Claude Fable 5: The AI Model That Spooked Governments

TL;DR:

Anthropic launched Claude Fable 5 on June 9, 2026 — the first Mythos-class model available to the general public. Claude Mythos 5, the version with safety restrictions removed, remains limited to approved organizations only.
On SWE-bench Pro, Fable 5 scores 80.3% against GPT-5.5's 58.6% and Gemini 3.1 Pro's 54.2%. On GPQA Diamond (doctoral-level science), Gemini leads at 94.3%.
Pricing ($10/$50 per million tokens) and safety filters (8–9% of tasks blocked or degraded in real-world usage) are the most documented friction points of this launch.

Three days before Anthropic released its most powerful model to the public, the company published a statement warning that artificial intelligence is advancing faster than society can adapt. Four days later, it shipped Claude Fable 5.

That contradiction is not accidental. It is the defining tension of the AI industry in 2026.

On June 9, 2026, Anthropic announced two models: Claude Fable 5, the first Mythos-class model available for general use, and Claude Mythos 5, the version with safety classifiers removed, accessible only to organizations approved through Project Glasswing. Same underlying model. Completely different rules.

Here is what the data actually shows.

What Mythos-Class Means and Why It Matters

Anthropic introduced the term "Mythos-class" to designate a capability tier above Opus, which until now represented the ceiling of what was publicly available. In this framework, Fable 5 is the entry point: same architecture as Mythos 5, with active safety classifiers filtering certain request types.

Back in April 2026 we analyzed Claude Mythos Preview in detail, when the model was still locked away and Anthropic maintained that its capabilities were too risky for open distribution. That changed. Fable 5 is the answer to that restriction: Mythos-class capabilities with active safeguards.

The official documentation is explicit: both models share the same underlying architecture. The difference lies in what they are allowed to execute.

Claude Fable 5 includes:

Active safety classifiers across three domains: cybersecurity, biology/chemistry, and model distillation.
When a request is rejected, the API can automatically fall back to Claude Opus 4.8.
Anthropic says this fallback occurs in fewer than 5% of sessions. Independent reviewers report 8–9% in real-world usage.

Claude Mythos 5 operates without:

Fable 5's cybersecurity classifiers.
Available only through Project Glasswing, the controlled-access program detailed below.
Direct successor to Claude Mythos Preview.

Technical Specifications

The concrete numbers that matter when evaluating the model:

Context window: 1 million tokens by default.

Maximum output tokens: 128,000 per request — a practical difference for tasks involving extensive code generation or long-document analysis.

Pricing: $10 per million input tokens, $50 per million output tokens. That is twice the cost of Claude Opus 4.5 through 4.8.

Knowledge cutoff: January 2026.

Adaptive thinking: always on, cannot be disabled. In previous models, extended thinking was optional. In Fable 5 and Mythos 5, it is the only available reasoning mode. Depth is controlled via the effort parameter.

There is a significant API behavior change: raw thinking chain content is never returned. Thinking blocks are empty by default (display: "omitted"). To receive summarized reasoning output, you must explicitly set display: "summarized". Any integration that relied on inspecting the model's reasoning process will need to be updated.

Features included at launch:

effort parameter for controlling reasoning depth.
task-budgets (beta): explicit resource limits per task.
Native memory tool.
Context editing with compaction.
Vision support.

Data retention: Claude Fable 5 and Mythos 5 are Covered Models, which means mandatory 30-day data retention. This applies even to accounts with active zero-retention agreements — making both models incompatible with many regulated-industry workflows.

Project Glasswing: The Program Nobody Mentions Enough

Project Glasswing is not a conventional early-access program. In Anthropic's own framing, it is an effort to "secure critical software before hostile actors can weaponize similar technologies."

The context matters: Claude Mythos Preview demonstrated the ability to autonomously identify software vulnerabilities, performing at or above all but the most elite human cybersecurity specialists. That creates an obvious problem: if the model exists, someone will have access to it. The question is who, and under what conditions.

Anthropic answered that question by building a controlled perimeter with US government involvement.

Program numbers as of June 10, 2026:

Over 150 organizations across more than 15 countries.
Founding partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.
$100 million in model usage credits donated.
$4 million in direct funding to open-source security organizations: $2.5M to Alpha-Omega and OpenSSF; $1.5M to the Apache Software Foundation.

Documented results:

Over 10,000 vulnerabilities identified in critical infrastructure software.
A 27-year-old flaw in OpenBSD.
A 16-year-old vulnerability in FFmpeg.
Multiple Linux kernel exploits.
Vulnerability reproduction rate: 83.1% — the previous best model achieved 66.6%.

These are the numbers that justify, according to Anthropic, why Mythos 5 without filters is not available to the general public. Post-preview access pricing for approved organizations: $25/$125 per million input/output tokens.

