The striking thing about Anthropic’s 9 June 2026 release was that two model names described two access policies for the same underlying system. Fable 5 was meant for general use; Mythos 5 was reserved for vetted work where stronger cyber or biology capability could be useful and risky.
That launch plan changed within three days. A US export control directive led Anthropic to suspend both models on 12 June. The controls were lifted later that month, and access returned from 1 July. Anthropic then released Fable 5.1 and Mythos 5.1 on 1 September. This guide explains the original split; our account of the US suspension follows the regulatory timeline.
At launch, Anthropic said Fable 5 and Mythos 5 came from the same Mythos-class system. The difference was the safeguards and who could use each version. The 5.1 successors follow the same broad pattern, though their pricing and safeguards changed.
That makes this release more interesting than a normal benchmark victory lap. Anthropic is trying to ship its highest-capability model to the public while carving out the parts it considers too risky for open access.
What were Fable 5 and Mythos 5 at launch?
Current status, checked 24 September 2026: Fable 5’s June outage ended on 1 July. Mythos 5 access resumed for approved US organisations. Fable 5.1 is now generally available; Mythos 5.1 remains limited to trusted access. The prices, platform rows and subscription dates below describe the 9 June launch terms, not an offer to rely on today.
- At launch, Anthropic presented Fable 5 as its most capable generally available model.
- At launch, Mythos 5 used the same underlying model with some safeguards lifted for vetted users.
- Fable 5 was announced for general access across several platforms on 9 June, then suspended on 12 June and redeployed from 1 July.
- Mythos 5 was limited to approved customers through Project Glasswing; restored access remained gated.
- Both models support a 1 million token context window and up to 128,000 output tokens per request.
- Both are priced at USD $10 per million input tokens and USD $50 per million output tokens.
- Fable 5 added safeguards for risky domains such as cybersecurity, biology and chemistry.
- Under the June safeguards, some Fable 5 requests could be refused or routed to Claude Opus 4.8.
- The June launch terms required 30-day data retention for both models; confirm current model and platform terms.
Fable 5 vs Mythos 5
| Area | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Public positioning | Widely released Mythos-class model | Restricted Mythos-class model |
| Underlying model | Same underlying model as Mythos 5 | Same underlying model as Fable 5 |
| Access | General availability | Limited to approved customers |
| Best fit | Long-running coding, research, knowledge work, enterprise workflows | Vetted defensive cybersecurity and selected scientific research |
| Safeguards | Extra classifiers and routing in high-risk domains | Some safeguards lifted for trusted programs |
| API model ID | claude-fable-5 | claude-mythos-5 |
| Context window | 1M tokens | 1M tokens |
| Max output | 128k tokens | 128k tokens |
| Price | $10 input / $50 output per million tokens | $10 input / $50 output per million tokens |
| Data retention | 30 days required | 30 days required |
In Anthropic’s 9 June announcement, the launch logic was clear: Fable was the public version of a Mythos-class model; Mythos was the trusted-access version for sensitive domains. That distinction remains useful when comparing the later 5.1 versions.
What Claude Fable 5 was designed for
Fable 5 is pitched at the jobs that usually break normal chat-model workflows: long projects, messy codebases, huge documents, multi-step analysis and agent runs that need to keep going without a human nudging every five minutes.
The model is meant for:
- large software migrations
- long-horizon autonomous coding
- complex implementation work
- deep research and analysis
- finance, legal and analytics workflows with dense documents
- visual work involving diagrams, charts, tables, PDFs and design comparison
- agent harnesses such as Claude Code and Claude Managed Agents
The part I would watch is not the benchmark table. It is the claim that Fable 5 can work for days in an agent harness, plan across stages, check its own progress and build tests or evaluations for its own work.
If that holds up in real use, this is less about a smarter chatbot and more about Anthropic pushing Claude further into project execution.
What Claude Mythos 5 was designed for
Mythos 5 was not intended as a model most teams could open in the console and start using. Its trusted-access design remains the important point of comparison.
It is restricted to approved customers, initially through Project Glasswing. That program is aimed at defensive work on critical software and infrastructure. Anthropic has also said it plans trusted access for selected biology researchers.
