# Claude Fable 5

> Source: https://aiwiki.ai/wiki/claude_fable_5
> Updated: 2026-09-04
> Fact-checked: 2026-09-04
> Categories: AI Models, Anthropic, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Claude Fable 5." aiwiki.ai, 4 Sept 2026. https://aiwiki.ai/wiki/claude_fable_5
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Claude Fable 5 is a large [language model](https://aiwiki.ai/wiki/language_model) developed by [Anthropic](https://aiwiki.ai/wiki/anthropic) and released on June 9, 2026, as the first publicly available model in a new "Mythos class" that the company describes as "a tier of Claude models that sit above our Opus class in capability." [1] It is the most capable generally available member of the [Claude](https://aiwiki.ai/wiki/claude) family, and Anthropic states that "Fable 5's capabilities exceed those of any model we've ever made generally available." [1] The model is served through the Claude API under the identifier claude-fable-5 at $10 per million input tokens and $50 per million output tokens, roughly twice the price of the previous flagship, [Claude Opus 4.8](https://aiwiki.ai/wiki/claude_opus_4_8). [1][6] Anthropic released its successor, Claude Fable 5.1, on September 1, 2026, retaining the $10 input and $50 output rates while cutting cached-input reads from $1.00 to $0.25 per million tokens.[17][18]

## What is Claude Fable 5?

Claude Fable 5 is a reasoning-oriented, multimodal [large language model](https://aiwiki.ai/wiki/large_language_model) that accepts text and image input and returns text. [4] It inaugurated Anthropic's Claude 5 generation and launched alongside a restricted sibling, [Claude Mythos 5](https://aiwiki.ai/wiki/claude_mythos_5), with which it shares the same underlying weights (see below). Anthropic positions the Mythos class as its first tier of models to exceed the Opus class, and frames the release around both capability and risk, writing that "Mythos-class models have reached a threshold where they present significant risks." [1] The company reports state-of-the-art results on nearly all tested benchmarks, with particular strength in software engineering, knowledge work, vision, and scientific research. [1] The most-quoted figure from the launch, an acceleration of "aspects of the drug design process by around 10 times," came from Anthropic's own internal protein design experts working with Mythos 5, not from an outside customer. [1] In August 2026, Anthropic researcher Levent Alpöge reported that Fable 5, running autonomously with a generic prompt and no internet access, had reproduced five of the ten mathematical results OpenAI announced on August 1, 2026 for its unreleased [Astra](https://aiwiki.ai/wiki/openai_astra) model, writing on August 2 that "after 24h i have half of them with fable" (problems 4 through 8 of OpenAI's set) and that he thought Ehrhart's volume conjecture was the only one so far where both models found essentially the same argument. [15] These are Alpöge's own attributed claims about his experiment, posted on X and reproduced in Zvi Mowshowitz's writeup the following day. [16]

The model supports a [context window](https://aiwiki.ai/wiki/context_window) of 1,000,000 tokens and can return up to 128,000 tokens in a single response. [4][12] Anthropic gives its reliable knowledge cutoff as January 2026. [12] On the [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) Intelligence Index version 4.1, an aggregate of nine evaluations that includes [GPQA Diamond](https://aiwiki.ai/wiki/gpqa_diamond), [Humanity's Last Exam](https://aiwiki.ai/wiki/hle), and [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1, the configuration Artificial Analysis labels "Claude Fable 5 (with fallback)" scores 59.9, which rounds to 60 and was the highest of any model the site tracked when Fable 5 launched. [4] As of July 27, 2026 that configuration ranks second on the same index, behind Claude Opus 5 at 60.7 and ahead of GPT-5.6 Sol at 58.9. [13] Because Artificial Analysis measures Fable 5 with its safety fallback to Opus 4.8 in place, and at maximum reasoning effort, the figure describes the deployed system rather than the raw model. [4]

| Attribute | Detail |
|---|---|
| Developer | Anthropic |
| Released | June 9, 2026 (redeployed July 1, 2026) |
| Model ID | claude-fable-5 |
| Model class | Mythos class (above Opus) |
| Type | Reasoning, multimodal (text and image in, text out) |
| Context window | 1,000,000 tokens |
| Maximum output | 128,000 tokens |
| Reliable knowledge cutoff | January 2026 |
| API price | $10 per 1M input, $50 per 1M output |
| Prompt caching | $12.50 write (5 minute), $20 write (1 hour), $1.00 read per 1M tokens |
| Artificial Analysis Intelligence Index v4.1 | 59.9 (first at launch; second behind Claude Opus 5 as of July 27, 2026) |

