Anthropic 发布 Claude Fable 5 和 Mythos 5，编程和科学能力大幅提升
2026-06-10 02:25·23天前·Matthias Bastian
AI 摘要
Anthropic 推出 Claude Fable 5 和 Mythos 5 两款新模型，声称全面超越当前 Opus 代，尤其在编程与科研方面。Fable 5 一天内完成 Stripe 的代码迁移，原需团队两个月工作量；Mythos 5 可自主设计候选药物，但因具备进攻性网络能力，目前仍保持封闭不开放。
原文 · 未翻译
Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science
Anthropic releases two new models in the fifth Claude generation. Claude Fable 5 claims the top spot in nearly all benchmarks, while Claude Mythos 5 (no longer in preview) is still only available to select partners.
Both models share the same base model. Fable 5 ships with conservative safety guardrails for general use. Mythos 5 drops those restrictions in areas like cybersecurity and is reserved for a small group of partners.
Anthropic says Fable 5 beats every generally available model the company has ever shipped and claims state-of-the-art results in nearly all benchmarks tested. The gap widens on long, complex tasks, the company states.
On SWE-Bench Pro, a benchmark for solving real software engineering tasks from public GitHub repos without help, Fable 5 hits 80.3 percent. Claude Opus 4.8 lands at 69.2 percent, GPT 5.5 at 58.6 percent, and Gemini 3.1 Pro at 54.2 percent.
On Cognition's FrontierCode benchmark, which tests demanding coding tasks under production standards, Fable 5 scores 29.3 percent. Claude Opus 4.8 manages 13.4 percent. GPT 5.5 gets just 5.7 percent.
Fable 5 is also more token-efficient than earlier Claude models, Anthropic claims. At medium effort, it posts the top score among all frontier models on FrontierCode. Payment processor Stripe says Fable compressed five months of engineering work into days. In a Ruby codebase with 50 million lines, the model finished a migration in one day that would have taken a full team over two months.
Knowledge work, vision, and long-term memory jump ahead
Fable 5 also tops the charts on complex analytical tasks, according to Anthropic. On Hebbia's Finance Benchmark, which tests AI reasoning at the level of seasoned financial analysts, it posted the highest score of any model, with gains in document-based reasoning and chart and table interpretation. Trading group IMC says Fable 5 passed their trading analysis evaluations almost across the board.
Anthropic 推出 Claude Fable 5 和 Mythos 5 两款新模型，声称全面超越当前 Opus 代，尤其在编程与科研方面。Fable 5 一天内完成 Stripe 的代码迁移，原需团队两个月工作量；Mythos 5 可自主设计候选药物，但因具备进攻性网络能力，目前仍保持封闭不开放。
原文 · 保持原样，未翻译
Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science
Anthropic releases two new models in the fifth Claude generation. Claude Fable 5 claims the top spot in nearly all benchmarks, while Claude Mythos 5 (no longer in preview) is still only available to select partners.
Both models share the same base model. Fable 5 ships with conservative safety guardrails for general use. Mythos 5 drops those restrictions in areas like cybersecurity and is reserved for a small group of partners.
Anthropic says Fable 5 beats every generally available model the company has ever shipped and claims state-of-the-art results in nearly all benchmarks tested. The gap widens on long, complex tasks, the company states.
On vision tasks, Fable 5 is the new state-of-the-art model, Anthropic says. It can pull precise figures from detailed scientific illustrations and rebuild a web app's source code from screenshots alone. As a demo, Fable 5 played through Pokemon FireRed using only game screenshots. Earlier models needed a complex helper framework with extra tools and access to additional game data like maps.
Anthropic says Fable 5 stays focused across millions of tokens and boosts its own results by taking notes. The company didn't share specific benchmarks here.
Drug design and autonomous genomics research
Anthropic's internal protein design experts say Mythos 5 sped up parts of the drug design process by 10x. In one test, the model, equipped with protein design and bioinformatics tools but no human help, matched or beat experienced human operators.
It handled every step a scientist would: it picked binding sites, launched and ran protein design tools, and fixed errors on its own. Nine out of 14 protein targets yielded strong drug design candidates, which are now being studied.
Anthropic's biggest claim is that Mythos 5 is the first model to consistently produce novel and convincing scientific hypotheses, something that's highly debated when it comes to current LLMs. In blinded comparisons, Anthropic's scientists preferred Mythos' molecular biology hypotheses over those from Opus-class models about 80 percent of the time. One hypothesis, a novel mechanism for an E. coli protein, was backed by an independent study, the company says.
