Claude Fable 5 and Claude Mythos 5 are Anthropic's newest frontier models, announced on 9 June 2026. The short version, traced to Anthropic's launch announcement: they share the same underlying model, but Fable 5 is the version made safe for general use, while Mythos 5 has its safeguards lifted in specific areas and ships first through a restricted programme called Project Glasswing. If you build on the Claude API, the practical facts are the model ID claude-fable-5, pricing of $10 per million input tokens and $50 per million output tokens, a 1M token context window, up to 128k output tokens, and a safeguard layer that quietly falls back to Claude Opus 4.8 on certain restricted requests.
This blog is the plain-English starting point: what each model is, how they relate, what the safeguards do, and who should actually reach for them. If you are already running an older Claude model in production, keep the Claude Opus 4.8 API migration guide open alongside this one, and for the wider picture see the frontier model comparison and the main AI tools and AI development hub.
Quick answer: Fable 5 vs Mythos 5
The naming trips people up, so start here. Mythos 5 is the raw, more capable configuration of the model. Fable 5 is that same model wrapped in production safeguards so it can be offered for general use. You will use Fable 5 in almost every real scenario, because Mythos 5 is restricted to Project Glasswing rather than handed out as a normal API model. Think of Fable 5 as the model you can call today, and Mythos 5 as the same engine with the guard rails removed in narrow, controlled settings.
| Fact | Detail | Why it matters |
|---|---|---|
| Announced | 9 June 2026 | Fresh enough that guides should use dated wording and avoid stale assumptions. |
| API model ID | claude-fable-5 | This is the string you call for general-use access. Mythos 5 is restricted access. |
| Pricing | $10 input / $50 output per 1M tokens | Anthropic describes this as less than half the price of the Claude Mythos Preview. |
| Availability | Immediate on the Claude API and consumption-based plans | You can test it now rather than waiting for a staged rollout. |
| Subscription access | Included in subscription plans through 22 June, then requires usage credits | Plan budgets before the window closes if you rely on subscription quota. |
| Safeguard fallback | Restricted requests fall back to Claude Opus 4.8 | Some answers may come from Opus 4.8, not Fable 5, on sensitive topics. |
| Safeguard activation | Less than 5% of sessions, per Anthropic | For most everyday work you are talking to Fable 5 directly. |
Claude Fable 5 release facts
For the concrete details, Anthropic's developer documentation says Claude Fable 5 and Claude Mythos 5 both support a 1M token context window by default and up to 128k output tokens per request. The same docs list claude-fable-5 as the Claude API model ID and claude-mythos-5 as the restricted sibling model ID.
- Release date: 9 June 2026.
- Claude API ID:
claude-fable-5. - AWS Bedrock ID:
anthropic.claude-fable-5, with regional forms such asus.anthropic.claude-fable-5in Bedrock examples. - Context and output: 1M input context and 128k max output on the synchronous Messages API.
- Where it is available: Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry, and GitHub Copilot rollout surfaces.
- Data retention: Fable 5 is a covered Mythos-class model with 30-day retention, so it is not a zero-data-retention model.
What is Claude Fable 5?
Anthropic's models overview describes Claude Fable 5 as its most capable widely released model for demanding reasoning and long-horizon agentic work. That wording matters: Fable 5 is not just a chat upgrade, it is positioned for the kind of multi-step work where a model plans, uses tools, checks outputs, and keeps context over a long run.
Claude Fable 5 is the general-use member of the pair. Anthropic positions it as a Mythos-class model — meaning it carries the full capability of Mythos 5 — but with safeguards layered on so it is appropriate to expose broadly. In Anthropic's framing, Fable 5 reaches state-of-the-art performance on nearly all the benchmarks it was tested on, with particular strength in software engineering, knowledge work, vision, and scientific research.
For developers, the interesting part is not just a higher score. It is that Fable 5 is described as able to operate autonomously across millions of tokens with improved long-context behaviour. That points at the same class of work where stronger models earn their keep: reading across a large codebase, holding a plan over many steps, calling tools, reacting to output, and producing a result a human can review. If you already use Claude Code in VS Code, that agentic loop is where you would feel the difference first.
The announcement also calls out scientific reach that goes beyond typical assistant work: novel hypothesis generation in molecular biology, protein design and drug-discovery acceleration, and genomics research. Treat those as headline capabilities rather than turnkey features — they describe where the model is strong, not a promise that any single prompt will deliver a research-grade result without expert review.
What is Claude Mythos 5, and what is Project Glasswing?
Claude Mythos 5 is the same underlying model with its safeguards lifted in specific areas. Because of that, it is not offered as an open API model. Anthropic says it is initially deployed through Project Glasswing, a restricted programme rather than a public endpoint. In other words, Mythos 5 exists so that controlled, supervised settings can use the model's full range, while Fable 5 is the version the rest of us call.
The formal model docs list Mythos 5 as available through Project Glasswing and as the successor to Claude Mythos Preview. If you do not already have approved access through Anthropic, AWS, or Google Cloud account teams, the practical model to plan around is Fable 5, not Mythos 5.
