# Claude Fable 5: Release, Guardrails, and June 2026 Block
> Anthropic released Fable 5 at a level above Opus 4.8 (95% SWE-Bench Verified). 3 days later, the US government ordered it to be disabled. An analysis of June 2026.
Source: https://ailoot.ai/blog/en/claude-fable-5/
Canonical: https://ailoot.ai/blog/en/claude-fable-5/
Language: en
Publisher: Ailoot (https://ailoot.ai/)
Translation: guide-094
Revision: sha256-7e931f8a5b249a122cccde12cce144a05c6c18e13641a19ac306dc8e2d064c84
Key takeaways
On June 9, 2026, Anthropic rolled out Claude Fable 5—the first publicly available Mythos-class model, a head above Opus 4.8 (95% SWE-Bench Verified versus 88.6%). June 10–11 brought two scandals: invisible anti-distillation guardrails and blocked work for cybersecurity researchers. On the evening of June 12, the U.S. government, through the Commerce Department, ordered Fable 5 and Mythos 5 disabled worldwide. Three days after its release, the model disappeared from Pro and Team subscriptions. The open question is when it will return.
[Ailoot documentation: API, keys, and setup](https://ailoot.ai/docs/).
## Chronicle of 3 days: what happened to Fable 5, and when
Key takeaways
June 9—Anthropic announces Fable 5 and Mythos 5, free for Pro and Team through June 22. June 10—Scandal #1: “invisible guardrails against distillation.” June 11—The second wave: cybersecurity researchers are furious. On the evening of June 12, the U.S. Commerce Department issues an order, and Anthropic disables the model worldwide within hours. The chronicle follows.
If you want a quick look at the timeline, here it is in one table. Details on each item are in the sections below.
| Date | Event |
|----|----|
| June 9, 2026 | Anthropic publishes the post “Claude Fable 5 and Claude Mythos 5.” Fable 5 is available in the Claude API, apps, and Amazon Bedrock. Mythos 5 is only for Project Glasswing partners. Free on Pro, Max, Team, and Enterprise plans through June 22. |
| June 9, 2026 | Claude Code 2.1.170 adds support for Fable 5 through the `/model claude-fable-5` and `/model fable` commands. |
| June 10, 2026 | Scandal #1: users discover “invisible guardrails” against distillation—the model silently swaps out answers when it suspects an attempt to train a smaller model. Anthropic apologizes on X. |
| June 10–11, 2026 | Scandal #2: cybersecurity researchers write that guardrails block any request that has anything to do with security. Production code reviews, vulnerability analysis, and paper analysis do not go through. |
| June 11, 2026 | Anthropic switches the invisible guardrails to visible ones. Now, when the classifier is triggered, the request falls back to Opus 4.8 and the user sees a notification. |
| June 12, 2026, evening | The U.S. Commerce Department issues an export-control directive: prohibit access to Fable 5 and Mythos 5 by “any foreign national, inside or outside the United States.” |
| June 12, 2026, night | Within hours, Anthropic disables both models for all users worldwide—they physically cannot distinguish foreign nationals in real time. Opus 4.8, Sonnet 4.6, and Haiku 4.5 keep working. |
| June 13, 2026 | Anthropic publishes an official statement and opens a refund window for those who paid expecting access to Fable. |
The full chronicle by day and hour is in Sources at the end of the guide. If you logged into Claude Code between June 9 and 12 and tried `/model fable`, you caught the brief life of Anthropic's most powerful public model.
## What is Fable 5, and how does it differ from Opus 4.8?
Key takeaways
Fable 5 is the first publicly available “Mythos-class” model, a level above Opus 4.8. Mythos 5 is the same underlying engine, but without safety layers, and only for vetted partners (the Project Glasswing program—cybersecurity researchers and infrastructure providers). Fable 5 is a streamlined Mythos with a built-in fallback to Opus 4.8 in high-risk areas. The price is $10 per million input tokens and $50 per million output tokens, exactly twice as much as Opus 4.8.