In genomics, Mythos 5 worked largely on its own for over a week, Anthropic claims. The model compiled single-cell data for millions of cells from 138 animal species, then designed and trained its own machine learning model to identify cells with the same function across distantly related organisms. The result reportedly outperformed a model recently published in Science, despite being 100 times smaller. Anthropic plans to publish these results in the coming months.
Mythos 5 stays locked to cyber defenders for now
Claude Mythos 5 will continue to be offered through Project Glasswing in partnership with the US government. It replaces the earlier Claude Myth Preview. Anthropic calls it the world's strongest cybersecurity model. It scored 78 percent on the ExploitBench benchmark, up from 69 percent for Mythos Preview and 40 percent for Opus 4.8.
All current Mythos Preview users can upgrade to Mythos 5. Access will expand gradually in coordination with the US government. Anthropic is also planning a Trusted Access Program for biology. Select researchers will get access to Fable 5 without biology and chemistry safeguards, though cyber safeguards will stay in place.
Price doubles compared to Opus, and Fable gets kicked out of subscription plans
Both models cost $10 per million input tokens and $50 per million output tokens. Anthropic says that's less than half the price of Claude Mythos Preview, but it's much more than current Claude Opus models. How expensive Fable and Mythos actually are in practice depends on token consumption per task relative to cost per million tokens. On Claude.ai plans, the new models count as 2x usage, though it's not clear whether usage here maps directly to token burn.
Fable 5 is available now through the Claude API and usage-based Enterprise plans. Subscription plans (Pro, Max, Team, seat-based Enterprise) follow a staggered rollout. Until June 22, Fable 5 is included at no extra cost. Starting June 23, access requires usage credits. Down the line, Anthropic plans to fold the model back into regular subscription plans once it has enough capacity.
Opus fallback is supposed to keep dangerous prompts in check
Anthropic says Mythos-class models are powerful enough to pose real risks, in cyberattacks or bioweapons research, for example. Fable 5 uses new AI classifiers that flag dangerous requests and automatically route them to the weaker Claude Opus 4.8 model. Over 95 percent of sessions aren't affected.
The classifiers cover three areas: cybersecurity, biology and chemistry, and distillation, where third parties try to extract model capabilities, something all major Western AI companies suspect Chinese labs of doing.
In the web interface and apps, users get a notification when the system falls back to Opus 4.8. In the Messages API, the request is blocked by default, though developers can turn on a server-side fallback. In cyber tests, Fable 5 scored a zero percent success rate on offensive tasks. External testers couldn't find a universal jailbreak in over 1,000 hours, Anthropic says.
The company also built in another layer of protection that's invisible to users. Prompts aimed at building frontier LLMs, like setting up pretraining pipelines, distributed training infrastructure, or ML accelerator designs, aren't blocked outright. Instead, their effectiveness is limited through methods like prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT).
Anthropic admits the filters are too aggressive right now and can block harmless requests. The company is also adding a 30-day data retention period for all Mythos models to detect new attacks. More security-related details are in the system card and the risk report.
On SWE-Bench Pro, a benchmark for solving real software engineering tasks from public GitHub repos without help, Fable 5 hits 80.3 percent. Claude Opus 4.8 lands at 69.2 percent, GPT 5.5 at 58.6 percent, and Gemini 3.1 Pro at 54.2 percent.
On Cognition's FrontierCode benchmark, which tests demanding coding tasks under production standards, Fable 5 scores 29.3 percent. Claude Opus 4.8 manages 13.4 percent. GPT 5.5 gets just 5.7 percent.
Fable 5 is also more token-efficient than earlier Claude models, Anthropic claims. At medium effort, it posts the top score among all frontier models on FrontierCode. Payment processor Stripe says Fable compressed five months of engineering work into days. In a Ruby codebase with 50 million lines, the model finished a migration in one day that would have taken a full team over two months.
Knowledge work, vision, and long-term memory jump ahead
Fable 5 also tops the charts on complex analytical tasks, according to Anthropic. On Hebbia's Finance Benchmark, which tests AI reasoning at the level of seasoned financial analysts, it posted the highest score of any model, with gains in document-based reasoning and chart and table interpretation. Trading group IMC says Fable 5 passed their trading analysis evaluations almost across the board.
