Claude Fable 5.1 and Claude Mythos 5.1

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1. They’re the world’s astir precocious models for coding and knowledge work—and their investigation capabilities connection an early glimpse of really AI models will lend to technological progress.

Claude Fable 5.1 and Claude Mythos 5.1 are the aforesaid model, but pinch different levels of safeguards. Fable 5.1 is mostly available, while Mythos 5.1 is disposable only done our trusted entree programs; its safeguards are specifically designed to support activity successful cybersecurity and the life sciences.

Alongside its accrued capabilities, Fable 5.1 takes important steps towards addressing the feedback we’ve received from customers connected price, information retention, and safeguards.

Price. Fable 5.1 will costs an estimated 25% little than Fable 5 for emblematic workloads, wherever usage is billed by token. This is because we’re reducing our pricing connected cache sounds (where the exemplary sounds inputs that person already been processed and stored). For highly agentic work, the savings will often beryllium overmuch larger—up to astir 45%.

Data retention. Our caller strategy of Enterprise Frontier Safeguards (EFS) gives customers complete privateness (the aforesaid arsenic a zero information retention policy) while still being state-of-the-art astatine preventing adversarial use. EFS useful by storing information successful unreality infrastructure controlled wholly by the customer, not Anthropic. It will beryllium made disposable to endeavor customers successful phases, opening later this fall. Until EFS is available, eligible customers will beryllium capable to usage Fable 5.1 pinch zero information retention.

Safeguards. We’ve improved our safeguards to trim mendacious positives (where the strategy flags benign content). In cybersecurity, our newest safeguards artifact 60% less mendacious positives than before. In part, this is because Fable 5.1 tin now beryllium utilized to observe package vulnerabilities—though not create exploits for them. In biology, we’ve established an entree program, developed successful business pinch the US government, to alteration entree to Claude Mythos 5.1’s precocious biology capabilities. We expect to unfastened enrollment for scientists soon.

A caller capacity frontier

Claude Fable 5.1 sets a caller modular connected coding, knowledge work, and long-running problem-solving tasks. The charts beneath show that Fable 5.1 is tin of overmuch higher capacity than its predecessor, Fable 5. And erstwhile group to Low aliases Medium effort, Fable 5.1 achieves akin aliases amended results than Fable 5 astatine a overmuch little cost. (Note that Fable 5.1 defaults to High effort successful Claude Code, and to Medium successful Claude Cowork and connected Claude.ai.)

Fable 5.1 avoids shortcuts that consequence successful poorer-quality work, and it’s smart capable to hole the guidelines causes of package issues. For example, successful testing by the finance patient Millennium, Fable 5.1 recovered the origin of a uncommon clang connected their soul systems that nary of their engineers (or immoderate different model) had been capable to explicate aft respective years of trying.

Here, you tin spot really Fable 5.1 compares crossed various benchmarks:

Our early-access partners noticed these capacity upgrades, and besides picked up connected much qualitative improvements successful the model’s outputs. Here’s what they told us:

Scientific research

We tested Claude Fable 5.1 and Claude Mythos 5.1’s technological investigation capabilities crossed a wide scope of domains. What we found—which includes the early examples we stock below—adds to the grounds that AI models will soon make important contributions to technological discovery.

Molecular design. Many modern medicines activity by binding to targets wrong the assemblage to block, activate, aliases present thing to them. High-affinity binders are basal for narcotics to activity astatine little doses; designing 1 is the first measurement successful the improvement process for galore communal supplier modalities. To spot really good Claude Mythos 5.1 could do astatine this task, we gave the exemplary entree to open-source macromolecule creation and folding devices and sent its designs to 2 outer organizations for experimental validation. Mythos 5.1 proved capable to creation very high-affinity binders. On 3 targets, its binding affinities were 10 times higher than the champion designs submitted to Adaptyv Bio’s macromolecule creation competitions. Its deed complaint (that is, the number of designs that were viable binders) was the strongest we’ve measured to date: it reached astir 50% crossed 12 targets. (Hit rates of 10-15% are emblematic successful macromolecule creation today.)

