Anthropic launched two models in 2026, on the 9th of June: Claude Fable 5, and Claude Mythos 5 These two are part of the same tier. “Mythos-class.” The Opus tier is the next tier up in terms of capability. Fable 5 has been deemed safe enough for everyday use. Mythos 5 – the same model but with certain safety safeguards weakened, and kept on limited release.
Claude Fable 5, Mythos 5.
Mythos class models are a tier above Claude-class models. The Opus-class models are superior in terms of capabilities. Project Glasswing published Claude Mythos Preview back in April.
Fable 5 shares the same model as Mythos 5. It’s the safety measures that make a difference. Safety classifiers are included in Fable 5, and can be used for all purposes. Mythos 5 still has a limited release, but some safety classifiers have been removed.
Names reflect the split. “Fable” Latin is the source of this word Fabula, “that which is told.” The Greek language is a good example. Mythos. They are distinguished by the safeguards, and so have different names.
The Anthropic Team calls Fable 5 the most advanced model that has been widely released. The model is designed for demanding reasoning, long-horizon work and agentic tasks. Fable 5, according to Anthropic, has more capabilities than any other model they have made available.
The default context for both models is 1M tokens. The models support up to 128k tokens of output per request. The price is $50 for every million output tokens, and $10 for each million input tokens. It is cheaper than Claude Mythos Preview.
The Capability Case
Fable 5 has been tested and found to be up-to-date on almost all benchmarks. The model shows excellent results for software engineering and knowledge work. It also performs well in vision research, scientific investigation, and scientific exploration. Its lead increases as the complexity and length of the task increase.
Stripe, a software engineer company, tested Fable 5 in early access. Model performed codebase wide migration on a Ruby codebase of 50,000,000 lines. Stripe claims that this migration took just one day. To do it manually, the team would need over two month.
Fable 5, too, is more efficient in terms of tokens than previous Claude-models. Fable 5, according to Cognition FrontierCode, is the best frontier model. Even at moderate effort, this holds true. It tests tough coding under the production codebase standard.
Anthropic’s Finance Benchmark is cited by Hebbia for reasoning at the senior level. Fable 5, the model with highest marks, is the best of them all. The gains are in chart interpretation and problem-solving, as well as document-based reasoning.
Anthropic describes Fable 5 as the latest state of the art in vision. The software can be used to get precise figures from complex scientific images. It can reconstruct a website’s code using only screenshots. The scaffolding is also less than previous models. Fable 5 beat Pokémon FireRed with a minimal, vision-only harness.
Fable 5 maintains focus on millions of tokens by using memory and context. The game improves on its own output using the notes it produces. The game The SpireOpus 4 helped three times more with persistent file-based memories.
Mythos 5, a new product, contains science-based claims. The internal protein designers accelerated drug development by ten to twenty times. Anthropic says Mythos 5 was its first model that consistently produced novel scientific hypotheses. In blinded tests, scientists preferred the molecular-biology hypotheses of Anthropic’s Mythos 5 model 80% out of 100 times.
Mythos 5 ran its own novel genomics research during a week-long, largely automated period. It developed a model using single-cell information from 138 species of animals. Anthropic claims that the model is superior to a model recently published in Scientists are able to explain the origins of science, despite being 100 times smaller.
What are the safeguards?
The release of such a powerful model is not without risk. Fable 5’s cyber-security capabilities can be abused to do serious harm if they are not protected. Anthropic launched Fable 5 using a brand new classifier set.
AI-classifiers work separately. They can detect possible misuse including attempts to jailbreak. These models prevent their main model from answering any flagged requests.
When Fable 5 flags a request for response, Claude Opus 4.8 handles the task. These areas include cybersecurity, biology and chemical engineering, and distillation. When a fallback happens, users will be notified.
Fable 5 is currently reverting to Opus 4.8.1 for most biology and chemistry requests. Anthropic has expressed concern about the possibility that dual-use queries might be used by malicious actors. The program will give approved Fable 5 users access to Fable 5 with no safeguards.
