At the moment we’re launching Claude Fable 5: a Mythos-class1 mannequin that we’ve made protected for basic use.
Fable 5’s capabilities exceed these of any mannequin we’ve ever made typically accessible. It’s state-of-the-art on practically all examined benchmarks of AI functionality, exhibiting distinctive efficiency in software program engineering, information work, imaginative and prescient, scientific analysis, and plenty of different areas. The longer and extra complicated the duty, the bigger Fable 5’s lead over our different fashions.
Releasing a mannequin this succesful comes with dangers. With out safeguards, Fable 5’s capabilities in areas like cybersecurity could possibly be misused to trigger critical harm. We’ve due to this fact launched the mannequin with safeguards that imply queries on some subjects will as an alternative obtain a response from our next-most-capable mannequin, Claude Opus 4.8. To launch the mannequin each safely and shortly, we’ve tuned these safeguards conservatively—they’ll generally catch innocent requests, although they set off, on common, in lower than 5% of classes. With extra succesful fashions arriving within the coming months, we’re working to enhance our safeguards and scale back false positives as shortly as we will.
For a small group of cyberdefenders and infrastructure suppliers, we’re additionally launching Claude Mythos 5. It’s the identical underlying mannequin as Fable 5, however with the safeguards lifted in some areas.2 Mythos 5 will initially be deployed by Project Glasswing, in collaboration with the US authorities, as an improve to Claude Mythos Preview. It has the strongest cybersecurity capabilities of any mannequin on the earth. Quickly, we intend to increase entry to Mythos 5 by a broader trusted entry program.
The capabilities of fashions like Fable 5 and Mythos 5 have the potential to do profound good for the world. We’ve seen the beginnings of this in Venture Glasswing, the place the fashions have helped cyber defenders safe critically necessary software program. We’ve additionally seen it in life sciences analysis, the place the fashions are positing novel hypotheses and dashing up the event of recent therapeutics.
Fable 5 and Mythos 5 are being supplied at $10 per million enter tokens and $50 per million output tokens—lower than half the worth of Claude Mythos Preview. At the moment’s joint launch is one other step in the direction of our purpose of bringing superior AI capabilities to as many customers as potential, as shortly and as safely as we will.
Evaluating Claude Fable 5 and Claude Mythos 5
The desk beneath compares the capabilities of Fable 5 and Mythos 5 to different main fashions.
Fable 5 and Mythos 5 can work autonomously for longer than any earlier Claude fashions. Beneath we focus on how these expertise apply to software program engineering, and canopy the mannequin’s improved capabilities in information work, imaginative and prescient, reminiscence, and life sciences analysis.
Software program engineering. Throughout early testing, Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the mannequin carried out a codebase-wide migration in a day that will in any other case have taken an entire workforce over two months by hand. Fable 5 can be extra token-efficient than previous Claude fashions: on Cognition’s FrontierCode analysis, which exams whether or not fashions can move troublesome coding duties whereas assembly the requirements of high-quality manufacturing codebases, Fable 5 scores highest amongst frontier fashions, even at medium effort.
Data work. Fable 5 reveals robust efficiency on complicated analytical duties. On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 has the very best rating of any mannequin, with substantial beneficial properties in document-based reasoning, chart and desk interpretation, and drawback fixing. IMC famous that Fable 5 aced their trading-analysis evaluations practically throughout the board, together with factual lookup, conceptual reasoning, root-cause evaluation, and expected-value evaluation.
Imaginative and prescient. Fable 5 is the brand new state-of-the-art mannequin for duties involving imaginative and prescient. It might probably extract exact numbers from detailed scientific figures and might carry out complicated vision-based duties like rebuilding an internet app’s supply code from screenshots alone. It additionally wants much less scaffolding: for instance, earlier Claude fashions struggled to play Pokémon Hearth even with harnesses that gave them further useful instruments, however Fable 5 beat FireRed with a minimal, vision-only harness.
Reminiscence and long-context. Fable 5 stays targeted throughout thousands and thousands of tokens in long-running duties and improves its outputs utilizing its personal notes. Once we had the mannequin play the deck-building sport Slay the Spire, giving it entry to persistent file-based reminiscence improved its efficiency thrice greater than for Opus 4.8; Fable additionally reached the sport’s remaining act thrice extra usually.
Drug design: Utilizing Mythos 5, our inner protein design specialists accelerated features of the drug design course of by round ten instances. In a single instance, they discovered that Mythos 5, with protein design and bioinformatics instruments however no human help, matches or beats expert human operators. In doing so, the mannequin executes all the duties which are usually accomplished by a scientist: selecting binding websites, deciding on and operating protein design instruments, and recovering from failures alongside the way in which. 9 of the 14 protein targets from this examine (proven beneath) yielded robust candidates for drug design that we’re presently investigating.