The reason is obvious from the system card. Anthropic says Mythos-class capability creates higher risk in areas such as cybersecurity, chemical and biological misuse, model distillation and advanced AI development assistance.
So Mythos 5 is the more direct version of the model, but it sits behind a trust gate.
Why Anthropic split Claude Fable 5 and Mythos 5
This is the important bit.
Anthropic wants to sell a model above Opus into mainstream coding and knowledge work, but it does not want to give every user the same level of help in high-risk technical domains.
Fable 5 is the compromise:
- You get the general reasoning, coding, long-context and agentic capability.
- Anthropic adds classifiers for sensitive areas.
- Some requests are refused.
- Some requests are routed to Opus 4.8 instead of Fable 5.
- Anthropic keeps 30 days of data for safety monitoring.
That is a fairly big product trade-off. For most business users, it may barely matter. For security researchers, bio researchers and developers building sensitive tooling, it matters a lot.
Fable 5 and Mythos 5 launch pricing and limits
| Item | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Input tokens | $10 / MTok | $10 / MTok |
| Output tokens | $50 / MTok | $50 / MTok |
| 5 minute cache write | $12.50 / MTok | $12.50 / MTok |
| 1 hour cache write | $20 / MTok | $20 / MTok |
| Cache hit / refresh | $1 / MTok | $1 / MTok |
| Context window | 1M tokens | 1M tokens |
| Maximum output | 128k tokens | 128k tokens |
At launch, the $10/$50 per-million-token price was double Anthropic’s listed Opus 4.8 rate. Those are historical launch prices. Anthropic says Fable 5.1 reduces typical workload cost, largely through lower cache-read pricing, so check current model and platform prices before budgeting.
One useful detail: Anthropic says the full 1M token context window is included at standard per-token pricing. A 900k-token request is not billed at a special long-context premium just because it is large, though normal token usage still applies.
Fable 5 and Mythos 5 availability: launch versus now
| Platform | Fable 5 at launch | Mythos 5 at launch |
|---|---|---|
| Claude API | Generally available | Limited access |
| Claude Platform on AWS | Generally available | Limited access |
| Amazon Bedrock | Generally available | Limited access |
| Vertex AI | Generally available | Limited access |
| Microsoft Foundry | Generally available | Not generally available |
| Claude subscription plans | Temporary launch access, then usage credits unless capacity changes | Not generally available |
At the June launch, Anthropic planned a staged subscription window: Fable 5 was included through 22 June on certain paid plans, with usage credits expected after that. The 12 June suspension interrupted that plan; Anthropic later announced a short included period after its July redeployment. These windows have passed, so use the current plan terms before making a buying decision.
The shifting windows show why launch-day availability is a weak foundation for a production plan. Confirm the model and billing route you can actually use, and keep a tested fallback.
Claude Fable 5 and Mythos 5 developer details
For builders, the headline specs are simple:
| Detail | Value |
|---|---|
| Fable API ID | claude-fable-5 |
| Mythos API ID | claude-mythos-5 |
| Context | 1M tokens |
| Max output | 128k tokens |
| Thinking mode | Adaptive thinking always on |
| Raw chain of thought | Not returned |
| Summarised thinking | Available through display settings |
| Data retention | 30 days required |
| Zero data retention | Not available |
A few practical points stand out.
First, adaptive thinking is always on. You cannot disable thinking on Fable 5 or Mythos 5. Developers can control effort, but the old mental model of switching thinking on or off does not apply here.
Second, raw chain-of-thought is not returned. If you want readable reasoning, you get summarised thinking rather than the model's private working.
Third, refusals are a normal API behaviour, not necessarily an error. For Fable 5, a refused request can return a successful HTTP 200 with stop_reason: "refusal". If you build around this model, you need to handle refusal and fallback paths properly.
The AWS wrinkle: data retention opt-in
AWS was one distribution route at launch, and its retention rules remain a procurement question. The terms have changed since June, so use the current Bedrock data-retention guide for an AWS deployment.
AWS currently lists Fable 5 and Fable 5.1 as requiring an explicit data-retention mode that allows human review. A zero-retention mode blocks requests to those models. The choice is set for the relevant AWS region or project, so an enterprise team should verify the exact configuration before testing.
