# Claude Fable 5 Is Back: Pricing and Benchmarks

> Claude Fable 5 returns after a US export-control suspension. Here is the $10/$50 pricing, the SWE-Bench Pro and FrontierCode scores, and what it means.

Source: https://capitalandcompute.net/blog/claude-fable-5-returns-pricing-benchmarks/
Published: 2026-07-01
Updated: 2026-10-12

Key takeaways

* Claude Fable 5 is coming back online in early July 2026 after the US Commerce Department lifted the export-control directive that had forced it offline since June 12.
* The price is unchanged: $10 per million input tokens and $50 per million output tokens, the most expensive model Anthropic sells and twice the Opus 4.8 rate.
* On the coding benchmarks Anthropic published at launch, Fable 5 leads the frontier: 80.3% on SWE-Bench Pro against Opus 4.8 at 69.2%, per third-party transcriptions of the launch numbers.
* The suspension was the first known US use of export-control authority against a specific commercial frontier model, and the reason it happened is baked into how Fable 5 is built.
* The jailbreak came from an Amazon security report: prompting Fable 5 to identify software vulnerabilities and, in one case, show how to exploit one. Anthropic found every model it tested, including Opus 4.8, GPT-5.5, and Kimi K2.7, could do the same, and shipped a classifier that blocks the specific technique in over 99% of cases.
* Anthropic, Amazon, Microsoft, and Google are adopting a shared framework for grading jailbreak severity, and Anthropic committed to giving the US government pre-release access to evaluate future national-security-relevant models.

Claude Fable 5, the most capable model Anthropic has ever put on general sale, spent most of June 2026 switched off by order of the US government. On June 30 the Commerce Department lifted the directive, and Anthropic said it would begin restoring access, per [CNBC’s reporting on the reversal](https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html). For anyone who was mid-migration when the model vanished, the practical questions are the same as they were on launch day: what it costs, how good it actually is, and whether the top-tier price is worth paying.

Here is the whole arc, the pricing that did not change while the model was dark, the benchmark case, and the reason a government pulled it in the first place.

## What happened: the two weeks Fable 5 was offline

Anthropic launched Fable 5 and its no-classifier sibling Mythos 5 on June 9, 2026\. Three days later the US government intervened. In [Anthropic’s own statement on the directive](https://x.com/AnthropicAI/status/2065597531644743999), the government “issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” Rather than try to fence off foreign users, Anthropic disabled both models for everyone, as [Fortune reported at the time](https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/).

The trigger was a jailbreak, and Anthropic’s own account of it, published in [Anthropic’s redeployment post](https://www.anthropic.com/news/redeploying-fable-5), is more specific than the early reporting: Amazon security researchers found they could bypass Fable 5’s safeguards by “prompting it so that it identified a number of software vulnerabilities,” and “in one case, the model produced code demonstrating how the relevant vulnerability could be exploited.” Those safeguards are the classifiers meant to keep general users away from the cybersecurity abilities of the Mythos-class models. Controls on Mythos 5 eased first, for a short list of approved US critical-infrastructure partners, before the full reversal at the end of the month. Commerce Secretary Howard Lutnick framed the resolution as cooperation, writing that “over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5 to ensure alignment across the US Government and strengthen America’s leadership in AI,” in [a statement posted to X](https://x.com/howardlutnick/status/2072100729603452965) and reported by [Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says).

1. June 9, 2026  
Fable 5 and Mythos 5 launch  
Anthropic's most capable widely released model ships at $10 / $50 per million tokens, generally available on the Claude API, AWS, Bedrock, Google Cloud, and Microsoft Foundry.
2. June 12, 2026  
US export-control directive  
The government orders access suspended for all foreign nationals, citing national security. To comply, Anthropic disables both models for every user worldwide.
3. June 26-27, 2026  
Mythos 5 controls eased  
Access to Mythos 5 is restored for a short list of approved US organizations defending critical infrastructure. Fable 5 stays dark.
4. June 30, 2026  
Commerce lifts the controls  
The Commerce Department removes the export-control directive on both models after two weeks of review with Anthropic.
5. Early July 2026  
Access restored  
Anthropic begins rolling general access back to Fable 5 and Mythos 5\. The rollout is staged, not instant.