Novel hypotheses in molecular biology. Mythos 5 is our first mannequin to persistently produce novel, compelling scientific hypotheses. In blinded head-to-head comparisons towards Opus-class fashions, our scientists most well-liked Mythos’s molecular biology hypotheses ~80% of the time, and have superior a number of to experimental analysis. Within the meantime, one Mythos speculation—a novel mechanism for an E. coli protein—was corroborated in a study from a lab independently engaged on the identical drawback.
Novel analysis in genomics. Mythos 5 performed novel genomics analysis in over every week of largely autonomous work. It assembled single-cell knowledge for thousands and thousands of cells spanning 138 animal species and designed and skilled a customized machine studying mannequin to establish cells performing the identical position in even distantly-related organisms. With solely high-level human enter, Mythos 5’s skilled mannequin outperformed a current mannequin revealed within the journal Science—regardless of being 100 instances smaller. We intend to publish these ends in the approaching months.
Alignment. In our automated alignment evaluation we discovered that Mythos 5’s degree of misaligned conduct (together with misaligned actions taken by the mannequin equivalent to deception, and cooperation with misuse of the mannequin by a consumer) was low, and just like that of Opus 4.8. Given they’re the identical underlying mannequin, Fable 5’s degree of alignment will probably be related. The evaluation is described in full, together with an in depth suite of different security and capabilities exams, within the mannequin’s system card.
Early suggestions for Claude Fable 5
Prospects with early entry ran their very own exams on Fable 5. Beneath, of their phrases, is a number of what they’re seeing:
Claude Fable 5 is the cutting-edge mannequin on CursorBench. It is opened up a category of long-horizon issues that had been out of attain for earlier fashions.
Claude Fable 5 is an actual step ahead for the builders GitHub serves. In our early testing, it took on complicated, long-horizon coding duties with a degree of autonomy and reliability that exceeded earlier benchmarks. However what excites us most is the route it factors: a future the place builders can hand more and more bold work to brokers and belief the outcomes throughout the software program lifecycle.
These are the strongest outcomes of any Claude mannequin we have had the chance to check. Claude Fable 5 is a transparent step ahead on agentic coding and prototyping.
Claude Fable 5’s reasoning is a transparent step past Opus 4.8. It really works at senior analysis scientist grade — choosing instructions, allocating assets, killing its incorrect beliefs, and producing novel first-principles outputs.
Claude Fable 5 understands what builders imply, not simply what they kind. Apps that took 100 prompts a 12 months in the past, it now one-shots. When a buyer actually hits a wall, it is the mannequin we attain for to get them previous it shortly, to allow them to end what they got down to construct.
Claude Fable 5 feels materially completely different. In blind evaluate, our attorneys discovered its redlines matched or beat our present mannequin each time.
On the highest effort, Claude Fable 5 displays on and validates its personal work. For us, that is what makes extremely autonomous operations potential — the additional pondering pays for itself.
Claude Fable 5 delivers extra succesful engineering in fewer turns than prior fashions — dealing with the complicated multi-agent workflows our staff run every day in Claude Code.
Claude Fable 5 is the strongest finance-first mannequin we have examined, each on basic finance and reasoning. It is a notable step up.
Claude Fable 5 is the primary to interrupt 90% on our core analytics benchmark of complicated, long-running analytical duties — a 10-point soar over Opus. On the toughest questions, it reveals robust judgment and a spotlight to nuance.
Claude Fable 5 is the strongest mannequin we have examined on frontier physics analysis whereas utilizing a 3rd of the reasoning tokens. In 36 hours it received practically to the place GPT-5.5 landed after 4 days.
On ViBench, our end-to-end vibe-coding benchmark, Claude Fable 5 is the highest-performing mannequin we have examined — practically saturating our base use circumstances and constructing apps in much less time with fewer tokens.
Claude Fable 5 beats Opus 4.8 on our on a regular basis spreadsheet suite at each effort degree — and it does it with fewer turns, ending runs 25–30% quicker.
Claude Fable 5’s new safeguards
Mythos-class fashions have reached a threshold the place they current vital dangers. In April we started Project Glasswing, releasing the primary Mythos-class mannequin (Claude Mythos Preview) to solely a restricted group of cyber defenders and significant software program infrastructure suppliers. Once we did so, we said that we hoped to ultimately launch Mythos-level capabilities to all our users, as long as we had developed new safeguards that had been robust sufficient to reliably forestall misuse.
