#Anthropic发布Fable5模型 A model, two names: Anthropic releases Fable 5 and Mythos 5, the most powerful models open to the public for the first time
June 9th, Anthropic simultaneously released two new models—Claude Fable 5 and Claude Mythos 5. Interestingly, they are actually based on the same underlying model, with the only difference being the "safety guard" tightness. This is the first time Anthropic's most capable model has been publicly available, and it’s also a rare "tiered release" experiment.
This article will help you understand: what makes it powerful, why it has two names, and how to use it now.
1. First, clarify: what is a "Mythos level" model?
Familiar with Claude? Readers know that Anthropic’s models are always divided into three tiers: Haiku (fast), Sonnet (balanced), Opus (most powerful). In April this year, Anthropic quietly introduced a new tier above Opus—Mythos level. The first Mythos-level model (Claude Mythos Preview) was not publicly sold but was released through a project called Project Glasswing, in cooperation with the U.S. government, and was only available to a select few cybersecurity agencies and critical infrastructure providers.
The reason is straightforward: this level of model is too powerful, capable of being misused in fields like cyberattacks and biological research, so Anthropic believes it should not be released directly.
The two models released this time are the official Mythos level versions:
◆ Claude Fable 5: an open-to-all version with a new set of safety guards;
◆ Claude Mythos 5: a version with some guards lifted, only available to trusted partners like Glasswing. Both are based on the same core model. The names are also carefully chosen: Fable comes from Latin fabula ("story told"), and shares roots with Greek mythos—same origin—differing only in "ability to speak freely."
2. How powerful is it?
According to Anthropic, Fable 5 surpasses any model the company has ever publicly released, and is state-of-the-art (SOTA) on nearly all AI benchmarks, especially excelling in software engineering, knowledge work, visual understanding, and scientific research. The longer and more complex the task, the greater its lead over older models.
Coding: compressing two months into one day. Payment company Stripe reported in early testing that Fable 5 completed a full migration of a 50 million line Ruby codebase in one day—something that would normally take a team over two months of manual work. In Cognition (Devin’s developer) FrontierCode evaluation, Fable 5 scored the highest among cutting-edge models and used fewer tokens.
Visual: beating Pokémon FireRed without assistance. Previous Claude models needed complex auxiliary tools (maps, navigation, game state info) to barely progress in Pokémon, but Fable 5 cleared the game just by looking at the raw game screen. It can also reconstruct source code for a web app directly from screenshots.
Memory: taking notes while playing Slay the Spire. In tasks with millions of tokens, Fable 5 can maintain focus and improve performance by using its own notes. In tests with the card game Slay the Spire, after being given persistent memory, its performance improved threefold over Opus 4.8.
Scientific research: speeding drug design by about 10 times. Internal protein design experts at Anthropic used Mythos 5 to accelerate parts of drug development processes by roughly tenfold—models independently completed tasks like selecting binding sites, running protein design tools, and recovering from failures, all without human assistance. Out of 14 protein targets, 9 yielded promising candidate drugs for further development. It also proposed a new mechanistic hypothesis about E. coli proteins, later confirmed by an independent research lab.
3. How does the safety guard work: triggers "downgrade" to Opus 4.8
This is the most intriguing aspect of the release. Fable 5 is equipped with a new set of classifiers—an AI system separate from the main model—designed to detect potential misuse. When the classifier detects that your request involves any of the following three sensitive areas, the response will no longer come from Fable 5 but will automatically switch to a less powerful model, Claude Opus 4.8 (with a prompt informing the user):
1. Cybersecurity: Mythos level models are highly capable of discovering and exploiting software vulnerabilities, and can "autonomous hacking"—reconnaissance and lateral movement in one package. These abilities significantly lower the cost of cyberattacks, hence the focus on defense.
2. Biology and chemistry: the model can now perform real-world scientific tasks (like predicting virus capsid assembly). Such dual-use capabilities are dangerous in malicious hands, so most requests in biotech are downgraded to Opus.
3. Distillation attacks: prevent large-scale "extraction" of Fable 5’s capabilities to train competing models.
Key figures: the guard triggers in less than 5% of sessions, meaning over 95% of users experience nearly the full capabilities of Mythos 5. Anthropic admits the guard is conservatively tuned and may produce false positives, and plans to reduce false alarms soon. Regarding jailbreaks, over 1,000 hours of external vulnerability bounty testing have found no "universal jailbreak."