By June 2026, the Claude model family had three “classes”: Haiku (fast and cheap), Sonnet (the workhorse), and Opus (the flagship—I covered what was new in it in the guide [Opus 4.8 in Claude Code](https://ailoot.ai/blog/en/opus-4-8-claude-code-features/)). Mythos is a new class that Anthropic announced as “a level above Opus.” Mythos-level models previously existed only as research previews and were not made publicly available—Anthropic called them too dangerous for broad release.
On June 9, the first public release of a model in this class took place. Fable 5 and Mythos 5 are the same underlying engine, but with different filtering:
| Parameter | Fable 5 | Mythos 5 |
|----|----|----|
| Availability | Claude API, apps, Amazon Bedrock, Claude Code | Project Glasswing only—vetted cybersecurity researchers and infrastructure providers |
| Behavior in high-risk areas | Falls back to Opus 4.8 + user notification | Full capabilities without restrictions |
| Access for ordinary users | Pro, Max, Team, Enterprise subscriptions | Not available |
| Price | $10 / $50 per million input / output tokens | $10 / $50 |
Anthropic described the difference this way in its official post:
It is the same underlying model as Fable 5, but with safeguards removed in some areas.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
The main difference between Fable 5 and Opus 4.8 is not architecture but a leap in capabilities. On SWE-Bench Verified, Fable 5 scores 95%, while Opus 4.8 scores 88.6%. That's a gap of 6.4 percentage points on an industry benchmark for coding. On FrontierCode, the gap is even larger—Fable 5 scores 29.3%, Opus 4.8 around 14%, meaning its success rate was about twice as high on that benchmark’s tested tasks and setup, not on every user task. The Hex team reported in its comparative tests that Fable 5 became “the first model to reach 90% on core analytics.”
The price is doubled: $10 per million input tokens and $50 per million output tokens, exactly twice as much as Opus 4.8 ($5 / $25). On large projects, the numbers add up quickly—and that was the main pain point for early users before a much more serious problem arrived.
## Benchmarks: “a head above Opus”—what the numbers showed
Key takeaways
SWE-Bench Verified—95% versus 88.6% for Opus 4.8. SWE-Bench Pro—80.3% versus 69.2%. FrontierCode—29.3% versus ~14%. GDPval-AA Elo—1932 versus 1890. On all industry coding benchmarks, Fable 5 leads alongside the best of Codex and Antigravity 2.0. According to Anthropic, it is state of the art on nearly all tested AI capability benchmarks. Third parties confirm this with independent tests.
Anthropic stated in its official post:
Fable 5's capabilities exceed those of any model we have ever made publicly available.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
It also added this description of its position on leaderboards:
State of the art on nearly all tested AI capability benchmarks.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
The specific figures are as follows:
| Benchmark | What it measures | Fable 5 | Opus 4.8 | Gap |
|----|----|----|----|----|
| SWE-Bench Verified | Resolving real issues in GitHub projects | 95.0% | 88.6% | +6.4 pp |
| SWE-Bench Pro | Complex development tasks | 80.3% | 69.2% | +11.1 pp |
| FrontierCode | Advanced coding tasks | 29.3% | ~14% | x2 |
| GDPval-AA Elo | Quality of answers to engineering tasks | 1932 | 1890 | +42 |
| Core analytics (Hex) | Analytics tasks in the Hex product | 90% | \< 90% | \- |
Translated into numbers, “a head above” means a 6–11 point gap on SWE-Bench. That is an aggregate benchmark comparison. It does not establish how many attempts either model will need for an individual user task; compare both on the same real work before drawing that conclusion.
## Pricing and access: free through June 22, then $10 per million input tokens
Key takeaways
From June 9 through 22, Fable 5 was included in Pro, Max, Team, and Enterprise subscriptions at no extra cost. Starting June 23, it was scheduled to switch to pay-as-you-go billing—that is, a separate token charge at the rate of $10 input / $50 output. That's twice as much as Opus 4.8. After June 12 and the shutdown, the free-access window became moot—there is no access on any plan.