The Fable 5 timeline: eleven days from general release to a government shutdown, then eighteen more to reinstatement.Source: [Anthropic; US Department of Commerce; CNBC; Al Jazeera; Gizmodo](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says)

The episode matters beyond one model. It is the first known instance of the US using export-control authority against a specific commercial frontier model, a precedent that now sits over every lab shipping at the capability frontier. The dual-use logic is the crux: a model strong enough to help defenders find software vulnerabilities is, by the same skill, strong enough to help attackers exploit them.

## The jailbreak was not unique to Fable 5

Anthropic’s most consequential disclosure in the redeployment post is not about Fable 5 at all. After the Amazon report, Anthropic re-ran the technique against every model it had on hand and found the vulnerability-identification step was common: “our testing confirmed that many less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities as Fable 5 did in the report,” per [Anthropic’s post](https://www.anthropic.com/news/redeploying-fable-5). The harder step, producing a working exploit demonstration, turned out to be just as common: “every model we tested could produce the same demonstration as Fable 5.”

| Model tested      | Could identify the vulnerability | Could produce the exploit demo |
| ----------------- | -------------------------------- | ------------------------------ |
| Claude Fable 5    | Yes                              | Yes                            |
| Claude Opus 4.8   | Yes                              | Yes                            |
| Claude Opus 4.7   | Not disclosed                    | Yes                            |
| Claude Opus 4.6   | Not disclosed                    | Yes                            |
| Claude Sonnet 4.6 | Not disclosed                    | Yes                            |
| Claude Haiku 4.5  | Not disclosed                    | Yes                            |
| GPT-5.5           | Yes                              | Yes                            |
| GPT-5.4           | Not disclosed                    | Yes                            |
| Kimi K2.7         | Yes                              | Yes                            |

That finding reframes the whole episode. A government export-control action was triggered by a capability Anthropic then showed was present in models several tiers down the capability ladder, including a Haiku-class model priced at a twentieth of Fable 5’s rate. It is also why Anthropic pushed back on the idea that the fix is a Fable-5-specific patch: the underlying skill is broadly distributed across the frontier, so the real defense has to be the classifier layer, not the base model’s raw capability.

## How much does Claude Fable 5 cost?

The price never moved while the model was offline. Fable 5 is $10 per million input tokens and $50 per million output tokens, per [Anthropic’s launch announcement](https://www.anthropic.com/news/claude-fable-5-mythos-5). That is the most expensive rate card Anthropic publishes, exactly double Opus 4.8 and more than three times Sonnet 5 on output.

| Model            | Input / Mtok | Output / Mtok | Position                            |
| ---------------- | ------------ | ------------- | ----------------------------------- |
| Claude Fable 5   | $10.00       | $50.00        | top tier, most capable              |
| Claude Opus 4.8  | $5.00        | $25.00        | flagship, half the Fable rate       |
| GPT-5.6 Sol      | $4.00        | $20.00        | OpenAI top tier, promo since Aug 21 |
| Claude Sonnet 5  | $3.00        | $15.00        | mid-tier daily driver               |
| Claude Haiku 4.5 | $1.00        | $5.00         | cheapest tier                       |

The specs justify the tier even if the price stings. Fable 5 carries a 1M-token context window, up to 128k output tokens per request, and adaptive-thinking-only reasoning (there is no way to turn thinking off), per [Anthropic’s platform documentation](https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5). Prompt-cache reads bill at roughly $1 per million tokens. For the dated, sourced rate card across every current model, the [AI model tracker](/ai-models/) keeps them current, and the [model comparison tool](/ai-model-comparison/) prices Fable 5 against Opus 4.8, Sonnet 5, and GPT-5.6 on the same task.

Subscription users get a short grace window on the return. Per [Anthropic’s redeployment post](https://www.anthropic.com/news/redeploying-fable-5), “Fable 5 will be included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits” on Pro, Max, and Team plans. After that date, Fable 5 usage on those plans draws down credits rather than counting toward the included weekly limit, the same metering already applied to other top-tier models.