Another notable change: all traffic from Mythos level models will be retained for 30 days (including enterprise clients) for defense against new attack methods and jailbreaks—Anthropic promises this data will not be used for training and has privacy protections like access audits.
4. Mythos 5: a "trusted circle" version without guards
Mythos 5 is essentially Fable 5 with some guards removed, claiming to have the strongest cybersecurity capabilities globally. Currently, it is only available through Project Glasswing for cybersecurity agencies, as an upgrade from Mythos Preview. Anthropic plans to gradually expand: negotiating with the U.S. government to increase partners; launching a more systematic trusted access program for cybersecurity organizations; and opening a biological research trusted plan (removing biotech guards but retaining cybersecurity ones) to accelerate new therapies. This "same model, tiered unlocking by identity" approach is a first in frontier AI releases.
5. Pricing and access: free API window until June 22
The API pricing during the free window: $10 per million tokens for input, $50 per million tokens for output—less than half of Mythos Preview. The model identifier is claude-fable-5, and the API and pay-as-you-go enterprise plans are available starting today. Subscribers should note:
◆ From now until June 22: Pro, Max, Team, and enterprise packages include Fable 5 for free;
◆ Starting June 23: it will be removed from packages, and continued use will require purchase of usage credits; if capacity allows, the free window may be extended;
◆ Afterward: as compute power catches up, Anthropic plans to re-integrate Fable 5 into standard subscription packages. In other words, the next two weeks are the best window to experience the most powerful model for free, so interested readers should seize the opportunity.
6. Final thoughts: a "new approach" to capability and safety
In recent years, frontier labs faced only two options for "what if the model is too strong": delay release or cut capabilities. Anthropic offers a third solution—full release with dynamic guards for the public version, while reserving the full version for trusted institutions.
This approach isn’t perfect: 5% false positives may frustrate some professional users (especially security researchers and biologists), and the 30-day data retention could raise privacy concerns. But it does allow the "most powerful capability" to not be locked behind government projects, instead entering everyone’s dialogue in a controlled way. When AI can complete two months of team effort in a day, or independently design protein drug candidates, "how to release" becomes as important as "what to build." The shared roots of Fable and Mythos may set a precedent for future frontier models’ release strategies.
June 9th, Anthropic simultaneously released two new models—Claude Fable 5 and Claude Mythos 5. Interestingly, they are actually based on the same underlying model, with the only difference being the "safety guard" tightness. This is the first time Anthropic's most capable model has been publicly available, and it’s also a rare "tiered release" experiment.
This article will help you understand: what makes it powerful, why it has two names, and how to use it now.
1. First, clarify: what is a "Mythos level" model?
Familiar with Claude? Readers know that Anthropic’s models are always divided into three tiers: Haiku (fast), Sonnet (balanced), Opus (most powerful). In April this year, Anthropic quietly introduced a new tier above Opus—Mythos level. The first Mythos-level model (Claude Mythos Preview) was not publicly sold but was released through a project called Project Glasswing, in cooperation with the U.S. government, and was only available to a select few cybersecurity agencies and critical infrastructure providers.
The reason is straightforward: this level of model is too powerful, capable of being misused in fields like cyberattacks and biological research, so Anthropic believes it should not be released directly.
The two models released this time are the official Mythos level versions:
◆ Claude Fable 5: an open-to-all version with a new set of safety guards;
◆ Claude Mythos 5: a version with some guards lifted, only available to trusted partners like Glasswing. Both are based on the same core model. The names are also carefully chosen: Fable comes from Latin fabula ("story told"), and shares roots with Greek mythos—same origin—differing only in "ability to speak freely."
2. How powerful is it?
According to Anthropic, Fable 5 surpasses any model the company has ever publicly released, and is state-of-the-art (SOTA) on nearly all AI benchmarks, especially excelling in software engineering, knowledge work, visual understanding, and scientific research. The longer and more complex the task, the greater its lead over older models.
Coding: compressing two months into one day. Payment company Stripe reported in early testing that Fable 5 completed a full migration of a 50 million line Ruby codebase in one day—something that would normally take a team over two months of manual work. In Cognition (Devin’s developer) FrontierCode evaluation, Fable 5 scored the highest among cutting-edge models and used fewer tokens.