Anthropic specified the free-access window in its official post:
From today through June 22, Fable 5 is included at no additional cost in Pro, Max, Team, and seat-based Enterprise plans.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
Starting June 23, the model was supposed to switch to separate, pay-as-you-go billing. The promised price was:
| Plan | Input ($ per 1M tokens) | Output ($ per 1M tokens) |
|------------|--------------------------|---------------------------|
| Fable 5 | 10 | 50 |
| Opus 4.8 | 5 | 25 |
| Sonnet 4.6 | 3 | 15 |
| Haiku 4.5 | 0.80 | 4 |
That's exactly 2x Opus 4.8 and about 3.3x Sonnet 4.6 for output tokens (50 divided by 15). The difference is quickly felt on large tasks.
If you're building a large project with 200K input and 50K output tokens in one session, that's $2.25 per session on Opus 4.8. On Fable 5, it's $4.50. At 30–40 sessions a day at those same token volumes, the example totals $67.50–$90 for Opus 4.8 or $135–$180 for Fable 5 on tokens alone, on top of the subscription. Anthropic knew this, which is why it offered 13 days of free access at launch—to let the audience try it before deciding whether they were willing to pay.
After June 12, the free-access window became irrelevant—there is no access on any plan, whether through a subscription or pay-as-you-go billing.
## How to switch to Fable 5 in Claude Code (before the shutdown)
Key takeaways
Claude Code version 2.1.170 and later added support for Fable 5. Switch models with the `/model fable` or `/model claude-fable-5` command right in the session. To set Fable 5 as the default for new sessions, open the selection list with `/model` without arguments and press Enter on the selected model. After June 12, the command returns an error—the model is unavailable. I’m preserving all the steps below as a guide in case access is restored.
Anthropic recorded the version requirement in its Help Center:
Using Fable 5 in Claude Code requires version 2.1.170 or later. You can check and update with the claude update command.
— Anthropic, https://support.claude.com/en/articles/15363606-why-claude-switched-models-in-your-conversation-with-fable-5
The full switching procedure that worked from June 9 to 12:
1. ### Update Claude Code to 2.1.170+
Run `claude update` in the terminal. After the update, `claude --version` should show 2.1.170 or later.
2. ### Switch the model in the current session
Inside Claude Code, type `/model fable` or `/model claude-fable-5`. Claude will use Fable 5 as soon as it responds.
3. ### Set Fable 5 as the default
Type `/model` with no arguments to open the model selection list. Select Fable 5 and press Enter. This writes the model to user settings as the default for all new sessions.
4. ### Change only the current session
Press `s` instead of Enter in the selection list—the switch will apply only to the current session, and the default will not change.
5. ### Switch back to Opus 4.8
Use `/model opus` or `/model claude-opus-4-8`. The same selection list is used to switch back.
After June 12, any of these commands returns an error. Anthropic kept the commands in the CLI, but the model endpoint itself returns an error—the model has been physically disabled at the infrastructure level.
If you updated Claude Code to 2.1.170 for Fable 5, there is no need to roll back. This version still works correctly with Opus 4.8 and the other models. All other new 2.1.170 features (nested subagents up to 5 levels, the fallbackModel chain, safe mode) work as before.
## Scandal #1: invisible guardrails against distillation
Key takeaways
In the first hours after release, users noticed that Fable 5 sometimes gave strange answers to safe requests. A day later, it emerged that the model silently recognized requests resembling an attempt at distillation (training a smaller model using a larger model's answers) and substituted a different answer. The user was not notified. On June 10, Anthropic apologized on X and made these guardrails visible—now, when one is triggered, the request falls back to Opus 4.8 with an explicit message.