Benchmarks: Where It Wins and Where It Doesn't

Benchmarks are the first level of evidence — not the only one, but the most comparable across models.

SWE-bench Pro (high-difficulty software engineering tasks):

Model	Release	SWE-bench Pro
Claude Fable 5	Jun 9, 2026	80.3%
GPT-5.5 (OpenAI)	Apr 23, 2026	58.6%
Gemini 3.1 Pro (Google)	Feb 19, 2026	54.2%

Fable 5's advantage over GPT-5.5 is 21.7 percentage points — larger than the gap between GPT-5.5 and Gemini 3.1 Pro combined.

GPQA Diamond (doctoral-level scientific reasoning):

Model	GPQA Diamond
Gemini 3.1 Pro	94.3%
GPT-5.5	92.8%
Claude Fable 5	91.3%

On high-complexity scientific reasoning, Gemini 3.1 Pro leads. The three-model spread is less than 3 percentage points — statistically close but consistent.

FrontierCode (code quality and efficiency): Claude Fable 5 leads across all evaluated frontier models.

Security (vulnerability reproduction): 83.1% with Mythos 5 versus the previous best of 66.6%.

The most-cited real-world case in the launch announcement: Stripe migrated a 50-million-line Ruby codebase in one day — a task the engineering team had estimated at over two months of manual work.

First Impressions: What Early Users Found

Simon Willison, developer and writer known for the datasette project and systematic AI tool reviews, published his observations on June 9 with real billing data attached.

His findings: five and a half hours of intensive work with the model cost $110.42, with the bulk ($99.26) going toward building a Datasette agent. His assessment: "It's a beast. It's slow, it's expensive, and it's been happily churning through everything I've thrown at it so far. As is frequently the case with current frontier models, the challenge is finding tasks that it can't do."

Technical work completed in those sessions:

Generated a 13.9MB Python wheel to run CPython in WebAssembly sandbox environments.
Implemented four interconnected improvements to the LLM library, including tool-call pause/resume mechanics and error handling.
Proactively identified Brett Cannon's WASI builds and debugged path resolution issues.

Willison also identified a concrete limitation: unpredictable token usage variation for SVG generation across identical requests at different thinking levels.

Andrej Karpathy described the release as "a major-version-bump-deserving step change forward."

These are not marketing assessments. They are from users measuring results against real costs.

Launch communications have a documented tendency to omit real friction. Independent reviewers document it instead.

Real production costs: The Decoder documented a case where monthly API costs jumped from $200 to $10,000 under enterprise billing, with potential to reach $20,000 with Fable 5. In markets where hourly development costs are lower, the comparison to hiring additional capacity may not favor the model.

Safety filters in practice: Anthropic's stated fallback rate (under 5%) does not match independent measurements (8–9% in real usage). The blocking patterns extend beyond what the intentional design suggests:

A medical physicist could not use the model because the term "nuclear" triggered classifiers, blocking legitimate MRI segmentation work.
Computational biology researchers encountered refusals on standard technical queries.
One Max plan user exhausted the entire five-hour session window without completing the assigned task.

These are not users attempting to circumvent security restrictions. They are legitimate use cases blocked by miscalibrated classifiers.

The invisible degradation: The Decoder reported that Anthropic deliberately manipulates responses in approximately 0.03% of traffic — specifically on topics related to ML accelerator design — to prevent model distillation by competitors. The percentage is small, but the principle matters for organizations evaluating the model for sensitive technical research.

Mandatory data retention: The 30-day retention requirement — which applies even to accounts with zero-retention agreements — makes Fable 5 a non-starter for finance, healthcare, and legal workflows under strict regulatory frameworks.

The Competitive Landscape: What Is Coming

Fable 5 does not compete in a static market. The capability and pricing comparison shifts every few weeks.

Model	Company	Release	Input ($/M tokens)	Output ($/M tokens)	SWE-bench Pro	GPQA Diamond
Claude Fable 5	Anthropic	Jun 9, 2026	$10	$50	80.3%	91.3%
GPT-5.5	OpenAI	Apr 23, 2026	$5	$30	58.6%	92.8%
Gemini 3.1 Pro	Google	Feb 19, 2026	$2	$12	54.2%	94.3%

Additional context:

OpenAI has GPT-5.6 in internal testing. Estimated launch: June 2026.
Google released Gemini 3.5 Live Translate on June 10: real-time speech-to-speech in 70+ languages.
xAI's Grok 4 appears as a fourth relevant competitor in benchmark comparisons.
OpenAI filed its confidential S-1 with the SEC on June 8. Anthropic had done the same days earlier. Two of the most capable AI labs in the world are in simultaneous pre-IPO mode.