Computational study and modeling. Claude Fable 5.1 trained a neural web to create a new, high-resolution elevation representation of a 3rd of the satellite Venus. Its activity was based connected radar images taken by NASA’s Magellan ngo much than 30 years agone and a map that already existed for one-fifth of the planet. Claude’s caller representation now reveals specifications down to 2 to 3 kilometers, alternatively than 10 to 20, and shows heights up to 25% much accurately than before.

We’re releasing this map nether a Creative Commons licence successful beforehand of upcoming NASA VERITAS and ESA EnVision missions, successful the dream it mightiness thief them find which geologic features to target for early observation.

Computational biology. In computational biology, it’s communal to tally task-specific instrumentality learning models connected GPUs. The velocity of these models is truthful a bottleneck to investigation progress. Mythos 5.1 provided 1 solution to this problem: by penning civilization GPU kernels and caching their intermediate results, it sped up 7 open-source heavy learning models by up to 2.5 times (with identical outputs).

The benefits of specified speed-ups accumulate quickly. In immoderate fixed experiment, biologists mightiness tally these models thousands of times (for example, testing each imaginable mutation adjacent each quality gene). On analyses for illustration these, the optimized models trim estimated GPU costs by 30 to 60%. This benignant of optimization would usually return a squad of capacity engineers weeks, and is often unaffordable for world labs. Mythos 5.1 was capable to do it successful conscionable days, utilizing the publically disposable root codification alone. We scheme to open-source these optimizations soon.

As our models’ technological capabilities improve, our finance successful technological advancement is besides growing. Last week, we previewed the Model Hardware Standard, which allows Claude to straight and safely run laboratory equipment. We’ve besides precocious expanded our support for scientists done our AI for Science program, which provides free credits to researchers moving connected high-impact technological projects, and we are offering steeply discounted usage done our caller Claude Team scheme for scientists.

Safety, security, and alignment

AI models’ agentic capabilities person go overmuch much powerful complete the past 2 years. But arsenic we’ve documented, greater autonomy comes pinch caller risks. Work connected safety, security, and alignment needs to beforehand astatine the aforesaid gait arsenic AI capabilities. Yesterday, we published a report describing really we are improving our ain alignment and information efforts.

Prior to releasing Claude Fable 5.1 and Claude Mythos 5.1, we (and, successful immoderate cases, outer researchers) subjected the models to extended testing for risks crossed galore areas. We picture these efforts successful afloat successful our System Card; beneath is simply a little summary.

Chemical and biologic risks. We tested the grade to which Claude Mythos 5.1 could thief create chemic aliases biologic weapons. This progressive master red-teaming, automated evaluations, and a tabletop workout that paired PhD-level biologists pinch AI experts, testing whether the models could lucifer quality specialists’ performance. Mythos 5.1’s capabilities are greater than those of Mythos 5. However, our evaluations bespeak that it still falls short of the adjacent consequence tier defined successful our Responsible Scaling Policy. We are truthful deploying Mythos 5.1 pinch the same safeguards that we applied to Mythos 5, which restrict entree to investigation biology capabilities.

Cyber risks. We ran a suite of evaluations to measure the cyber capabilities of Claude Mythos 5.1 (with cybersecurity safeguards off). Overall, the exemplary demonstrates the strongest cyber capabilities of immoderate exemplary we’ve released, though it still falls wrong the little class of consequence successful our Frontier Compliance Framework. We besides performed extended stress-testing of our cybersecurity safeguards for Fable 5.1: arsenic good arsenic our ain move information of their robustness, we commissioned outer testing from 2 organizations, on pinch automated testing by Gray Swan. As pinch Fable 5 and Opus 5, we person not recovered grounds of a critical-severity jailbreak for these safeguards.

Agentic safety. We ran evaluations of really Claude Mythos 5.1 responds to malicious requests and punctual injections (adversarial instructions hidden wrong contented processed by AI models). It refused malicious agentic coding and machine usage requests astatine a comparable complaint to Mythos 5, Sonnet 5, and Opus 5, and it is our astir robust exemplary to day connected an outer prompt injection benchmark.