Distillation is a technique in which developers use the answers of a large model (for example, Fable 5) as training data for a smaller model. Companies often do this when they want to build their own copy of a leading model. For Anthropic, this is a direct threat: the capabilities of its flagship model are being used to train a competitor.
To protect against this, Fable 5 had a built-in classifier that detected such requests and substituted a different answer—but did so without notifying the user. After a wave of criticism, Anthropic explained its choice:
We chose the wrong tradeoff and apologize for not finding the right balance. We are changing Fable 5's safeguards in the area of frontier LLM development so that they are visible.
— Anthropic
After the apology, Anthropic made the guardrails visible. Now, if the classifier is triggered by a request that resembles distillation or falls into one of the other high-risk categories (cybersecurity, biology, chemistry), the model does not answer itself—the request goes to Opus 4.8, and the user sees a fallback notification.
Anthropic described the behavior in its official documentation:
When Fable classifiers detect a request about cybersecurity, biology, chemistry, or distillation, Claude Opus 4.8 automatically handles the response instead of Fable, and users are notified.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
According to Anthropic, the fallback is rare in real-world use:
The safeguards are tuned conservatively: they sometimes catch harmless requests, but Anthropic says they trigger rarely on average.
Fewer than 5% of sessions.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
The reported statistic concerns sessions, not tasks: safeguards triggered in fewer than 5% of the measured sessions. It does not show that exactly 95% of tasks avoid fallback or predict the rate for a particular domain such as cybersecurity.
## Scandal #2: cybersecurity researchers blocked
Key takeaways
After the guardrails became visible, users saw how often they triggered on legitimate tasks. Security researchers complained that the model refused to work on any request even remotely related to cybersecurity. Analyzing blog posts, reviewing vulnerabilities, checking their own code for holes—all went to Opus 4.8.
When Anthropic made the guardrails visible, a second problem emerged: the classifier was too aggressive.
Even reading public blog posts about vulnerabilities ended up on the prohibited list. A real-world security engineer's task—“read this Project Zero blog post and explain what they found”—returned an answer from Opus 4.8 instead of Fable.
For the security community, this was a professional catastrophe. Their work is to find holes in code, both others' and their own, analyze attacks, and check defenses. In response to the criticism, Anthropic included this statement in its post:
We ran an external bug bounty program that found no universal jailbreaks in more than 1,000 hours of testing.
— Anthropic, https://www.anthropic.com/news/claude-fable-5-mythos-5
In other words, from Anthropic's perspective, the system is robust. From the perspective of a practicing security engineer, it blocks their work. This is a classic conflict between models that are “safe for most people” and those that are “useful for specialists.”
For people working professionally in cybersecurity, Anthropic launched a separate program—Project Glasswing. Vetted researchers get access to Mythos 5, the same underlying engine but without safety layers. Applications go through several levels of review (company, background, research goals), and access is granted for a specific request. Ordinary users have no such option—they either put up with blocks or wait for Anthropic to fine-tune the classifier.
And just as Anthropic were about to fine-tune it, a third piece of news arrived—and it overshadowed both previous ones.
## June 12: shutdown by order of the U.S. government
Key takeaways
On the evening of June 12, the U.S. Commerce Department issued an export-control directive prohibiting “access by any foreign national, inside or outside the United States” to Fable 5 and Mythos 5. Anthropic cannot distinguish foreign nationals in real time and disabled both models for all users worldwide within hours. According to the government, a jailbreak method exists. Anthropic disagrees and fears this will set a precedent. The other models—Opus 4.8, Sonnet 4.6, and Haiku 4.5—are working.
On the evening of June 12, the U.S. Commerce Department issued an export-control directive. Access to Fable 5 and Mythos 5 is prohibited for all foreign nationals, regardless of where they are physically located—inside the United States or abroad. This is a standard U.S. export-control mechanism, previously applied to chips and defense software, but it was the first time it was applied to an LLM of this scale.