## What is the difference between Fable 5 and Mythos 5?

Fable 5 and Claude Mythos 5 are the same underlying model; the difference is safety configuration. Fable 5 runs with all safeguards in place, while Mythos 5 is "the same underlying model as Fable 5, but with the safeguards lifted in some areas" and is available only to approved organizations. [1] Because the raw model has dual-use capabilities, Fable 5 uses classifiers that reroute sensitive requests, for example in [cybersecurity](https://aiwiki.ai/wiki/ai_in_cybersecurity) and in biology and chemistry, to Claude Opus 4.8. Anthropic reports that "more than 95% of Fable sessions involve no fallback at all." [1] The company explains the split by noting that "the same queries that are beneficial in the hands of cybersecurity professionals and biology researchers could be dangerous if available to malicious actors," and that "without safeguards, Fable 5's capabilities in areas like cybersecurity could be misused to cause serious damage." [1]

Mythos 5 is not generally available. At launch it was restricted to partners in Anthropic's Project Glasswing, with cybersecurity safeguards lifted, pending a broader trusted-access program. [1] Glasswing began in early April 2026 with roughly 50 partners running the earlier Claude Mythos Preview, and on June 2, 2026 Anthropic extended it by approximately 150 new organizations across more than 15 countries, so membership after that expansion was on the order of 200 rather than 150. [11] Anthropic's announcement is inconsistent about the planned biology track: its availability section lists Mythos 5 as going "soon to select biology researchers (with biology and chemistry safeguards lifted)," while the section describing the trusted access programs says that program "will provide access to Fable 5 with the biology and chemistry safeguards removed (but the cyber safeguards still in place)." [1] Both models are priced identically at $10 and $50 per million input and output tokens. [1]

## How does Fable 5 compare to Opus 4.8?

Claude Fable 5 sits one tier above Claude Opus 4.8, which had been Anthropic's most capable generally available model and now serves as the fallback target for Fable 5's safeguards. [1] Anthropic's system card reports the two Mythos-class configurations separately, and the difference matters: on SWE-bench Pro, a harder variant of [SWE-bench](https://aiwiki.ai/wiki/swe_bench) drawn from actively maintained repositories, Mythos 5 scored 80.3% and Fable 5 scored 80%, against 69.2% for Opus 4.8 and 58.6% for OpenAI's GPT-5.5. [9] On [SWE-bench Verified](https://aiwiki.ai/wiki/swe-bench_verified) the same table gives Mythos 5 95.5%, Fable 5 95%, and Opus 4.8 88.6%, and third-party leaderboards that aggregate vendor submissions repeat those figures. [9][5] All of these are Anthropic's own reported numbers, averaged over five trials, not independent measurements. [9]

The safeguards can show up in the scores. On Terminal-Bench 2.1 the system card reports 88% for Mythos 5 but 84.3% for Fable 5, noting that 20.9% of Fable 5 trials hit a safety refusal and fell back to Opus 4.8 for the rest of the trajectory. [9] Anthropic reports Mythos 5 at 94.1% on [GPQA](https://aiwiki.ai/wiki/gpqa) Diamond and says it now regards that benchmark as saturated; the 92.6% figure widely quoted for Fable 5 is an Artificial Analysis measurement of the public, fallback-equipped configuration rather than an Anthropic-published result. [9][8]

| Benchmark (Anthropic system card, June 2026) | Claude Mythos 5 | Claude Fable 5 | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| SWE-bench Pro | 80.3% | 80% | 69.2% | 58.6% |
| SWE-bench Verified | 95.5% | 95% | 88.6% | not reported |
| Terminal-Bench 2.1 | 88.0% | 84.3% | 82.7% | 83.4% (Codex CLI) |
| GPQA Diamond | 94.1% | not reported | not reported | not reported |