The naming convention is worth internalising: Mythos is the capability tier, and Fable is the safe, general-use packaging of that tier. That is also why pricing is quoted against the earlier "Claude Mythos Preview" — the Mythos line is the capability lineage, and Fable 5 is the broadly available product built on it.
The three safeguards and the Opus 4.8 fallback
The most practically important detail for builders is how Fable 5 stays safe for general use. Anthropic describes three classifier systems. When any of them fires on a request, the model does not simply refuse — it falls back to Claude Opus 4.8 to handle that query instead. The three classifiers cover:
- Cybersecurity: blocks exploitation and offensive cyber tasks.
- Biology and chemistry: prevents bioweapon-related misuse.
- Distillation: prevents extraction of the model's capabilities into a copycat model.
Anthropic reports that these safeguards activate in less than 5% of sessions. That number matters for two reasons. First, for the overwhelming majority of everyday work, you are talking to Fable 5 directly. Second, if your application happens to live near one of those boundaries — say, defensive security tooling or life-sciences work — some responses may quietly come from Opus 4.8 rather than Fable 5. If consistency of model behaviour matters to you, that is something to test for explicitly rather than assume away.
Pricing and availability
Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. Anthropic frames that as less than half the price of the Claude Mythos Preview, so on a capability-per-dollar basis it is positioned as a meaningful improvement over the preview line rather than a price increase. It is available immediately on the Claude API and on consumption-based plans, which means you can run real evaluations today instead of waiting for a phased release.
There is one timing detail not to miss. Fable 5 is included in subscription plans through 22 June, after which continued use draws on usage credits. If your team relies on subscription quota rather than pay-as-you-go billing, plan your testing and any production cutover with that date in mind. To keep spend sensible once it is live, pair this with my AI API cost optimisation guide — route Fable 5 to the genuinely hard tasks and cache repeated context wherever you can.
How to access Claude Fable 5
On the Claude API, use claude-fable-5 wherever your SDK or Messages API call accepts a model string. Anthropic's docs list Fable 5 as generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. If you use AWS, the AWS launch post shows Bedrock examples using anthropic.claude-fable-5 and regional Bedrock IDs such as us.anthropic.claude-fable-5.
For GitHub Copilot, GitHub says Fable 5 is rolling out to Copilot Pro+, Max, Business, and Enterprise users across VS Code, Visual Studio, Copilot CLI, GitHub Copilot cloud agent, github.com, mobile clients, JetBrains, Xcode, and Eclipse. Business and Enterprise administrators must enable the Fable 5 policy, and GitHub notes that the policy is off by default.
Is Claude Fable 5 free?
Claude Fable 5 is not a free API model. It is a premium model priced at $10 per million input tokens and $50 per million output tokens. Anthropic included it in Pro, Max, Team, and seat-based Enterprise subscriptions through 22 June 2026, but said use after 23 June would require usage credits unless capacity allows an extension or later restoration to standard subscription access. For budgeting, treat the early subscription window as a trial period, not as the long-term price.
Data retention and enterprise caveats
The biggest operational caveat is retention. Anthropic's model docs designate Fable 5 and Mythos 5 as covered models with 30-day data retention, and GitHub's Copilot changelog says Fable 5 prompts and outputs may be retained for up to 30 days to operate Anthropic's safety classifiers. GitHub also says other Claude models in Copilot, including Opus 4.8, continue under zero data retention there. If your company has strict data-handling rules, this may decide whether Fable 5 can be used at all.
On AWS, the launch post says Bedrock users must opt into provider data sharing before invoking Fable 5, because Anthropic requires 30-day input and output retention for Mythos-class traffic. That does not make the model unusable, but it means regulated teams should route sensitive prompts carefully and keep Opus 4.8 or another ZDR-compatible model available where retention is not acceptable.
What the benchmarks suggest
Benchmark claims should always be read as direction, not gospel, but the ones Anthropic highlights are consistent with a model aimed at serious knowledge work. Fable 5 is reported to post the highest score on Hebbia's Finance Benchmark for senior-level reasoning, and top performance on Cognition's FrontierCode evaluation for software engineering. On vision, the announcement notes it can perform complex tasks — including reconstructing a web app from a screenshot — without additional scaffolding.
Early customer quotes point the same way: Stripe says testing "compressed months of engineering into days," Cursor's CEO calls it a state-of-the-art coding model, and Bloomberg describes it as the strongest finance-first model it has tested. Useful signals, but treat them as the vendor's launch evidence. The honest way to decide is still to run the model against your own real tasks, which is exactly what the test plan below is for.