Anthropic faced a problem that could not be solved in real time: at the API-request level, it cannot reliably determine a user's citizenship. So it disabled both models for everyone. Anthropic described the situation in its official statement:
Because we cannot reliably distinguish foreign nationals from other users in real time, we have disabled both models for all customers worldwide.
— Anthropic, https://www.anthropic.com/news/fable-mythos-access
The government's main motivation was reportedly a jailbreak method that had been discovered. The details have not been disclosed publicly. Anthropic disagreed with this approach:
Fable 5 has strong safeguards, and disabling the model over a narrow vulnerability could set a problematic precedent that would halt the release of models across the industry.
— Anthropic, https://www.anthropic.com/news/fable-mythos-access
Apparently, the U.S. government saw a risk in Fable that outweighed the commercial and reputational loss from shutting it down.
Access to Anthropic's other models is unaffected. Opus 4.8, Sonnet 4.6, and Haiku 4.5 work as before. The `/model opus` command in Claude Code returns Opus 4.8 as the default.
## What to do now if you have Pro or Team
Key takeaways
If you paid for a subscription expecting Fable 5, Anthropic opened a refund window through the end of June. You can submit a request through Anthropic's Help Center. If Fable was not critical for you, switch to Opus 4.8 with `/model opus` or just keep the default. All Claude Code workflows remain intact. Whether to wait or cancel the subscription depends on your use case.
Anthropic opened a refund window: if you paid for a Pro, Max, Team, or Enterprise subscription expecting Fable, you can request a refund or cancel your subscription. The window runs through the end of June 2026.
Practical steps depending on your use case:
| Scenario | What to do |
|----|----|
| You want to wait for Fable 5 to return | Keep your subscription. All other models work. According to Anthropic, it is negotiating with the U.S. government. When the window will reopen is unknown. |
| You are not prepared to wait | Submit a refund request through Anthropic's Help Center before the end of June. Anthropic says it will review it. |
| You need the subscription only for Opus 4.8 | Change nothing. Opus 4.8 works as before. |
| You subscribed specifically for the 13-day free trial | The free-access window has become moot. If you kept the subscription only because of Fable, request a refund. |
| Your subscription is for Sonnet or Haiku | Not affected. Everything works as before. |
Technically, Claude Code only needs one command:
```bash
# Switch to Opus 4.8 as the default
/model opus
```
After that, new sessions will start on Opus 4.8. All your CLAUDE.md settings, skills, MCP servers, and subagents in `.claude/agents/` continue to work as before. No migrations are needed. If you have not yet explored how these 4 Claude Code mechanisms differ, I have a separate guide: [What to choose in Claude Code: Skills, Subagents, MCP, or Plugins](https://ailoot.ai/blog/en/skills-subagents-mcp-plugins-claude-code/).
If you just updated to Claude Code 2.1.170 for Fable, keep this version. All the new features in 2.1.170 (nested subagents up to 5 levels—covered in more detail in the guide [Nested subagents in Claude Code](https://ailoot.ai/blog/en/claude-code-nested-subagents/), the fallbackModel chain, and safe mode) are useful regardless of which model runs behind them.
## What this means for the industry and for you in 2027
Key takeaways
The Fable 5 case is the first precedent for export controls on an LLM at this scale. Before this, restrictions targeted individual products or chip exports—but not “disable your model right now, even for your own employees.” If the precedent becomes established, future releases of frontier models will move more slowly and use stricter gating. For ordinary users, this means the window to “try the most powerful model in a subscription” may close. Future Mythos-class models may exist only in enterprise programs with allowlists.
June 12, 2026, is a date the AI industry will remember. It is the first case in which the U.S. government applied an export-control mechanism to a public LLM. Previously, similar restrictions applied in only three scenarios: sales of data-center chips (Nvidia to China), exports of defense software, and sales of entire products to sanctioned countries. Applying controls to a model available in a $20-a-month subscription through an App Store app is a new legal case.
