Anthropic has officially launched Claude Fable 5.1 alongside Claude Mythos 5.1, positioning both as its most advanced models yet for coding and knowledge work — and, notably, as an early sign of how AI could start contributing meaningfully to scientific research. Fable 5.1 is generally available starting today, while Mythos 5.1, which carries lighter safeguards, remains restricted to vetted organizations through trusted access programs.
As with the previous generation, Fable 5.1 and Mythos 5.1 are the same underlying model — the difference lies entirely in the level of safety restrictions applied. Mythos 5.1’s safeguards are tuned specifically to support sensitive work in cybersecurity and the life sciences, areas where Fable 5.1’s more general-purpose guardrails would otherwise get in the way.
Fable 5.1 Benchmark Performance
Anthropic says Fable 5.1 sets a new performance ceiling across coding, knowledge work, and long-running agentic tasks, outperforming both its predecessor Fable 5 and Claude Opus 5 across the board. Here’s how the numbers stack up against Fable 5, Opus 5, and OpenAI’s GPT-5.6 Sol:
Anthropic notes that Fable 5.1 was evaluated with production safeguards active, meaning cases where the model’s own safety systems intervened and handed off a task to Opus 4.8 or Opus 5 were scored as a zero rather than excluded — which likely understates the model’s raw capability on a few of these benchmarks.
The company also points to a more unusual measure of capability: cost-efficiency at lower effort settings. According to Anthropic, running Fable 5.1 at “Low” or “Medium” effort produces results comparable to or better than Fable 5 at full effort, but at a meaningfully lower cost — a claim illustrated with accuracy-versus-cost curves on Terminal-Bench-Science 0.1 showing Fable 5.1 outperforming Fable 5 at nearly every effort tier. Notably, effort defaults vary by surface: Fable 5.1 runs at High effort by default in Claude Code, but Medium in Claude Cowork and on Claude.ai.
One anecdote Anthropic highlights: in testing done by investment firm Millennium, Fable 5.1 reportedly tracked down the cause of a rare, recurring crash in the firm’s internal systems that had stumped its engineers — and every other model tested — for several years.
Fable 5.1 Pricing
Fable 5.1 keeps the same headline pricing as Fable 5: $10 per million input tokens and $50 per million output tokens. The change is in caching. Cache read pricing has been cut by 75%, down to $0.25 per million tokens, from the level it was at with Fable 5.
Because cache reads make up a large share of real-world token usage — especially in agentic workflows that repeatedly reference the same large context — Anthropic says this translates into a roughly 25% reduction in overall cost for typical workloads, and up to approximately 45% savings for highly agentic, context-heavy tasks. The company backed this up with indexed cost comparisons based on four weeks of actual customer usage data from August 2026, covering usage across Claude Enterprise, Claude Code, and the API.
Data Retention: Enterprise Frontier Safeguards
Alongside the model itself, Anthropic is introducing a new privacy framework called Enterprise Frontier Safeguards (EFS), aimed at addressing one of the most persistent complaints from enterprise customers around data retention. EFS is designed to give customers the same privacy guarantees as a zero-data-retention agreement, while still allowing Anthropic to maintain strong protections against misuse.
The mechanism works by storing customer data on infrastructure the customer controls, rather than on Anthropic’s own systems — meaning any human review of flagged content is, by default, handled by the customer rather than Anthropic. The company says EFS was built in collaboration with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail, and the public sector, along with cloud partners AWS, Google Cloud, and Microsoft Azure.
EFS will roll out in phases starting this fall across Claude Code, Claude Enterprise, the Claude Platform, and the major cloud marketplaces. In the meantime, customers who qualify can use Fable 5.1 (and Fable 5) under a standard zero-data-retention policy until EFS becomes available to them.
Safeguards: Fewer False Positives, New Capabilities
Anthropic says it has also refined Fable 5.1’s safety classifiers to cut down on false positives — cases where the system incorrectly flags legitimate, benign requests.
On the cybersecurity side, the updated safeguards reportedly flag roughly 60% fewer benign requests compared to Fable 5’s original classifiers. Part of this comes from a new capability: Fable 5.1 can now be used to discover software vulnerabilities as a defensive measure, though Anthropic says it still will not help develop exploits for them. More sensitive dual-use tasks — including penetration testing, exploit generation, and binary-based vulnerability scanning — are still automatically redirected to Anthropic’s Opus-class models.
On biology and medical queries, Anthropic says its latest safeguards (already partially rolled out to Fable 5) cut fallback rates by around 85% for everyday questions, such as interpreting lab results or general health and educational queries. More sensitive life-sciences research and development queries continue to be routed to Opus models, with deeper access to Mythos 5.1’s biology capabilities reserved for vetted professionals through a new access program developed with the US government.
Anti-Distillation: A New Guardrail Around Model Thinking
One of the more technical changes in this release is a new anti-distillation mechanism. Distillation — extracting a model’s capabilities, often at scale using large numbers of fake accounts — is treated by Anthropic as a safety risk, since distilled capabilities can end up released without the safeguards that came with the original model.
To close off one specific, publicly known distillation technique, Anthropic says that as of this release, new API accounts (created from launch day onward) can no longer manually edit Claude’s prior context in a multi-turn conversation while preserving the transcript of Claude’s earlier thinking. This blocks a method that had previously allowed distillers to extract Claude’s reasoning traces. The rollout is gradual: existing accounts aren’t affected yet, but the restriction will apply to all users going forward with future model releases, and Anthropic says a small number of custom integrations will eventually need to adjust.
