted enough to justify the switch [1].
Set against Fable 5 specifically, the .1 release reads less like a new model family and more like Anthropic tightening the same design: identical base pricing, identical two-tier safeguard structure, identical positioning above Opus 5 in the lineup, with the actual movement concentrated in three places — a real jump on coding and reasoning benchmarks, a real and specific cut to cache-read pricing that disproportionately rewards the agentic workloads Anthropic is most visibly courting with this release, and a formalized second on-ramp into the restricted Mythos tier for life sciences work that previously existed in a narrower form [1] [2] [4]. 9to5Mac’s own summary of the release lands on close to the same read, describing the update as delivering “around 60 percent fewer cybersecurity false positives” for Claude Code users specifically alongside the pricing change — a safety-tooling detail that belongs to this series’ safety-measures piece rather than to a launch overview, but one worth flagging as a sign that this release touches more surface area than pricing and benchmarks alone [6].
The model-comparison table also puts a date on how current this release actually is: Fable 5.1 carries a reliable knowledge cutoff of June 2026, a month later than Opus 5’s May 2026 and eight months ahead of Sonnet 5’s January 2026, and Anthropic’s retirement commitment for the model runs “not sooner than September 1, 2027” — a full year of guaranteed availability from launch day, on Anthropic-operated platforms specifically, matching the pattern Anthropic has set for the rest of the current generation [3]. None of that is a capability claim in itself, but it is the kind of detail that answers a question a team evaluating the release will ask before anything about benchmarks: how fresh is what this model actually knows, and how long can we build against this exact identifier before Anthropic requires a migration.
What is not in this article
Four things are deliberately absent above, each because another piece in this series is built to carry it properly rather than to summarize it in a paragraph here. How the Cyber Verification Program and Life Sciences Verification Program actually vet an applicant, and what specifically changes in the model’s behavior once a safeguard is lifted, is a governance question the next article in this series takes on directly. The complete benchmark scorecard Anthropic published — beyond the three headline figures used here — including how each evaluation is constructed and what it does and does not measure, belongs to this series’ science deep dive. The mechanics a developer actually needs to ship against this release — prompt-caching patterns that capture the new cache-read rate, the effort parameter’s behavior at each tier, the removal of forced tool-choice that a few outlets flagged as a breaking change worth testing before migrating [9] — is the API practitioner guide’s job. And the safety architecture behind both models, including the Enterprise Frontier Safeguards Anthropic says will begin rolling out this fall and the invisible text watermarking added for EU AI Act compliance, is this series’ safety-measures analysis [1].
What today actually was, stripped down to what Anthropic itself put on the record: one model, released under two names, at two different levels of restraint, with the public-facing half of it priced the same as its predecessor except for a specific and fairly large cut to what agentic workloads pay for reused context, and a small set of benchmark numbers Anthropic is confident enough in to lead with before the fuller evidence comes out. Everything downstream of that — whether the governance model holds up, whether the benchmarks generalize, whether the pricing change actually moves a real invoice, and whether the safety claims survive scrutiny — is what the rest of this series exists to check.