Fable 5 has a billing quirk no other Claude model has

Fable 5 ships with safety classifiers that can decline a request. When one does, the Messages API returns `stop_reason: "refusal"` as a normal HTTP 200 response, and you are not billed for a refused call before any output is generated. If you retry on another Claude model, fallback credit refunds the prompt-cache cost of the switch so you do not pay it twice. Budget for the refusal path: it is rare, but it changes how you handle responses and reconcile spend.

## Is it actually the best model? The benchmarks

The reason a team pays double is the capability gap, and on the numbers Anthropic published at launch it is real. On SWE-Bench Pro, the harder agentic-coding benchmark, Fable 5 scores 80.3% against Opus 4.8 at 69.2% and GPT-5.5 at 58.6%, per figures [transcribed by The Decoder](https://the-decoder.com/anthropic-releases-claude-fable-5-and-mythos-5-with-major-gains-in-coding-and-science/) from Anthropic’s announcement. These are vendor-reported results relayed by secondary coverage, not an independent evaluation, so read them as the launch claim rather than a settled fact.

__SWE-Bench Pro: Fable 5 versus Opus 4.8 and GPT-5.5__
| Item            | Value |
| --------------- | ----- |
| Claude Fable 5  | 80.3% |
| Claude Opus 4.8 | 69.2% |
| GPT-5.5         | 58.6% |

SWE-Bench Pro, Anthropic's harder agentic-coding benchmark, as reported at the Fable 5 launch. Fable 5 leads the frontier by roughly eleven points over Opus 4.8 and more than twenty over GPT-5.5\. Vendor-reported figures, relayed by The Decoder.Source: [Anthropic launch benchmarks, as transcribed by The Decoder](https://the-decoder.com/anthropic-releases-claude-fable-5-and-mythos-5-with-major-gains-in-coding-and-science/)

The gap widens on harder tasks. On FrontierCode, a tougher coding evaluation, Fable 5 reportedly scores 29.3% against Opus 4.8 at 13.4% and GPT-5.5 at 5.7%, more than doubling the flagship, per the same launch transcription. For the wider context on that benchmark and what it measures, the [DeepSWE versus FrontierCode breakdown](/blog/deepswe-vs-frontiercode/) covers the field. Anthropic also reported the highest score of any model on Hebbia’s Finance Benchmark and state-of-the-art vision, and Stripe said Fable 5 compressed a roughly five-month, 50-million-line Ruby migration into days. Those are customer-reported and vendor-reported claims, credible as signals but not independently benchmarked.

| Benchmark     | Fable 5 | Opus 4.8 | GPT-5.5 |
| ------------- | ------- | -------- | ------- |
| SWE-Bench Pro | 80.3%   | 69.2%    | 58.6%   |
| FrontierCode  | 29.3%   | 13.4%    | 5.7%    |

## Fable versus Mythos, and the safety architecture

The distinction that got Fable 5 suspended is the one that makes it usable at all. Fable 5 ships with safety classifiers that route flagged requests, on cybersecurity, biology, chemistry, and model distillation, to a response from Claude Opus 4.8 instead. Those classifiers trigger in fewer than 5% of sessions, per Anthropic. Mythos 5 is the same model without the classifiers, offered only in limited release through Anthropic’s Project Glasswing to approved customers.

That is the whole story of the ban in one design decision. The government’s concern was that Fable 5’s classifiers could be bypassed to reach the raw Mythos-class capability underneath, which is exactly why Mythos itself is gated to vetted partners. The safeguards are not window dressing; they are the load-bearing wall between a generally available model and one the government treats as controlled technology.

## What Anthropic changed, and the new industry rules

Anthropic did not claim the fix closes the gap for good. Its own framing, from the redeployment post, is blunt: “it is probably impossible to make any AI model fully robust (that is, impervious) to jailbreaks.” What shipped instead is narrower and more measurable. A new safety classifier now blocks “the specific technique described in the Amazon report” in “over 99% of cases,” built on a “defense in depth” approach that layers multiple safety mechanisms rather than relying on one classifier alone. Anthropic also widened the safety margin around ambiguous requests, which trades a higher false-positive rate (more legitimate requests getting flagged) for a lower chance the specific bypass slips through.