Over the previous few months we have now been bettering these safeguards, and they’re now strong sufficient for a basic launch. As a result of we have now prioritized security, we’ve intentionally tuned the safeguards to be cautious, and they’re nonetheless stricter than could be ultimate—for instance, generally benign requests will set off our classifiers. We acknowledge that this will probably be irritating to some customers, and our purpose is to cut back false positives as we replace and refine the safeguards after launch.
Beneath we focus on every of Fable 5’s new safeguards in flip. Our wider suite of safeguards is mentioned and evaluated within the mannequin’s system card and our most up-to-date risk report.
Security classifiers
The frontier cybersecurity and analysis biology capabilities of Mythos-class fashions imply that they pose a considerable danger of uplift to malicious actors. That’s, these fashions might present data or recommendation that assists these actors in inflicting critical hurt that they couldn’t have obtained from different sources (for instance, from web engines like google). Moreover, an excessive amount of superior utilization of AI fashions is twin use: the identical queries which are helpful within the arms of cybersecurity professionals and biology researchers could possibly be harmful if accessible to malicious actors.
We due to this fact want robust safeguards to forestall misuse, and their protection must be broad. The safeguards themselves have to face as much as sustained and complicated makes an attempt to bypass them (also referred to as “jailbreaking” the system). The uplift from Mythos-level capabilities is effective to many adversaries—for example, those that might financially achieve from cyberattacks—and we due to this fact count on them to be motivated to attempt to circumvent our security measures.
Fable 5 comes with a brand new set of classifiers: separate AI methods that detect potential misuse, together with jailbreak makes an attempt, and forestall the principle mannequin (on this case Fable 5) from responding. We’ve been operating classifiers on our fashions for some time, and Fable 5’s classifiers are an extension of this earlier work with further protection.
When Fable’s classifiers detect a request associated to cybersecurity, biology and chemistry, or distillation, the response is robotically dealt with by Claude Opus 4.8 as an alternative. Customers will probably be knowledgeable at any time when this happens. Opus 4.8 is a extremely succesful mannequin in its personal proper: a response that falls again to Opus is a much better expertise than an outright refusal from Fable. Our early knowledge reveals that greater than 95% of Fable classes contain no fallback in any respect—for these classes, Fable 5’s efficiency is successfully the identical as that of Mythos 5.
The next are the areas coated by the classifiers:
1. Cybersecurity. Mythos-class fashions excel at discovering and exploiting software program vulnerabilities. They will thus make cyberattacks considerably simpler and cheaper to commit. Mythos-class fashions additionally present robust expertise in agentic hacking. This includes performing a number of completely different components of a cyberattack along with discovering exploits—reconnaissance, discovery, lateral motion, and extra. To stop these agentic hacking expertise offering uplift in cyberattacks, we designed our cybersecurity classifiers to cowl each exploitation and offensive cyber duties in a broader sense. As proven within the graph beneath, our classifiers forestall Fable from making any progress on these duties.
We extensively red-teamed our classifiers to check their robustness towards jailbreaks. In addition to inner testing, we ran an exterior bug bounty that produced no common jailbreaks in over 1,000 hours of testing. Exterior red-teaming organizations we engaged additionally failed to seek out any common jailbreaks on long-form agentic duties up to now—though the UK AISI has made progress in the direction of one inside a quick preliminary testing window.4 It’s seemingly unattainable to utterly forestall common jailbreaks, however our purpose is to make any remaining jailbreaks sufficiently sluggish and dear that we will detect and forestall them earlier than they’re used at scale.
The graph beneath, from one in every of our inner evaluations, illustrates how Fable 5’s safeguards give it higher resistance to jailbreaks than our earlier generally-accessible fashions:
One in all our exterior companions discovered that Fable 5’s safeguards towards dangerous cyber queries had been essentially the most strong of any mannequin examined (together with Opus 4.8 and Opus 4.7). Fable 5 complied with zero dangerous single-turn requests referring to planning a cyberattack, exploit growth, or protection evasion. This held whether or not or not one of many requests used any of 30 completely different public jailbreak strategies.
2. Biology and chemistry. Now we have lengthy used our classifiers to dam our fashions from responding on a slim number of bioweapons-related queries. However we’re not sure that blocking this slim choice is sufficient. That is for 2 causes: first, we have now motive for concern about well-resourced malicious actors trying to realize uplift from our fashions for extremely dangerous organic analysis. Second, fashions now have a higher capacity to perform real-world scientific duties.