The model choice depends on the task. For software engineering, Fable 5 has a measurable and substantial advantage. For scientific research, Gemini 3.1 Pro competes at a quarter of the price. For general use with budget constraints, GPT-5.5 offers the best cost-to-capability ratio.

The Question Benchmarks Cannot Answer

Import AI #460, published on June 8 — one day before Fable 5 launched — included an observation from Jack Clark, Anthropic co-founder: code merged into Anthropic's internal codebase in 2026 increased eightfold compared to the 2021–2024 period, with acceleration beginning in 2025.

Clark was deliberate in his framing: this is not maximalist recursive self-improvement — there is no model autonomously designing its own successor. What exists is a quantifiably documented productivity acceleration at the lab level, which Clark describes as potentially "existential" in importance.

The missing piece, according to Clark himself, is whether that acceleration includes sufficient creativity for "paradigm-shifting ideas," or only efficient execution of known tasks.

That is the question Fable 5, as a product, does not answer. But it makes it more urgent to ask.

For Developers: What Changes in the API

If you are migrating from Claude Opus 4.8 or from Claude Mythos Preview, there are concrete API changes that affect existing integrations.

The most important change: thinking: {"type": "disabled"} is no longer valid.

In Fable 5 and Mythos 5, adaptive thinking is mandatory. Sending this parameter will return an error. The replacement is the effort parameter:

# BEFORE — Opus 4.8 with optional extended thinking
response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=8096,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000
    },
    messages=[{"role": "user", "content": prompt}]
)

# NOW — Fable 5 with effort parameter
response = client.messages.create(
    model="claude-fable-5",
    max_tokens=32000,
    effort="high",  # options: "low" | "medium" | "high"
    messages=[{"role": "user", "content": prompt}]
)

Refusal handling: the new stop_reason: "refusal".

When a classifier rejects a request, the API returns HTTP 200 with stop_reason: "refusal" — not an HTTP error. If your code only checks for stop_reason: "end_turn" or "max_tokens", it may silently ignore refusals. Correct handling:

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    fallbacks=["claude-opus-4-8"],  # automatic fallback (beta)
    messages=[{"role": "user", "content": prompt}]
)

if response.stop_reason == "refusal":
    print(f"Request rejected. Classifier: {response.stop_details}")
    # System automatically fell back to Opus 4.8 if `fallbacks` was configured
else:
    print(response.content)

The fallbacks parameter and fallback credit.

Including fallbacks=["claude-opus-4-8"] instructs the API to automatically retry a rejected request on the specified model. Anthropic provides a fallback credit that reimburses the prompt cache cost of the model switch, preventing double billing for the same context.

Thinking output: receiving summarized reasoning.

By default, thinking blocks are empty. To inspect the reasoning process:

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    thinking={"display": "summarized"},  # returns readable reasoning summary
    messages=[{"role": "user", "content": prompt}]
)

for block in response.content:
    if block.type == "thinking":
        print("Reasoning:", block.thinking)  # human-readable summary
    elif block.type == "text":
        print("Response:", block.text)

Task budgets (beta): explicit resource limits per task.

For long-horizon agentic tasks, the new task-budgets parameter allows setting token limits at the task level. Activate it with the task-budgets-2026-03-13 header:

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=128000,
    extra_headers={"task-budgets-2026-03-13": "true"},
    messages=[{"role": "user", "content": prompt}]
)

Migration guide from Opus 4.8: The official documentation includes a step-by-step migration guide covering parameter changes, new stop_reasons, and how to handle thinking output in multi-turn conversations.

What This Means in Concrete Terms

Claude Fable 5 is the most capable coding model available to the public as of June 10, 2026. SWE-bench Pro data confirms this with a 21-point lead over GPT-5.5. That capability has a literal price: double the cost of its closest competitor, safety filters that create documented friction on legitimate use cases, and a data retention policy incompatible with regulated industries.

For teams building software agents, automated development pipelines, or code analysis tooling, Fable 5 offers documented capabilities that justify evaluation. For scientific research or multimodal use cases, Gemini 3.1 Pro competes at a fraction of the cost.

Project Glasswing and Claude Mythos 5 represent a different kind of bet: the hypothesis that some capabilities should not be publicly available yet, and that the right way to manage that situation is controlled distribution with government oversight. Whether that hypothesis reflects genuine safety reasoning or is institutional scarcity framing with government backing is something the currently available data does not conclusively answer.

What is verifiable: the model exists, it performs, it has documented limitations, and it launched four days after the same company warned that AI is advancing faster than society can adapt.

That contradiction is the real story.

Tincho Fuentes — Tech journalist and investigative researcher 🚀

Sources: Anthropic official documentation · Project Glasswing · Simon Willison's analysis · The Decoder: full critical review · TechCrunch: launch context · Import AI #460 · AWS Blog · GitHub Copilot Changelog