Anthropic's Claude 5 generation also includes Claude Sonnet 5, a lower-cost everyday model announced on June 30, 2026 that sits below Fable 5 in capability and price, and [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5), announced on July 24, 2026. [14] Opus 5 is priced like Opus 4.8, at $5 and $25 per million tokens, or half the price of Fable 5, and Anthropic's documentation still describes Fable 5 as its most capable widely released model while pointing developers to Opus 5 for complex agentic coding and enterprise work. [6][12]

## How much does Claude Fable 5 cost?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens, double the $5 and $25 charged for Claude Opus 4.8. [1][6] Prompt caching reduces costs substantially: cached input reads are billed at $1.00 per million tokens, a 90% discount, while cache writes cost $12.50 per million tokens with the five-minute cache and $20 per million with the one-hour cache. [12] Requests routed to US-only inference through the inference_geo parameter carry a 1.1x multiplier on every token category, including cache writes and reads. [2][12] Anthropic requires 30-day data retention for safety monitoring on all Mythos-class traffic and says the data will not be used to train new Claude models. [1] Billing is identical whether the model is accessed as Fable 5 or, for approved organizations, as Mythos 5. [1]

## What is the availability of Claude Fable 5?

At launch, Claude Fable 5 was available on the web to Pro, Max, Team, and seat-based Enterprise subscribers, through the Claude API, and on [Amazon Bedrock](https://aiwiki.ai/wiki/amazon_bedrock), Claude Platform on AWS, Google Cloud, and Microsoft Foundry. [1][12] Anthropic included the model on those subscription plans at no extra cost from June 9 through June 22, 2026, with usage credits required from June 23 onward. [1]

The rollout was interrupted by an [export-control](https://aiwiki.ai/wiki/export_controls) episode. On Friday, June 12, 2026, the US government applied export controls to Fable 5 and Mythos 5 that required Anthropic to restrict access by foreign nationals, whether inside or outside the United States. [3] Because the order took effect immediately and the company had no reliable way to verify nationality in real time, Anthropic suspended access to both models for all users. [3]

Anthropic says the directive followed a report by Amazon researchers describing a method of prompting Fable 5 so that it identified a number of software vulnerabilities, and in one case produced code demonstrating how a vulnerability could be exploited. [3] The characterization is contested. Anthropic asked Katie Moussouris, founder of Luta Security, to assess the report; she published her own account arguing that "that is not a guardrail bypass" but rather "the most valuable thing an AI model can do for defensive security: executing the find, fix, and test loop defenders run every day," and concluding that the capability could not be removed without degrading legitimate security work. [10] Anthropic's own testing found that less capable models, including Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities, and that every model it tested could reproduce the exploit demonstration; none of those models faced comparable restrictions. [3][10]

To restore access, Anthropic trained a new cybersecurity classifier that it says blocks the specific technique described in the Amazon report in more than 99% of cases, notifying the user and routing the flagged request to Opus 4.8 instead. [3] Researchers at the Center for AI Standards and Innovation (CAISI), part of the US Department of Commerce, tested both the previous and the new safeguards; Anthropic says they agreed the safeguards were extraordinarily strong. [3] Anthropic also committed to expanded pre-release access for government evaluators, rapid disclosure of significant jailbreaks, dedicated staff and compute for joint research, a HackerOne program for cyber jailbreak submissions, and work toward a shared voluntary security standard. [3] Separately, more than 100 cybersecurity professionals signed an open letter organized by Alex Stamos and addressed to Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross, warning that the controls risked doing more harm than good; the signatories included executives from Nvidia, Adobe, Zoom, Google, and Sophos. [10]

