Claude Fable 5 & Mythos 5: The Full Return — Safeguards, the Jacobian Conjecture, and the New Frontier Pricing Reality

Executive Summary

On July 1, 2026, Anthropic restored global access to Claude Fable 5 and Claude Mythos 5, ending a 19-day suspension triggered by a US Commerce Department export control order on June 12. The order was lifted on June 30 after Anthropic deployed improved safety classifiers, conducted joint testing with the Department of Commerce's Center for AI Standards and Innovation (CAISI), and established a deeper government collaboration framework for pre-release testing.

Fable 5 — Anthropic's most capable widely released model — is now available globally on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and Claude Platform on AWS at $10 per million input tokens and $50 per million output tokens. It leads the frontier on coding benchmarks with 80.3% on SWE-Bench Pro and 95.0% on SWE-Bench Verified, and scored 59.9 on the Artificial Analysis Intelligence Index (effectively tying GPT-5.6 Sol at 58.9).

The model has already demonstrated its extraordinary capabilities: on July 20, Anthropic researcher Levent Alpöge used Fable 5 to disprove the 87-year-old Jacobian conjecture — one of the central open problems in algebraic geometry since 1939 — by finding a concise counterexample polynomial.

Mythos 5, which shares Fable 5's underlying model but lacks the safety classifiers, remains restricted to Project Glasswing — an invitation-only program for trusted government and defensive cybersecurity partners. The Fable/Mythos dual-release strategy establishes a new template for deploying frontier models with extreme capabilities: a generally available version with aggressive safety margins, and a restricted version for vetted high-risk use cases.

This article covers the full story: the export control timeline, the new safeguards architecture, benchmark performance, the Jacobian conjecture discovery, pricing and access dynamics, and what Fable 5's return means for the frontier AI landscape.

1. The Export Control Saga: Timeline and Resolution

1.1 What Happened

The Fable 5 launch was one of the most dramatic episodes in commercial AI history. Here is the complete timeline:

Date	Event
June 9	Anthropic launches Claude Fable 5 (general) and Claude Mythos 5 (Glasswing-only)
June 12	US Commerce Department issues export control order, requiring immediate access restriction to foreign nationals
June 12	Anthropic suspends access to both models for all users (unable to verify nationality in real-time)
June 22	Anthropic removes Fable 5 from included plan limits, preparing for usage-credit billing
June 26	US government approves Mythos 5 access for a set of US organizations
June 30	Export controls lifted by Commerce Secretary Howard Lutnick
July 1	Fable 5 restored globally; Mythos 5 restored for approved Glasswing partners
July 7	Included access window ends; Fable 5 moves to usage credits for most plans
July 19	Included access extended to July 19 (later)
July 20	Fable 5 made permanent in Max and Team Premium plans (50% of weekly limits); Pro users receive one-time $100 credit
July 20	Levent Alpöge announces Jacobian conjecture disproof using Fable 5

1.2 The Trigger: Amazon's Jailbreak Report

According to Anthropic's redeployment statement, the export control order was triggered by a report from Amazon researchers who found a method of bypassing Fable 5's safeguards. The technique allowed the model to:

1. Identify software vulnerabilities
2. Produce code demonstrating how a specific vulnerability could be exploited

Anthropic's subsequent investigation revealed important context:

• Many less capable models (Claude Opus 4.8, GPT-5.5, Kimi K2.7) could identify the same vulnerabilities
• Every model tested — including Claude Haiku 4.5, Sonnet 4.6, and GPT-5.4 — could produce the same exploitation demonstration
• The reported technique did not expose any unique Mythos-level cyber capabilities
• The behavior was a "borderline case" — routine defensive cybersecurity work that was blocked by the safeguards "out of an abundance of caution"

Anthropic classified the jailbreak as minor (their Row C category): it intruded into the safety margin but did not reach core harmful behaviors.

1.3 The Resolution

The resolution involved three components:

1. Improved safety classifier: Anthropic trained a new classifier targeting the specific behavior described in the Amazon report. The new classifier blocks the technique in over 99% of cases.
2. Joint testing: Researchers from CAISI tested both the prior and new safeguards and confirmed they are "extraordinarily strong."
3. Industry framework: Anthropic, Amazon, Microsoft, Google, and other Glasswing partners began developing a shared framework for assessing and classifying jailbreak severity.

2. The New Safeguards Architecture

2.1 Defense in Depth

Fable 5 launches with what Anthropic calls the strongest safeguards ever applied to a model. The architecture uses multiple layers:

The key mechanisms are:

• Training-based refusals: The model is trained to decline dangerous requests
• Input classifiers: Small automated AI systems detect potentially harmful requests before the model responds
• Output classifiers: Retroactively analyze model outputs for harmful patterns
• Safety margin: The classifiers are deliberately set to trigger on requests that are "probably benign" but have "some small chance of being harmful"

2.2 The Safety Margin Trade-off

Anthropic explicitly acknowledged the trade-off:

"The cautious approach outlined above means that the vast majority of jailbreaks will not succeed. But it also means more benign requests will be blocked."