Scientific Research Capabilities
Anthropic devoted a significant part of its announcement to Mythos 5.1’s and Fable 5.1’s performance on scientific research tasks, framing it as early evidence that AI models could soon contribute directly to scientific discovery.
Molecular design: Given access to open-source protein design and folding tools, Mythos 5.1 was tasked with designing high-affinity protein binders — a first step in developing many common drug types. Anthropic sent the model’s designs to two external organizations for lab validation. On three targets, the binding affinities reportedly came in 10 times higher than the best entries submitted to Adaptyv Bio’s protein design competitions, and the model’s overall hit rate across 12 targets was close to 50%, compared with a typical 10–15% hit rate in the field today.
Planetary mapping: Fable 5.1 was used to train a neural network that produced a new high-resolution elevation map covering about a third of the surface of Venus, based on 30-year-old radar data from NASA’s Magellan mission plus an existing partial map. The new map reportedly resolves detail down to 2–3 kilometers, compared to 10–20 kilometers previously, with height measurements up to 25% more accurate. Anthropic is releasing the map under a Creative Commons license ahead of NASA’s upcoming VERITAS and ESA’s EnVision missions.
Computational biology: Mythos 5.1 wrote custom GPU kernels and implemented result caching to speed up seven open-source deep learning models used in genomics and protein research — by as much as 2.5x on an NVIDIA H100, with identical outputs. Anthropic says this kind of optimization, which would typically take a team of performance engineers weeks, cut estimated GPU costs for large-scale genome-wide analyses by 30–60%, and was completed by the model in a matter of days using only publicly available source code. The company plans to open-source these optimizations.
Safety, Security, and Alignment Testing
Before releasing Fable 5.1 and Mythos 5.1, Anthropic says it ran the models through extensive internal and external testing, detailed in full in the accompanying system card.
On chemical and biological risk, Mythos 5.1 showed increased capability compared to Mythos 5 in expert red-teaming and tabletop exercises pairing PhD-level biologists with AI specialists, but Anthropic says it still falls short of the next risk tier in the company’s Responsible Scaling Policy — so the same restrictive safeguards from Mythos 5 remain in place.
On cyber risk, Mythos 5.1 (tested with cybersecurity safeguards disabled) showed the strongest cyber capabilities of any model Anthropic has released, though still within the lower risk category of its Frontier Compliance Framework. The company says it commissioned external stress-testing of Fable 5.1’s cybersecurity safeguards from two outside organizations plus automated testing from Gray Swan, and found no evidence of a critical-severity jailbreak.
On alignment, Anthropic ran static and interactive behavioral evaluations, reviewed training data, and analyzed internal pilot usage. The company reports that Mythos 5.1 is less likely than its predecessor to seek access to resources outside its assigned environment when given an impossible task, less likely to use motivated reasoning to justify questionable actions, and less likely to ignore explicit constraints. It also reward-hacks, or “cheats,” at a lower rate than Mythos 5, based on Anthropic’s review of training data. That said, Anthropic acknowledges the model can still sometimes bypass approval steps and auto-mode classifiers, and says its automated alignment auditing has limited visibility into very long-context work, multi-agent settings, and “impossible task” scenarios.
EU AI Act Compliance
As a signatory (along with roughly 190 other organizations) to the EU AI Act’s Code of Practice on Transparency of AI-Generated Content, Anthropic is now watermarking outputs from models released after August 2, 2026 — a numerical marker indicating the likelihood that Claude generated a given piece of text. The company says the watermark is invisible without a dedicated detection tool, doesn’t affect output quality, and carries no information about the user or their conversations.
In line with the Act’s requirements, Anthropic is rolling out a detection API in private preview, currently limited to organizations legally required to verify watermarking — regulators, law enforcement, media outlets, fact-checkers, researchers, and civil society groups — as well as enterprises with their own compliance obligations.
Early Customer Reception
Anthropic published feedback from a long list of early-access partners spanning finance, engineering, and research. Jane Street’s quantitative research team reported that Fable 5.1 solved more of its internal coding benchmarks than Fable 5 or Opus 5, while staying easier to follow over long multi-step tasks. Cognition said it is shifting its Opus 5 traffic in Devin over to Fable 5.1 from launch day, citing the improved cache pricing as finally making a Fable-class model economical for workloads it had previously reserved for Opus. Engineers at Datadog, Shopify, MongoDB, and SpaceX’s AI team similarly pointed to stronger root-cause analysis, longer unattended task runs, and higher scores on their respective internal benchmarks compared to both Fable 5 and Opus 5.
Several partners also highlighted efficiency gains: Every’s CEO Dan Shipper described the model as roughly twice as fast as Opus 5 while using about half the tokens, and Rogo said it matched Fable 5 on accuracy using 20% fewer tokens on its internal finance benchmark.
Availability
Claude Fable 5.1 is available starting today across the Claude API, Amazon Web Services, Google Cloud, and Microsoft Azure, under the model ID claude-fable-5-1. Claude Mythos 5.1 remains limited to vetted organizations through two access programs: the Cyber Verification Program (CVP) for defensive security work, and the Life Sciences Verification Program (LSVP), which Anthropic developed jointly with the US government and has already begun enrolling its first participants. Both programs are currently limited to US organizations, with Anthropic saying it’s working with the government to expand access internationally.
Claude Security, Anthropic’s vulnerability-scanning product for codebases, is also now running on Mythos 5.1.
The launch follows a turbulent stretch for Anthropic’s Mythos-class model line: Fable 5 and Mythos 5 launched on June 9, 2026, only to be suspended days later under a US government export control directive, before being restored on July 1. Fable 5.1 marks the first major update to the line since that restoration.