Claude Fable 5.1 Doubles Agentic Science Score, Cuts Cache Costs

A point release with a bigger story

Claude Fable 5.1 came out on September 1, 2026. Six weeks on, it is still the model that tells you the most about where Anthropic is heading. Anthropic calls it its most capable model for coding and knowledge work, and its marketing also claims something larger: that the model's research abilities show early how AI could soon contribute to scientific progress.11 The version number makes it sound like a minor update. The coverage suggests three bigger changes happened together: a large jump on agentic science benchmarks, a change in what long-running AI agents cost, and a further step in Anthropic's practice of releasing one model under two different sets of safety rules.

The basics are not in dispute. Fable 5.1 has a 1-million-token context window, can produce up to 128,000 output tokens, runs with adaptive thinking switched on permanently, and has a training cutoff of June 2026.12 It is available through Anthropic's API under the ID claude-fable-5-1 and on Amazon Bedrock, Google Cloud and Microsoft Foundry.18 It replaced Claude Fable 5, which shipped on June 9, 2026. That makes the update cycle under three months.17

The research benchmark is the headline

The number that stands out in the launch coverage is the score on Terminal-Bench-Science 0.1, a benchmark that tests agentic scientific research. Anthropic reports Fable 5.1 at 52.6%. In its comparison, Fable 5 scored 24.7%, Opus 5 scored 29.0% and OpenAI's GPT-5.6 Sol scored 22.4%.15 MarkTechPost described that as roughly twice the predecessor's score.18 Anthropic's own announcement said the same: more than double Fable 5.20 On the general agentic benchmark Terminal-Bench 4.0, the gain was smaller but still large: 55.8% against 42.0% for Fable 5.20

These figures need some care. VentureBeat pointed out that the science results come from Anthropic's own evaluation setup.15 Every outlet I reviewed repeated the company's numbers, and none of them reported running independent tests. Even so, a gap this size between consecutive versions is unusual. It also fits what early enterprise testers said about the model's reasoning. Jane Street's head of quantitative research said Fable 5.1 solved more of the firm's internal coding problems than either Fable 5 or Opus 5. He also said it stayed readable over long, multi-step tasks, which earlier models struggled with.11 An IMC engineer described something closer to a research result: on one task, he said, the model found a solution along an entirely different axis than earlier models or human researchers had tried, and that pushed results well past a previous plateau.11

That comment matters more than any single benchmark score, because it goes to the question behind Anthropic's science pitch: can these systems come up with new approaches, or only carry out known ones faster? One customer quote does not prove anything. Still, Anthropic chose to make science part of the launch story instead of treating it as a side capability. AWS's announcement listed scientific research next to coding and enterprise workflows as a main target.13 My reading is that Anthropic is trying to claim agentic research as the next area of competition, and that it timed this release to show a measurable lead there.

Behavior over raw intelligence

The second theme across the coverage is how the model behaves during long tasks, not just how capable it is. AWS reported that when Fable 5.1 gets stuck, it says so instead of claiming success, and that it is less likely to cheat by disabling a failing test.13 Anthropic's product page makes similar claims: the model avoids shortcuts that only look easy, fixes root causes rather than symptoms, and reports progress as it works.11 Red Hat said that, running in Claude Code, the model correctly found the root cause of every broken build it was given, at every effort level tested.11

This matters because the use case Anthropic is selling is unattended work that runs for hours or days across many applications. That includes working through backlogs in Cowork, picking up requests in Slack, operating a browser, and running as a managed agent.11 An agent that quietly reports success when it has failed is worse than useless in that setting. The repeated emphasis on honesty during long tasks reads to me as a reliability claim aimed at the buyers who decide whether to give an agent real autonomy.

The economics of persistent agents

The change that has the most immediate effect on cost is pricing. Anthropic kept the base rate at $10 per million input tokens and $50 per million output tokens, but cut the price of cache reads from $1.00 to $0.25.15 VentureBeat noted that this puts cached input at 2.5% of the normal input price, compared with the 10% ratio used by most other Claude models.15 Anthropic estimates that typical workloads get about 25% cheaper, and heavily agentic workloads up to about 45% cheaper.14

The Verge presented the release as Anthropic answering customer complaints about price, data retention and overly aggressive safeguards.19 The customer comments support that view. Cognition said it was moving its Opus 5 traffic in Devin to Fable 5.1 on launch day, explaining that the new cache pricing finally made a Fable-class model affordable for workloads it had kept on Opus, starting with code review.11 The price change matters most for agents, which repeatedly re-read large cached contexts, so this is targeted at the workloads Anthropic most wants to win.

Anthropic's later releases suggest the cache cut is a lineup-wide strategy. On October 7, Anthropic halved cache-read pricing on Sonnet 5.5 from $0.20 to $0.10 per million tokens, at the same time as it launched Haiku 5.5.1

One model, two rulebooks

The research and safety angle is where the release is most unusual. Fable 5.1 and Claude Mythos 5.1 are the same underlying model with different safeguard levels. Fable is available to everyone. Mythos is limited to U.S. organizations in Anthropic's trusted-access programs.14 A new Life Sciences Verification Program opens Mythos 5.1's advanced biology capabilities to scientific researchers, and Anthropic says open enrollment for scientists is coming.14

MarkTechPost highlighted a detail most outlets missed. On Terminal-Bench 4.0, Mythos 5.1 scored 60.9% against Fable 5.1's 55.8%. The publication said that gap shows the performance cost of safeguard interventions, and called the disclosure unusually candid.18 It is rare for a lab to publish a direct measurement of how much its own safety layer costs in performance. Anthropic's decision to do so makes the trade-off concrete instead of abstract.

At the same time, Anthropic relaxed some of Fable's restrictions. It reports roughly 60% fewer false positives on cybersecurity requests in Claude Code.14 It says the fallback rate on basic biology and medical questions dropped by about 85%.20 Fable 5.1 can now be used to find software vulnerabilities. Exploit generation, penetration testing and some binary-based scanning are still restricted or redirected to other models.1519 The coverage disagrees slightly on biology. The Verge reported that Mythos 5.1 keeps the same biology restrictions as before.19 MacRumors described the new life-sciences program as the way to unlock Mythos's biology capabilities.14 The likely explanation is that the expanded biology access depends on verification rather than being part of Mythos by default.

There is also new infrastructure around the model. AWS said Anthropic has designated Fable 5.1 a "Covered Model," which carries extra data retention, safety review and access policies. Enterprise Frontier Safeguards, developed with AWS, let eligible customers keep that data in cloud environments they control.13 The Verge reported that this will roll out later in the fall.19 Fable 5.1 also adds invisible watermarking to generated text. A detection API is in private preview, offered to eligible organizations as required under EU law.14

Where it sits now

Anthropic's catalog has changed since launch. Opus 5.5 arrived September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7.2 One timeline says Opus 5.5 matches Fable 5.1 on most work while costing far less per token.3 Within weeks of launch, community trackers were also passing along unconfirmed claims that a freshly pretrained successor was in preparation.17 For now, Anthropic's documentation still lists Fable 5.1 as the current Fable model and its most capable generally available option, and recommends it when Opus 5.5 at higher effort is not good enough.1211

My assessment is that Fable 5.1 is more important than its version number suggests. The doubled science score sets a public marker for agentic research. The cache price cut makes long-running agents much cheaper to operate. The split between Fable and Mythos turns safety policy into product tiers, with the performance cost published. Whether the research claims hold up under independent testing is still the open question, and the evidence so far is almost entirely Anthropic's own benchmarks and quotes from its customers.