Claude Fable 5 vs Opus 4.8
The simple rule is this: use Opus 4.8 as the default strong model, and use Fable 5 when the job is long, ambiguous, high-value, or needs more autonomous planning. Anthropic's model overview still points unsure users toward Opus 4.8 for complex tasks, then points to Fable 5 for the highest available capability. That is a useful default because Fable 5 costs twice as much as Opus 4.8 and has the retention caveat above.
| Use case | Better first choice | Why |
|---|---|---|
| Short coding question | Opus 4.8 | Strong enough and cheaper. |
| Long repo migration or multi-file bug | Fable 5 | Better fit for long-horizon planning and agentic coding. |
| High-volume summaries or rewrites | Cheaper model or Opus 4.8 | Fable 5's premium is usually wasted. |
| Defensive security or life-sciences work | Test both carefully | Fable 5 may fall back to Opus 4.8 near safeguard boundaries. |
| Strict zero-data-retention workload | Opus 4.8 or another ZDR-compatible path | Fable 5 requires 30-day retention. |
| Complex document, chart, or screenshot analysis | Fable 5 for hard cases | Vision and long-context work are part of the launch positioning. |
Who should care first
Fable 5 is a premium model, so the best use cases are the ones where a stronger answer saves enough time or risk to justify the token price. If a cheaper model already classifies a ticket or rewrites a paragraph well, keep using it. Reach for Fable 5 where judgement, planning, long context, tool use, or genuinely hard analysis change the outcome.
- Developers should test it on multi-file bugs, migration planning, code review, and repo-wide explanation, especially through Claude Code.
- AI app builders should test it on tool-calling workflows where the model must choose tools, handle failures, and keep a stable output shape — my tool calling guide is a good companion.
- Finance and research teams should test it on the long-context, senior-reasoning tasks the benchmarks highlight, with expert review kept firmly in the loop.
- Product and operations teams should test it on long documents, audit trails, and synthesis work where context size and coherence matter.
A practical test plan
The fastest way to decide whether Fable 5 belongs in your stack is to compare it against your current model on five real tasks — not toy prompts.
- Pick one multi-file coding task and compare Fable 5 with your current model on correctness and review time.
- Pick one long-context task, such as a repo audit or a document comparison, and judge coherence across the whole input.
- Pick one tool-calling workflow and measure whether it handles errors and keeps output shape stable.
- Pick one task that sits near a safeguard boundary and confirm whether responses come from Fable 5 or fall back to Opus 4.8.
- Record cost per successful completion, not just the token price.
What to test if you have access today
If Fable 5 is visible in your account, do not waste the first session on generic prompts. Use it where it has a chance to prove something Opus 4.8 cannot. A good first test is a real repository task: ask it to map the architecture, identify the riskiest files for a planned change, propose a staged implementation, then make the smallest useful edit. Measure how much human steering it needs, not just whether the final answer sounds impressive.
For a second test, give it a long document or codebase context and ask for decisions that require synthesis across distant sections. For a third, test a tool-calling workflow where failure handling matters: bad inputs, missing files, schema validation, retries, and final output formatting. Those are the tasks where a frontier model can save hours instead of merely writing a nicer paragraph.
When not to reach for Fable 5
A new flagship is not a reason to route everything through it. For small transformations, simple classification, short summaries, or high-volume background jobs, a cheaper model is usually the smarter default. Keep Fable 5 for the work where its strengths actually move the needle, and lean on routing rules so routine requests never touch a premium model. If you need a cheaper, broadly capable fallback for those routine paths, Opus 4.8 — the very model Fable 5 falls back to on restricted queries — is a sensible second tier.
FAQs
What is Claude Fable 5? A Mythos-class Claude model announced on 9 June 2026, made safe for general use, with the API model ID claude-fable-5 and safeguards that fall back to Claude Opus 4.8 on certain restricted queries.
How is Mythos 5 different from Fable 5? Same underlying model, but Mythos 5 has safeguards lifted in specific areas and ships through the restricted Project Glasswing programme rather than as a public API model.
What does Fable 5 cost? $10 per million input tokens and $50 per million output tokens — described by Anthropic as less than half the price of the Claude Mythos Preview.
What is the Claude Fable 5 API model ID? The Claude API model ID is claude-fable-5. On Amazon Bedrock, examples use anthropic.claude-fable-5 and regional IDs such as us.anthropic.claude-fable-5.
Is Claude Fable 5 free? No, not as an API model. It was included in some paid Claude subscription plans through 22 June 2026, but Anthropic said use after 23 June would require usage credits unless capacity allowed an extension.
How large is the Claude Fable 5 context window? Anthropic lists a 1M token context window and up to 128k output tokens per synchronous Messages API request.
Does Claude Fable 5 require data retention? Yes. Fable 5 and Mythos 5 are covered models with 30-day retention, so they are not zero-data-retention models. That matters for enterprise, legal, finance, healthcare, and regulated workflows.
Should I use Fable 5 or Opus 4.8? Use Opus 4.8 as the default strong model and Fable 5 for harder, longer, higher-value tasks where planning, tool use, vision, or long-context reasoning justify the higher price and retention trade-off.
Where should I read next? If you are migrating an app, read the Claude Opus 4.8 API migration guide; for the broader landscape, see the frontier model comparison.
Related guides on this site
Stay inside the Claude cluster to go from release context to hands-on workflows.
- Read the Claude Opus 4.8 explainer — the model Fable 5 falls back to, with its own pricing and use cases.
- Continue with the API migration guide when you are ready to update production code.
- Set up Claude Code in VS Code if you need the coding-agent foundation first.
- Review tool-calling patterns before building Fable-powered agents.
- Return to the AI hub for the full set of AI development guides.