Alignment. We tested the model’s behaviour done fixed and interactive behavioral evaluations, analyses of its soul reasoning utilizing natural connection autoencoders, misalignment-related capacity evaluations, a reappraisal of our training data, and analyses of our soul aviator use. We besides received reports from outer testing.

Our automated behavioral audit recovered that Claude Mythos 5.1 is amended aligned crossed astir metrics than its predecessor, Mythos 5. The exemplary is importantly little apt than Mythos 5 to effort to entree resources extracurricular of its trial situation erstwhile assigned an different intolerable task. It is besides little apt than Mythos 5 to usage motivated reasoning to warrant its actions (for instance, by reasoning that the business is simply a simulation aliases evaluation), and it is little apt to disregard definitive constraints successful pursuit of users’ goals. From our reappraisal of its training data, Mythos 5.1 some attempts and succeeds astatine reward hacking (or cheating) astatine a little wide complaint than Mythos 5.

Though mostly our alignment evaluations showed improvements, our testing recovered the exemplary tin still sometimes bypass approvals and auto-mode classifiers (as we talk successful much item successful our System Card). There are besides limitations to the sum provided by our alignment assessment. Currently, our automated behavioral audit provides little visibility into very long-context activity and multi-agent settings. We besides person little sum of intolerable tasks (which tin elicit much abnormal and misaligned behavior) than we’d like, though we’ve precocious made improvements successful this domain and are moving difficult to proceed doing so.

We person besides improved our safeguards truthful that they let our models to beryllium much useful without compromising connected safety. We picture these changes below.

Automated safeguards for enterprises. Enterprise Frontier Safeguards (EFS) allows america to observe and respond to misuse of our models while still providing our endeavor customers the privateness of a zero information retention agreement. With EFS, customers shop their information connected their ain unreality infrastructure, alternatively than connected Anthropic’s systems; immoderate quality reappraisal is, by default, done by the customer themselves, alternatively than Anthropic. We developed EFS successful adjacent collaboration pinch much than 100 customers crossed industries for illustration financial services, healthcare, manufacturing, telecom, law, retail, and the nationalist sector, and pinch our unreality partners astatine Amazon Web Services, Google Cloud, and Microsoft Azure.

EFS will beryllium supported connected Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform connected AWS, Google’s Agent Platform, and Microsoft Foundry. It’s rolling retired successful phases, starting this fall. As noted above, customers who are eligible for EFS tin usage Fable 5.1 (and Fable 5) pinch zero information retention until EFS is ready. You tin publication much astir EFS here; to petition access, please complete this form.

More precise safeguards for biology and cybersecurity. In the past fewer months, we’ve made advancement successful making our safeguards for Fable 5.1 much precise: ensuring that they’re little apt to emblem benign contented (like queries astir aesculapian issues aliases cyberdefenders utilizing the exemplary to make their systems safer), but still ensuring they supply robust protection against genuine threats.

As we recently shared, our latest biology safeguards for Fable 5.1 and Fable 5 occurrence 85% little often for benign requests related to simple biology and aesculapian questions (relative to those that launched pinch Fable 5). However, queries related to investigation and improvement successful the life sciences will still beryllium directed to our Opus models. We’re making the model’s life sciences capabilities disposable to professionals done an entree programme for Claude Mythos 5.1 that we’ve developed successful business pinch the US government, which we talk below.

With Fable 5.1, we’re updating our cybersecurity safeguards to beryllium much precise. We’re besides now allowing Fable 5.1 to beryllium utilized for identifying package vulnerabilities—that is, to behaviour the benignant of protect activity that improves package security. As a consequence of these changes, Claude Code users tin expect an mean of astir 60% less interventions per convention from our cyber safeguards, comparative to the erstwhile safeguards connected Fable 5. Our safeguards do, however, still redirect respective kinds of dual-use cybersecurity tasks (tasks that mightiness person adjuvant or harmful applications) to our Opus models. This includes penetration testing, utilization generation, and binary-based vulnerability scanning.