Visual: beating Pokémon FireRed without assistance. Previous Claude models needed complex auxiliary tools (maps, navigation, game state info) to barely progress in Pokémon, but Fable 5 cleared the game just by looking at the raw game screen. It can also reconstruct source code for a web app directly from screenshots.
Memory: taking notes while playing Slay the Spire. In tasks with millions of tokens, Fable 5 can maintain focus and improve performance by using its own notes. In tests with the card game Slay the Spire, after being given persistent memory, its performance improved threefold over Opus 4.8.
Scientific research: speeding drug design by about 10 times. Internal protein design experts at Anthropic used Mythos 5 to accelerate parts of drug development processes by roughly tenfold—models independently completed tasks like selecting binding sites, running protein design tools, and recovering from failures, all without human assistance. Out of 14 protein targets, 9 yielded promising candidate drugs for further development. It also proposed a new mechanistic hypothesis about E. coli proteins, later confirmed by an independent research lab.
3. How does the safety guard work: triggers "downgrade" to Opus 4.8
This is the most intriguing aspect of the release. Fable 5 is equipped with a new set of classifiers—an AI system separate from the main model—designed to detect potential misuse. When the classifier detects that your request involves any of the following three sensitive areas, the response will no longer come from Fable 5 but will automatically switch to a less powerful model, Claude Opus 4.8 (with a prompt informing the user):
1. Cybersecurity: Mythos level models are highly capable of discovering and exploiting software vulnerabilities, and can "autonomous hacking"—reconnaissance and lateral movement in one package. These abilities significantly lower the cost of cyberattacks, hence the focus on defense.
2. Biology and chemistry: the model can now perform real-world scientific tasks (like predicting virus capsid assembly). Such dual-use capabilities are dangerous in malicious hands, so most requests in biotech are downgraded to Opus.
3. Distillation attacks: prevent large-scale "extraction" of Fable 5’s capabilities to train competing models.
Key figures: the guard triggers in less than 5% of sessions, meaning over 95% of users experience nearly the full capabilities of Mythos 5. Anthropic admits the guard is conservatively tuned and may produce false positives, and plans to reduce false alarms soon. Regarding jailbreaks, over 1,000 hours of external vulnerability bounty testing have found no "universal jailbreak."
Another notable change: all traffic from Mythos level models will be retained for 30 days (including enterprise clients) for defense against new attack methods and jailbreaks—Anthropic promises this data will not be used for training and has privacy protections like access audits.
4. Mythos 5: a "trusted circle" version without guards
Mythos 5 is essentially Fable 5 with some guards removed, claiming to have the strongest cybersecurity capabilities globally. Currently, it is only available through Project Glasswing for cybersecurity agencies, as an upgrade from Mythos Preview. Anthropic plans to gradually expand: negotiating with the U.S. government to increase partners; launching a more systematic trusted access program for cybersecurity organizations; and opening a biological research trusted plan (removing biotech guards but retaining cybersecurity ones) to accelerate new therapies. This "same model, tiered unlocking by identity" approach is a first in frontier AI releases.
5. Pricing and access: free API window until June 22
The API pricing during the free window: $10 per million tokens for input, $50 per million tokens for output—less than half of Mythos Preview. The model identifier is claude-fable-5, and the API and pay-as-you-go enterprise plans are available starting today. Subscribers should note:
◆ From now until June 22: Pro, Max, Team, and enterprise packages include Fable 5 for free;
◆ Starting June 23: it will be removed from packages, and continued use will require purchase of usage credits; if capacity allows, the free window may be extended;
◆ Afterward: as compute power catches up, Anthropic plans to re-integrate Fable 5 into standard subscription packages. In other words, the next two weeks are the best window to experience the most powerful model for free, so interested readers should seize the opportunity.
6. Final thoughts: a "new approach" to capability and safety
In recent years, frontier labs faced only two options for "what if the model is too strong": delay release or cut capabilities. Anthropic offers a third solution—full release with dynamic guards for the public version, while reserving the full version for trusted institutions.
This approach isn’t perfect: 5% false positives may frustrate some professional users (especially security researchers and biologists), and the 30-day data retention could raise privacy concerns. But it does allow the "most powerful capability" to not be locked behind government projects, instead entering everyone’s dialogue in a controlled way. When AI can complete two months of team effort in a day, or independently design protein drug candidates, "how to release" becomes as important as "what to build." The shared roots of Fable and Mythos may set a precedent for future frontier models’ release strategies.