Under the current AWS guide, prompts and outputs for these models may be retained within the AWS boundary for up to 30 days and reviewed by AWS. AWS says this content is not shared with the model provider, including under the legacy provider_data_share setting. That is a material change from the June launch wording and should be checked again before procurement.
If your organisation requires zero retention, check whether the model is available under that policy before putting it in a design. Retention is a gating technical condition, not a footnote to accept after development.
What the safeguards actually mean
Fable 5 is not just Mythos 5 with a different label. The public model has extra routing and refusal behaviour in domains Anthropic treats as high risk.
That includes:
- cybersecurity
- biology
- chemistry
- health-related high-risk prompts
- distillation attempts
- some frontier AI development assistance
For some categories, Fable 5 can route the request to Opus 4.8 instead. Anthropic says users are not charged Fable prices for rerouted requests, and API requests refused before output generation are not billed.
This is sensible, but it also means Fable 5 may feel inconsistent if your work sits near those boundaries. A normal product security task might be fine. A dual-use security request might trigger a different model or refusal. Developers need to test their actual workload, not just read the benchmark page.
The safety card is unusually blunt
The system card is worth reading because it makes the release feel less like marketing and more like a controlled exposure of a model Anthropic is not fully relaxed about.
A few points stand out:
- Anthropic evaluates Mythos 5 as its most capable model to date.
- It describes chemical and biological risk as higher and more uncertain than prior models.
- It says Mythos 5 can significantly assist production of non-novel chemical and biological weapons, while not crossing Anthropic's threshold for novel catastrophic biological weapon capability.
- It says Mythos 5 does not cross its automated AI R&D threshold.
- It still applies ASL-3 protections, including access controls, classifiers, threat intelligence, bug bounty work and model-weight security.
That is not normal SaaS release-note language. Anthropic is telling customers the model is useful enough to justify a new public tier, but risky enough that the unrestricted version cannot be broadly shipped.
Where this leaves Opus, Sonnet and Haiku
Fable 5 did not make the rest of the Claude family irrelevant at launch. The family has since changed again, including Opus 5 and the Fable/Mythos 5.1 successors. The table below captures the June positioning rather than a current model-selection ranking.
| Model family | Likely role after this release |
|---|---|
| Haiku | Fast, lower-cost work where latency matters more than frontier capability |
| Sonnet | Everyday coding, writing, analysis and production workflows |
| Opus | Complex synchronous work, strong reasoning and a lower price than Fable |
| Fable | Long-running, high-capability project execution with broader access |
| Mythos | Restricted high-risk research and defence work |
A lower-priced model can still be the better default for many workloads. Reserve Fable-class pricing and retention terms for tasks where a measured improvement justifies them, and compare the current model generation rather than relying on June’s lineup.
What changed with Fable 5.1 and Mythos 5.1
The June release drew a new line in the Claude stack. The July restoration and September 5.1 launch show that the line between general access and trusted access survived the short-lived interruption.
Fable 5 is the public face of Mythos-class capability. Mythos 5 is the restricted version for organisations Anthropic trusts with more dangerous capability. Same underlying model, different risk envelope.
That is both exciting and uncomfortable.
The attraction at launch was clear: longer agent runs, huge context, stronger coding and visual or document work were all areas where teams wanted progress. Those benefits still need testing against a specific workload.
The uncomfortable part is that the release only makes sense because the model is powerful enough to need a split personality. Anthropic is effectively saying: this is useful enough to ship widely, but not safely in every domain.
For builders, the practical advice is simple:
- Try Fable 5 on the hardest project work you already know how to verify.
- Do not start with sensitive workloads until you understand the retention and fallback behaviour.
- Test refusal and routing behaviour before putting it into production.
- Keep Opus, Sonnet or Haiku in the stack for cheaper and less sensitive jobs.
- Treat Mythos 5 as a trusted-access research model, not a general product option.
At launch, Fable 5 was the generally available face of Mythos-class capability, while Mythos 5 showed the access and safety constraints attached to more sensitive uses. Today, compare Fable 5.1 with the models actually available in your plan, and confirm retention and safeguards for your workload.