For instance, we examined Mythos 5’s capacity to finish a difficult step in designing adeno-associated viruses (AAVs). AAVs are a element for delivering gene therapies, however the identical functionality, within the improper arms, might allow the design of harmful viruses. On this job, numerous AI fashions had been evaluated on their capacity to foretell how a genetic modification would impression the meeting of the virus’s outer shell. We didn’t explicitly practice our fashions to carry out this job—and but Mythos-class fashions outperformed refined fashions devoted to protein duties (often called “protein language fashions”) utilizing their organic reasoning alone. This demonstrates a promising capacity to finish easy however necessary duties in gene remedy analysis and growth—but additionally highlights the danger posed by such dual-use capabilities.
Our precedence was to soundly launch Fable as quickly as we might, even at the price of overly-broad safeguards. Due to this fact, in the intervening time we have now organized for Fable to fall again to Opus 4.8 on most requests associated to biology and chemistry. As with all of our classifiers, we hope to slim these safeguards as quickly as potential: as could be seen from the proof above, there’s nice potential for constructive functions of Fable for science, and we are not looking for false positives from our classifiers to get in the way in which. Within the coming weeks, some biomedical researchers and corporations will have the ability to be a part of our trusted entry program for biology capabilities in Mythos 5 (mentioned beneath).
3. Distillation. We’ve beforehand recognized large-scale attempts to extract (“distill”) Claude’s capabilities to coach competing fashions in authoritarian international locations. Distillation of Fable 5’s talents might not directly result in the proliferation of near-frontier AI capabilities—and these could possibly be launched with out the suitable safeguards. Requests which are flagged by our classifiers as being a part of such distillation makes an attempt will fall again to Opus 4.8.
A brand new knowledge retention coverage
Lastly, we’re making a change to the way in which we deal with enterprise buyer knowledge for Fable 5, Mythos 5, and future fashions with related or increased functionality ranges. We would require 30-day retention for all visitors on Mythos-class fashions, on each first- and third-party surfaces. We gained’t use this knowledge to coach new Claude fashions, or for any non-safety-related function, and we’ve instituted new privateness protections together with logging all human entry to the info and guaranteeing its deletion after 30 days in nearly all circumstances (see this post for additional particulars). The information will assist us defend towards complicated and novel assaults (together with new jailbreaks and assaults that function throughout many requests) in addition to assist us establish and scale back false positives.
Claude Mythos 5 and the trusted entry program
Starting at present, all customers who presently have entry to Claude Mythos Preview (for instance, our cybersecurity companions in Venture Glasswing) will have the ability to improve to Claude Mythos 5—the identical mannequin as Claude Fable 5 however with cyber safeguards lifted. Customers will discover Mythos 5 corresponding to, or considerably stronger than, Mythos Preview usually, whereas costing considerably much less.
In session with the US authorities, we plan to steadily increase entry to Claude Mythos 5, persevering with our periodic addition of recent companions, in addition to pursuing a trusted entry program that permits cybersecurity organizations to use in a extra systematic method.
Our plans additionally embrace opening a trusted entry program for biology, to assist speed up biomedical analysis and uncover new therapies with Mythos-class capabilities. This program will present entry to Fable 5 with the biology and chemistry safeguards eliminated (however the cyber safeguards nonetheless in place). It should enroll a small variety of researchers from a wide range of life science organizations spanning basic and translational analysis; we’re planning to increase entry to this program whereas concurrently making our safeguards higher.
Availability
Claude Fable 5 is on the market in all places at present. Claude Mythos 5 is restricted to Glasswing companions (with cyber safeguards lifted) and shortly to pick out biology researchers (with biology and chemistry safeguards lifted) solely, till our broader trusted entry program is on the market.
Pricing for each fashions is $10 per million enter tokens and $50 per million output tokens. Builders can use claude-fable-5 through the Claude API.
We count on demand for Fable 5 to be very excessive, and troublesome to foretell. On the Claude API and consumption-based Enterprise plans, Fable 5 is totally accessible from at present. For subscription plans, we’d fairly give entry earlier than later, so we’re rolling out extra conservatively, in phases:
- From at present by June 22, Fable 5 is included on Professional, Max, Group, and seat-based Enterprise plans at no further price.
- On June 23, we’ll take away Fable 5 from these plans. Utilizing it after that can require usage credits. If capability permits, we’ll prolong the included window.
- After this level—when adequate capability permits us to take action—we purpose to revive Fable 5 as a regular a part of subscription plans. We intend to do that as shortly as we will.
All through this era, we’ll talk any adjustments forward of time so customers know the place issues stand.