A 99% block rate is not a promise of zero jailbreaks

Anthropic’s own position is that full robustness is not achievable. The 99%-plus figure covers the one disclosed technique, not jailbreaks in general. Treat future minor jailbreaks against Fable 5 or any other frontier model as expected, not as a sign the classifier layer failed; the bar Anthropic is holding itself to is that a jailbreak should not unblock the model’s most dangerous core capabilities, not that no jailbreak will ever be found.

The more durable outcome of the episode is procedural, not technical. Anthropic, Amazon, Microsoft, and Google (alongside Project Glasswing partners) are converging on a shared framework for grading how serious a jailbreak actually is, scored across four criteria: capability gain, breadth of capability gain, ease of weaponization, and discoverability. A severity scale that several labs use the same way makes it possible to compare incidents across vendors instead of every lab describing its own jailbreaks in its own vocabulary, which is exactly the ambiguity that made the Fable 5 episode hard to reason about from the outside.

Anthropic also committed to four specific practices with the US government going forward: pre-release government access and evaluation for models relevant to national security, rapid information sharing when jailbreaks or new safeguards are found, dedicated resources for joint research with government partners, and work toward a common industry bar so safety expectations do not vary lab to lab. Read together with the export-control precedent, this is Anthropic trying to convert a one-off, ad hoc shutdown into a standing process, one where the government sees the risk before a public incident forces a suspension rather than after.

## Should you use it?

Fable 5 is the right tool for a narrow, expensive slice of work, and the wrong default for everything else. Three rules cover the decision:

1. **Reserve it for the hardest autonomy work.** Where a few extra points of reliability on long-horizon agentic tasks pay for themselves, Fable 5 earns its rate. Its token efficiency can also claw back part of the sticker premium: a model that finishes in fewer tokens or fewer retries costs less per task than the per-token gap suggests. The [Fable 5 versus Opus 4.8 cost-per-task math](/blog/claude-fable-5-vs-opus-4-8-cost-per-task/) works that trade through.
2. **Default to Opus 4.8 or Sonnet 5 for volume.** At half and a fifth of the output rate, they carry the daily-driver load. The [Sonnet 5 pricing and benchmark rundown](/blog/claude-sonnet-5-pricing-benchmarks/) shows how far the mid-tier now reaches. Cost the specific matchup on your own task shape with the [cost-per-task calculator](/ai-coding-cost-calculator/).
3. **Weigh the access risk.** The suspension is lifted, but it happened once, and the export-control precedent is now live. Anthropic’s pre-release-access commitment to the US government should make a repeat less likely for the next model, not impossible for this one. If your workload cannot tolerate a sudden outage of a single model, keep a fallback wired. The [GPT-5.6 Sol versus Fable 5 comparison](/blog/gpt-5-6-vs-claude-opus-4-8-vs-fable-5/) sizes up the nearest rival, itself a government-gated preview.

Fable 5 comes back as the same thing it launched as: the most capable model on the market, at the highest price on the market, now carrying a track record of being switched off by the state. That last fact is not priced into the rate card, but for anyone building on a single model it belongs in the risk column.

## Frequently asked questions

Is Claude Fable 5 available again?

Yes. The US Commerce Department lifted the export-control directive on June 30, 2026, and Anthropic began restoring general access in early July. The rollout is staged rather than instant, so availability may return gradually across the Claude API and cloud platforms.

How much does Claude Fable 5 cost?

API pricing is $10 per million input tokens and $50 per million output tokens, with prompt-cache reads at roughly $1 per million. It is the most expensive model Anthropic sells, exactly double Opus 4.8 and more than three times Sonnet 5 on output.

Why was Claude Fable 5 suspended?

On June 12, 2026, the US government issued an export-control directive barring foreign nationals from accessing Fable 5 and Mythos 5 on national-security grounds. The trigger was an Amazon security report showing Fable 5 could be prompted to identify software vulnerabilities and, in one case, produce code exploiting one. To comply, Anthropic disabled both models for all users. It was the first known US export-control action against a specific commercial frontier model.

Was the Fable 5 jailbreak unique to that model?