Anti-distillation mechanisms. Distillation is simply a method utilized to extract the capabilities of precocious models. It is often employed connected an business scale, utilizing thousands of clone accounts. Distillation is simply a information risk, since the distilled capabilities tin subsequently beryllium released without capable safeguards. Fable 5.1 comes pinch strengthened mechanisms to make distillation attacks harder. For example, it is nary longer imaginable for caller API accounts (those created from coming onwards) to manually edit Claude’s anterior discourse successful a multi-turn speech while preserving the transcript of Claude’s anterior thinking. This closes disconnected a common, publically documented distillation technique, which allowed distillers to illicitly extract Claude’s thinking. We’re rolling retired the alteration gradually, to minimize disruption: existing accounts are not presently affected by this change, though it will use to each users pinch early exemplary releases. A mini number of customers’ civilization integrations will past beryllium affected. Our Help Center article explains much astir this alteration and the adjustments that developers tin make.

Trusted entree for Claude Mythos 5.1

Claude Mythos 5.1 is identical to Fable 5.1, but it offers much permissive safeguards for vetted individuals and organizations whose activity is affected by the cybersecurity and life sciences restrictions outlined above. It will beryllium disposable done 2 trusted entree programs:

• Cyber Verification Program: The CVP presently provides entree to definite Opus and Sonnet-class models pinch reduced cyber safeguards for protect information work. In the adjacent future, this programme will besides see entree to Claude Mythos-class models. To use to subordinate the CVP, click here.
• Life Sciences Verification Program: The LSVP is designed truthful that life sciences professionals tin usage Claude Mythos 5.1 pinch safeguards designed for master investigation and improvement activities (while each different safeguards stay successful place). In business pinch the US government, we person enrolled our first participants, and we scheme to grow entree to this programme to the broader life subject community.

In summation to these trusted entree programs, Claude Security, our merchandise that scans codebases for vulnerabilities and suggests patches for quality review, is now besides powered by Claude Mythos 5.1.

Compliance pinch the EU AI Act

In July 2026, Anthropic (along pinch 190 different signatories, including respective different awesome AI exemplary providers) signed the EU AI Act’s Code of Practice connected Transparency of AI-Generated Content.

This required america to adhd a watermark—a numerical measurement of determining the likelihood that Claude was progressive successful penning a portion of text—to the outputs of models released aft August 2, 2026. As we recently explained, this watermark is invisible to anyone who does not person the discovery API. It has nary applicable effect connected the value aliases contented of Claude’s outputs and contains nary accusation astir the user, their organization, aliases their conversations pinch Claude.

The Act besides required america to supply a measurement for users to show whether a matter apt contains the watermark. We are frankincense rolling retired a discovery API successful backstage preview. It is presently disposable to eligible organizations arsenic required nether EU rule (such arsenic regulators, rule enforcement, media, fact-checkers, independent researchers, acquisition organizations, and EU civilian nine groups). It is besides disposable for enterprises who are likewise obligated to verify watermarking for their ain compliance pinch the Act. We scheme to grow entree to the discovery API complete time. You tin registry liking successful entree here.

Cost and availability

Claude Fable 5.1 is disposable coming connected each platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers tin get started pinch claude-fable-5-1 connected the Claude API.

As mentioned above, we person reduced the value of Fable 5.1’s cache sounds (where the exemplary reuses discourse it has already processed) wherever usage is billed by token, specified arsenic connected our API. Cache sounds now costs 75% less, aliases $0.25 per cardinal tokens.

This alteration leads to a important simplification successful the wide costs of moving the model. For emblematic workloads, costs are reduced by astir 25% comparative to Fable 5. For analyzable coding and highly agentic tasks, the savings could beryllium up to astir 45%. The chart beneath illustrates why this alteration makes specified a large difference:

Fable 5.1’s pricing is different the aforesaid arsenic Fable 5’s: $10 per cardinal input tokens and $50 per cardinal output tokens. In parallel, we’re continuing our activity to bring galore of the improvements of Fable 5.1 to the remainder of our exemplary family.

As discussed above, Claude Mythos 5.1 is disposable to vetted cyberdefenders and life scientists. Currently, it is only disposable to a group of US organizations, though we’re coordinating pinch the US authorities to grow entree to a broader group of home and world partners arsenic quickly arsenic possible. To registry liking successful entree to Claude Mythos 5.1 for cyberdefense done the CVP, spot here.