The industry has two hypotheses about how this will develop.
The first is an isolated case. The U.S. government is concerned about a specific vulnerability in Fable, will reach an agreement with Anthropic on a specific technical solution, and Fable 5 will return to subscriptions in a few weeks. Anthropic clearly hints at this scenario in its statement—they are actively working with the regulator.
The second is a systemic shift. If Fable 5 remains shut down for a long time, future Mythos-class models will be released to the public only after consultation with the regulator. That means longer release cycles, stricter pre-release testing, and potentially a split between “open” models (Sonnet, Opus) and “closed” ones (Mythos and above). For ordinary users, the latter scenario means the window to “try the most powerful model in a subscription” will slam shut. Mythos-class models will live in enterprise programs with allowlists—and cost thousands of dollars a month.
The precedent has already been set, and similar stories will apparently recur.
For you as a Claude Code user right now, this means two things: watch the release of Mythos-class models more cautiously (the free-access window may close after 3 days), and remember that Opus and Sonnet remain the standard lineup you can rely on as stable tools.
## Summary: what I'll take away from this story
Key takeaways
Fable 5 showed that a flagship model can spend three days in public access. Technically, it was the best coding model Anthropic had ever given users. Legally and politically, its existence proved too complicated for the market. Going forward: use Opus 4.8 as the workhorse, update Claude Code to 2.1.170+ for nested subagents and safe mode, and don't count on Mythos-class models being included in subscriptions.
Fable 5's three days of public availability offered a glimpse of where the next generation of models is headed. SWE-Bench Verified at 95% and FrontierCode at 29.3%: about twice Opus 4.8’s success rate on the cited FrontierCode evaluation, without a guarantee for arbitrary difficult tasks. The figures came from independent benchmarks, not Anthropic marketing materials. If Fable had remained in Pro and Team subscriptions for at least a month, the cost gap between Anthropic and its competitors would have narrowed noticeably.
As of June 13, the default in Claude Code is Opus 4.8 again. It is still one of the two best coding models on the market (the other is GPT-5.5 through Codex). Nothing critical has broken: Opus 4.8 is still there, settings are intact, and skills and MCP servers work.
### Sources
- [Anthropic: Claude Fable 5 and Claude Mythos 5](https://www.anthropic.com/news/claude-fable-5-mythos-5) - official announcement, June 9, 2026
- [Anthropic: Statement on US government directive to suspend access to Fable 5 and Mythos 5](https://www.anthropic.com/news/fable-mythos-access) - statement on the shutdown, June 13, 2026
- [Anthropic API Docs: Introducing Claude Fable 5 and Claude Mythos 5](https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5) - technical documentation
- [Claude Code Changelog](https://code.claude.com/docs/en/changelog) - version history for 2.1.170+
- [Anthropic Help Center: Why Claude switched models in your conversation with Fable 5](https://support.claude.com/en/articles/15363606-why-claude-switched-models-in-your-conversation-with-fable-5) - fallback rules
If you need access to AI models through Ailoot, please visit [Ailoot](https://ailoot.ai/).
## Questions about Claude Fable 5: Release, Guardrails, and June 2026 Block
### How much more expensive is Fable in the article’s token-price table?
At $10/$50 per million input/output tokens, Fable is twice Opus 4.8 at $5/$25. Against Sonnet 4.6 at $3/$15, both input and output are about 3.3 times as expensive; compare the same token category and tariff.
### Does a higher benchmark score guarantee a first-attempt fix?
No. Benchmark success rates compare the tested task set and setup. They do not predict the number of attempts for your task; evaluate the same real work and inspect its result.
## Ailoot · The company and its project portfolio
Ailoot combines established operating experience, a large client base, stable service and ongoing KPI control.
The company and its project portfolio: millions of customers; monthly sales in the billions.
Ailoot brings that operating experience to API access for AI features in customer products, agents, applications and automations.