No. Anthropic retested the Amazon report against other models and found the vulnerability-identification step was replicated by Claude Opus 4.8, GPT-5.5, and Kimi K2.7, and that every model it tested, including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, GPT-5.4, GPT-5.5, and Kimi K2.7, could produce the same exploit demonstration Fable 5 did.

Is Claude Fable 5 the best coding model?

On Anthropic's launch benchmarks, Fable 5 leads: 80.3% on SWE-Bench Pro against Opus 4.8 at 69.2% and GPT-5.5 at 58.6%, and 29.3% on FrontierCode against 13.4% and 5.7%. These are vendor-reported figures relayed by secondary coverage, not independent evaluations.

What is the difference between Fable 5 and Mythos 5?

They are the same underlying model. Fable 5 includes safety classifiers that route flagged requests to Claude Opus 4.8 and is generally available. Mythos 5 has no classifiers and is offered only in limited release through Anthropics Project Glasswing to approved customers.

What did Anthropic change after the Fable 5 jailbreak?

Anthropic shipped a new safety classifier that blocks the specific disclosed technique in over 99% of cases, using a defense-in-depth approach layering multiple safety mechanisms, and widened its safety margin around ambiguous requests. It also joined Amazon, Microsoft, and Google in adopting a shared framework for grading jailbreak severity, and committed to giving the US government pre-release access to evaluate future national-security-relevant models.

## Sources

* Anthropic. (2026, July 1). _Redeploying Claude Fable 5_ \[primary; Amazon-report jailbreak mechanism, cross-model retesting results, new classifier block rate, defense-in-depth approach, jailbreak severity framework, government collaboration commitments, post-July-7 usage-credit pricing\]. Verified 2026-07-01\. [anthropic.com/news/redeploying-fable-5](https://www.anthropic.com/news/redeploying-fable-5)
* Anthropic. (2026). _Claude Fable 5 and Claude Mythos 5_ \[vendor launch post; release date, positioning, pricing, capability claims, classifier fallback\]. Verified 2026-07-01\. [anthropic.com/news/claude-fable-5-mythos-5](https://www.anthropic.com/news/claude-fable-5-mythos-5)
* Anthropic. (2026). _Introducing Claude Fable 5 and Claude Mythos 5_ \[Claude Platform docs; `claude-fable-5` API id, $10 / $50 pricing, 1M context, 128k output, adaptive thinking, refusal and fallback billing, 30-day retention\]. Verified 2026-07-01\. [platform.claude.com/docs/en/about-claude/models/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)
* Anthropic \[@AnthropicAI\]. (2026, June 12). _Statement on the export-control directive_ \[primary; the directive suspended access for all foreign nationals\]. Verified 2026-07-01\. [x.com/AnthropicAI/status/2065597531644743999](https://x.com/AnthropicAI/status/2065597531644743999)
* CNBC. (2026, June 30). _Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5_ \[secondary; reversal date and Anthropic restoration statement\]. Verified 2026-07-01\. [cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html](https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html)
* Al Jazeera. (2026, July 1). _US lifts restrictions on powerful AI models Fable, Mythos, Anthropic says_ \[secondary; restoration framing, Lutnick statement, foreign-national directive\]. Verified 2026-07-01\. [aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says)
* Lutnick, H. \[@howardlutnick\]. (2026). _Statement on approving Fable 5_ \[primary; Commerce Secretary quote\]. Verified 2026-07-01\. [x.com/howardlutnick/status/2072100729603452965](https://x.com/howardlutnick/status/2072100729603452965)
* Fortune. (2026, June 13). _Anthropic disables Fable and Mythos AI models after U.S. government bars it from giving foreigners access_ \[secondary; compliance response\]. Verified 2026-07-01\. [fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/](https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/)
* The Decoder. (2026). _Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science_ \[secondary; transcription of launch benchmarks SWE-Bench Pro and FrontierCode\]. Verified 2026-07-01\. [the-decoder.com/anthropic-releases-claude-fable-5-and-mythos-5-with-major-gains-in-coding-and-science/](https://the-decoder.com/anthropic-releases-claude-fable-5-and-mythos-5-with-major-gains-in-coding-and-science/)