Anthropic tunes these safeguards cautiously. Sometimes they will catch innocent requests. They trigger on average in less that 5% of all sessions. According to Anthropic, more than 95 percent of Fable sessions have no fallback. Fable 5 is performing at a level that matches Mythos 5 for those sessions.
Anthropic heavily red-teamed the classifiers. In over 1,000 hour, an external bug bounty did not yield any universal jailbreaks. An external bug bounty produced no universal jailbreaks in over 1,000 hours. The UK AISI has made some progress in this area during a short testing period, according to Anthropic.
Mythos 5 has the same features as Mythos 4, but without cyber protection. Anthropic claims it to have the most robust cybersecurity capability of any existing model. Project Glasswing, a collaboration between the US Government and Anthropic, deploys this model.
Use Cases
This capability maps to a number of concrete workflows that technical teams can use:
- Large-scale code migrationThe Long-horizon Coding is ideal for big refactors or cross-repo Migrations. Stripe is a good example of this on a scale of 50,000,000 lines.
- The pipelines for agentic code generationThe agent can run more efficiently with fewer turns. GitHub has reported reliability and autonomy on long-term, complex coding tasks.
- Work in finance and analyticsStrong reasoning on charts and documents is ideal for financial analysis at the senior level. Hebbia, IMC and others cited improvements in tasks involving reasoning and trading analysis.
- Vision-to-code tasksThe front-end reconstruction is a good fit for the source reconstruction. This harness is designed to reduce tooling costs.
- Research agents that are long-runningThe persistent memory of millions of tokens is ideal for multi-day loops. Mythos 5 carried out novel genomics research over a period of autonomous, mostly self-directed work.
Comparison Table: Fable 5 vs. Mythos 5 vs. Opus 4.8
| Attribute | Claude Fable 5 | Claude Mythos 5 | Claude Opus 4.8 |
|---|---|---|---|
| Model tier | Mythos-class | Mythos-class | Class Opus |
| Model underlying | Same as Mythos 5 | The same as Fable 5. | Opus 4.8 |
| Available | Available to all | Limited (Project Glasswing). | Available to all |
| Safety Classifiers | Active (cyber, bio/chem, distillation) | Cyber safeguards lifted | Opus-level safeguards |
| Target fallback | Fallback to Opus 4. | No, this is not applicable | No, this is not applicable |
| API model ID | claude-fable-5 |
claude-mythos-5 |
Existing Opus ID |
| Context Window | 1M tokens default | 1M tokens default | Per Opus specs |
| Max. output | 128k tokens/request | 128k tokens/request | Per Opus specs |
| Input price (per 1M) | $10 | $10 | (per Opus pricing) |
| Output price (per 1M) | $50 | $50 | (per Opus pricing) |
| The Thinking Mode | Only adaptive, Always on | Only adaptive, Always on | Configurable |
| Data retention | 30-day (Covered Model) | 30-day (Covered Model) | Standard sizes |
Note: The Fable 5 launch source does not provide specific Opus 4.8 specifications and pricing. Table cells are marked accordingly.
What you need to know
- Fable 5 and Mythos 5 both share a common model. The only differences are in the safeguards.
- Fable 5 has been rated as the best in its class by Anthropic on almost all benchmarks.
- In less than 5% sessions, Fable 5’s classifiers revert to Opus version 4.8.
- Each model offers a window with a million tokens context at $10 in input and $50 output.
- Fable 5 can be found on most major platforms. Mythos 5 is limited to Project Glasswing.
| Src | Author/Post | Theme | Reach | Sentiment |
|---|
Check out the Technical details, Docs You can also find out more about the following: Sentiments Analytics from the AI Community. Also, feel free to follow us on Twitter Join our Facebook group! 150k+ ML SubReddit Subscribe Now our Newsletter. Wait! What? now you can join us on telegram as well.
Want to promote your GitHub repo, Hugging Face page, Product release or Webinar?? Connect with us