The controls were lifted on June 30, 2026, and Anthropic restored Fable 5 to users globally on Wednesday, July 1, 2026 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, saying it would re-enable access on AWS, Google Cloud, and Microsoft Foundry as quickly as possible. [3][7] The Record described the outage as a roughly three-week shutdown; the interval from the June 12 suspension to the July 1 restoration is 19 days. [10] Mythos 5 was a partial exception: Anthropic restored it to a set of US organizations after US government approval on June 26, and it remains limited to vetted US organizations through Project Glasswing while the company negotiates wider domestic and international access. [3][10] CNBC reported that the letter formally lifting the ban was addressed to Anthropic co-founder [Tom Brown](https://aiwiki.ai/wiki/tom_brown) rather than to chief executive Dario Amodei. [7][10] On redeployment, Anthropic included Fable 5 for up to 50% of weekly usage limits on Pro, Max, Team, and select Enterprise plans through July 7, 2026, after which it moved to usage credits. [3]

## Claude Fable 5.1 update

Anthropic released Claude Fable 5.1 on September 1, 2026 as the successor to Fable 5; the model is covered in full in its own article, [Claude Fable 5.1](https://aiwiki.ai/wiki/claude_fable_5_1). The new model keeps Fable 5's $10 input and $50 output prices and cuts cache reads from $1.00 to $0.25 per million tokens, which Anthropic estimates lowers typical costs by about 25 percent and highly agentic costs by up to about 45 percent, based on its own August 2026 usage.[17][18] It uses the API identifier `claude-fable-5-1`, keeps the one-million-token context window, the 128,000-token maximum output, and always-on adaptive thinking, and moves the reliable knowledge cutoff to June 2026.[18] Like Fable 5, it shares its weights with a restricted sibling, Claude Mythos 5.1, and falls back to Opus 4.8 or Opus 5 when its safeguards fire; the safeguards were made more precise so that vulnerability discovery in source code is allowed and biology-related false positives are rarer, with about 60 percent fewer cyber interventions per Claude Code session by Anthropic's count.[17][19] Anthropic reported higher scores than Fable 5 on every launch benchmark, including 52.6 percent against 24.7 percent on Terminal-Bench-Science 0.1 and 55.8 percent against 42.0 percent on Terminal-Bench 4.0, all vendor measurements taken with production safeguards enabled and subject to the same fallback caveats that apply to Fable 5.[17] Artificial Analysis measured the deployed system, with fallback, at 66 on its Intelligence Index, the highest score it had recorded, at $3.76 per index task, about 20 percent more than Fable 5 because the newer model produced roughly 1.7 times as many output tokens.[20][21][22] The release also introduced Enterprise Frontier Safeguards for customer-controlled retention, a statistical text watermark for EU AI Act compliance, and anti-distillation limits on editing the context around thinking blocks.[17] A Vals AI post reported that Fable 5.1 produced a proposed decoding of Sir Thomas Urquhart's Cyphral Distich; the reading depends on the 1834 printing of the text rather than the 1653 edition, which does not contain the numbered distich, and the evidence is examined on the Fable 5.1 page.[23][24][25][26][27][28]

## See also

- [AI API cost and context planner](https://aiwiki.ai/tools/ai_cost_calculator?model=claude-fable-5)
- [AI model comparison finder](https://aiwiki.ai/tools/compare_models?models=claude-fable-5,claude-mythos-5)
- [AI model lifecycle calendar](https://aiwiki.ai/tools/model_deprecations?provider=anthropic)

## References

[1] Anthropic. "Claude Fable 5 and Claude Mythos 5." June 9, 2026. https://www.anthropic.com/news/claude-fable-5-mythos-5

[2] Anthropic. "Claude Fable" (model overview). https://www.anthropic.com/claude/fable

[3] Anthropic. "Redeploying Claude Fable 5." June 30, 2026, updated July 1, 2026. https://www.anthropic.com/news/redeploying-fable-5

[4] Artificial Analysis. "Claude Fable 5" model page. https://artificialanalysis.ai/models/claude-fable-5

[5] SWE-bench Verified leaderboard. LLM Stats. https://llm-stats.com/benchmarks/swe-bench-verified

[6] Finout. "Claude Fable 5 and Mythos 5: Pricing, API Costs, and Benchmark Comparison vs Opus 4.8 and GPT-5.5." June 2026. https://www.finout.io/blog/claude-fable-5-mythos-5-pricing-benchmarks