On vision tasks, Fable 5 is the new state-of-the-art model, Anthropic says. It can pull precise figures from detailed scientific illustrations and rebuild a web app's source code from screenshots alone. As a demo, Fable 5 played through Pokemon FireRed using only game screenshots. Earlier models needed a complex helper framework with extra tools and access to additional game data like maps.
Anthropic says Fable 5 stays focused across millions of tokens and boosts its own results by taking notes. The company didn't share specific benchmarks here.
Drug design and autonomous genomics research
Anthropic's internal protein design experts say Mythos 5 sped up parts of the drug design process by 10x. In one test, the model, equipped with protein design and bioinformatics tools but no human help, matched or beat experienced human operators.
It handled every step a scientist would: it picked binding sites, launched and ran protein design tools, and fixed errors on its own. Nine out of 14 protein targets yielded strong drug design candidates, which are now being studied.
Anthropic's biggest claim is that Mythos 5 is the first model to consistently produce novel and convincing scientific hypotheses, something that's highly debated when it comes to current LLMs. In blinded comparisons, Anthropic's scientists preferred Mythos' molecular biology hypotheses over those from Opus-class models about 80 percent of the time. One hypothesis, a novel mechanism for an E. coli protein, was backed by an independent study, the company says.
In genomics, Mythos 5 worked largely on its own for over a week, Anthropic claims. The model compiled single-cell data for millions of cells from 138 animal species, then designed and trained its own machine learning model to identify cells with the same function across distantly related organisms. The result reportedly outperformed a model recently published in Science, despite being 100 times smaller. Anthropic plans to publish these results in the coming months.
Mythos 5 stays locked to cyber defenders for now
Claude Mythos 5 will continue to be offered through Project Glasswing in partnership with the US government. It replaces the earlier Claude Myth Preview. Anthropic calls it the world's strongest cybersecurity model. It scored 78 percent on the ExploitBench benchmark, up from 69 percent for Mythos Preview and 40 percent for Opus 4.8.
All current Mythos Preview users can upgrade to Mythos 5. Access will expand gradually in coordination with the US government. Anthropic is also planning a Trusted Access Program for biology. Select researchers will get access to Fable 5 without biology and chemistry safeguards, though cyber safeguards will stay in place.
Price doubles compared to Opus, and Fable gets kicked out of subscription plans
Both models cost $10 per million input tokens and $50 per million output tokens. Anthropic says that's less than half the price of Claude Mythos Preview, but it's much more than current Claude Opus models. How expensive Fable and Mythos actually are in practice depends on token consumption per task relative to cost per million tokens. On Claude.ai plans, the new models count as 2x usage, though it's not clear whether usage here maps directly to token burn.
Fable 5 is available now through the Claude API and usage-based Enterprise plans. Subscription plans (Pro, Max, Team, seat-based Enterprise) follow a staggered rollout. Until June 22, Fable 5 is included at no extra cost. Starting June 23, access requires usage credits. Down the line, Anthropic plans to fold the model back into regular subscription plans once it has enough capacity.
Opus fallback is supposed to keep dangerous prompts in check
Anthropic says Mythos-class models are powerful enough to pose real risks, in cyberattacks or bioweapons research, for example. Fable 5 uses new AI classifiers that flag dangerous requests and automatically route them to the weaker Claude Opus 4.8 model. Over 95 percent of sessions aren't affected.
The classifiers cover three areas: cybersecurity, biology and chemistry, and distillation, where third parties try to extract model capabilities, something all major Western AI companies suspect Chinese labs of doing.
In the web interface and apps, users get a notification when the system falls back to Opus 4.8. In the Messages API, the request is blocked by default, though developers can turn on a server-side fallback. In cyber tests, Fable 5 scored a zero percent success rate on offensive tasks. External testers couldn't find a universal jailbreak in over 1,000 hours, Anthropic says.
The company also built in another layer of protection that's invisible to users. Prompts aimed at building frontier LLMs, like setting up pretraining pipelines, distributed training infrastructure, or ML accelerator designs, aren't blocked outright. Instead, their effectiveness is limited through methods like prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT).
Anthropic admits the filters are too aggressive right now and can block harmless requests. The company is also adding a 30-day data retention period for all Mythos models to detect new attacks. More security-related details are in the system card and the risk report.