[7] CNBC. "Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5." June 30, 2026. https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html

[8] Requesty. "Anthropic claude-fable-5 API Pricing, Context Window and Benchmarks." https://www.requesty.ai/models/anthropic/claude-fable-5

[9] Anthropic. "Claude Fable 5 System Card," sections 8.1, 8.2, 8.3 and 8.8. June 2026. https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf

[10] Martin, Alexander. "US lifts export controls on Anthropic's frontier cybersecurity AI models." The Record (Recorded Future News), July 1, 2026. https://therecord.media/us-lifts-export-controls-anthropic-cyber-models

[11] Anthropic. "Expanding Project Glasswing." June 2, 2026. https://www.anthropic.com/news/expanding-project-glasswing

[12] Anthropic. "Models overview" and "Pricing," Claude developer documentation. Accessed July 27, 2026. https://platform.claude.com/docs/en/about-claude/models/overview

[13] Artificial Analysis. Model leaderboard, Intelligence Index v4.1. Accessed July 27, 2026. https://artificialanalysis.ai/models

[14] Anthropic. "Introducing Claude Opus 5." July 24, 2026. https://www.anthropic.com/news/claude-opus-5

[15] Alpöge, L. (@__alpoge__). X thread on reproducing Astra results with Claude Fable 5. August 2, 2026. https://x.com/__alpoge__/status/2083855298239078748

[16] Mowshowitz, Zvi. "OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems." August 3, 2026. https://thezvi.substack.com/p/openais-unreleased-model-astra-solves

[17] Anthropic. "Claude Fable 5.1 and Claude Mythos 5.1." September 1, 2026. https://www.anthropic.com/claude-fable-and-mythos-5-1

[18] Anthropic. "Claude Fable 5.1" (developer documentation). Accessed September 1, 2026. https://platform.claude.com/docs/en/models/fable-5-1/overview

[19] Brandom, Russell. "Anthropic's new Fable release is cheaper, less restrictive." TechCrunch, September 1, 2026. https://techcrunch.com/2026/09/01/anthropics-new-fable-release-is-cheaper-less-restrictive/

[20] Artificial Analysis. "Claude Fable 5.1 tops the Artificial Analysis Intelligence Index but costs 20% more per task than Fable 5 despite a 75% cache read price cut." September 1, 2026. https://artificialanalysis.ai/articles/claude-fable-5-1

[21] Artificial Analysis. "Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) Intelligence, Performance & Price Analysis." Accessed September 3, 2026. https://artificialanalysis.ai/models/claude-fable-5-1

[22] Artificial Analysis. "Artificial Analysis Intelligence Benchmarking Methodology." Accessed September 3, 2026. https://artificialanalysis.ai/methodology/intelligence-benchmarking

[23] Jaff, Geby. "Claude Fable 5.1 Solves the Cyphral Distich." Vals AI, August 31, 2026. https://www.vals.ai/blogs/fable-solves-cyphral-distich

[24] Maitland Club. *The Works of Sir Thomas Urquhart of Cromarty, Knight*. Edinburgh, 1834, pp. 412-417. https://archive.org/details/worksofsirthomas00mait

[25] "Cipher." *Notes and Queries*, Series 9, vol. 3, February 18, 1899, p. 124. https://en.wikisource.org/wiki/Page:Notes_and_Queries_-_Series_9_-_Volume_3.djvu/134

[26] Schmeh, Klaus. "Revisited: Thomas Urquhart's encrypted poems." Cipherbrain, July 28, 2019. https://scienceblogs.de/klausis-krypto-kolumne/2019/07/28/revisited-thomas-urquharts-encrypted-poems/

[27] Urquhart, Thomas. *Logopandecteision, or an Introduction to the Vniversal Language*. London, 1653. Digitized microfilm copy. https://archive.org/details/bim_early-english-books-1641-1700_logopandecteision-or-an_urquhart-sir-thomas_1653

[28] Eason, James. "Logopandecteision: The Epilogue." Transcription of the 1653 edition. https://penelope.uchicago.edu/urquhart/logo14.html