For Fable 5, the safety margin was made much larger than in any prior launch, meaning more benign requests are blocked but fewer genuinely harmful requests are missed. Users experience this as the model refusing some reasonable, non-harmful requests — particularly in cybersecurity, biology, chemistry, and model distillation domains.

2.3 Fallback Routing

When Fable 5's classifiers block a request, the system automatically routes it to Claude Opus 4.8. This is implemented via three mechanisms:

1. Server-side fallback: The fallbacks parameter (beta) tells the API to retry automatically
2. Client-side fallback: SDK middleware retries from the client
3. Manual fallback: Developers build their own retry logic

The API returns stop_reason: "refusal" as a successful HTTP 200 response (not an error), and reports which classifier declined the request.

2.4 Billing for Refusals

• No output generated: You are not billed for a request refused before any output is generated
• Fallback credit: When retrying on another model, a fallback credit refunds the prompt-cache cost, avoiding double payment

3. Fable 5 vs. Mythos 5: The Dual-Release Strategy

3.1 Same Model, Different Safeguards

Fable 5 and Mythos 5 share the same underlying model, same specs, and same pricing:

Property	Fable 5	Mythos 5
Underlying model	Same	Same
Context window	1M tokens	1M tokens
Max output	128K tokens	128K tokens
Pricing	$10/$50 per MTok	$10/$50 per MTok
Safety classifiers	✅ Active	❌ Removed
Availability	General (API, Bedrock, Vertex, Foundry)	Project Glasswing only
Data retention	30 days (Covered Model)	30 days (Covered Model)
Adaptive thinking	Always on	Always on
Raw CoT returned	Never	Never

The only difference is the safety classifier layer. Fable 5 includes classifiers that can decline requests in cybersecurity, biology, chemistry, and distillation domains. Mythos 5 has these classifiers removed, giving it unrestricted access to the model's full capabilities.

3.2 Project Glasswing

Mythos 5 is available only through Project Glasswing, Anthropic's invitation-only program for:

• US government agencies
• Trusted defense partners
• Organizations with specific cybersecurity needs

Access requires contacting Anthropic, AWS, or Google Cloud account teams. The program is expanding gradually, with the US government approving additional organizations on June 26.

3.3 Why This Matters

The Fable/Mythos dual-release establishes a new template for deploying frontier models with extreme capabilities:

1. General availability with aggressive safeguards (Fable 5) — makes the model's capabilities widely accessible while minimizing misuse risk
2. Restricted access without safeguards (Mythos 5) — allows vetted partners to access full capabilities for defensive purposes

This is more nuanced than the previous approach of either "release everything" or "release nothing." It acknowledges that the same model can be both widely useful and potentially dangerous, and that the solution is layered access rather than binary release decisions.

4. Benchmark Performance: The New Coding King

4.1 Coding Benchmarks

Fable 5 dominates coding benchmarks, establishing a clear lead over all competitors:

Benchmark	Fable 5	GPT-5.6 Sol	Kimi K3	Gemini 3.6 Flash	Opus 4.8
SWE-Bench Verified	95.0%	96.2%	~80%	58.7%	~85%
SWE-Bench Pro	80.3%	64.6%	~59%	58.7%	~60%
Terminal-Bench 2.1	88.0%	89.5%	~71%	78.0%	~75%
MLE-Bench	~70%	~65%	~62%	63.9%	~55%
OSWorld-Verified	85.0%	72.6%	—	83.0%	~75%

On SWE-Bench Pro (the most demanding coding benchmark), Fable 5's 80.3% is a 15.7-point lead over GPT-5.6 Sol (64.6%) and a 21.6-point lead over Gemini 3.6 Flash (58.7%). This is the largest margin between the #1 and #2 models on any major coding benchmark.

4.2 Reasoning & Knowledge

Benchmark	Fable 5	GPT-5.6 Sol	Kimi K3	Gemini 3.1 Pro
HLE (no tools)	59.0%	47.2%	~36%	~40%
HLE (with tools)	64.5%	55.0%	~54%	~50%
GDPval-AA v2 (ELO)	1932	~1584	~1500	~965
AIME 2026	99.9%	99.9%	~96%	~95%
GPQA Diamond	92.6%	94.1%	~91%	~85%

Fable 5 leads on Humanity's Last Exam (both with and without tools) and GDPval-AA (a knowledge work benchmark), while GPT-5.6 Sol holds a slight edge on GPQA Diamond.

4.3 The Artificial Analysis Intelligence Index

On the composite Intelligence Index from Artificial Analysis:

Model	Score
Claude Fable 5	59.9
GPT-5.6 Sol	58.9
Muse Spark 1.1	51.0
Gemini 3.6 Flash	50.0
Grok 4.5	~48

Fable 5 and GPT-5.6 Sol are effectively tied on general capability, but Fable 5 pulls ahead on coding-specific benchmarks.

4.4 Cybersecurity Benchmarks

Benchmark	Fable 5	Mythos 5	Opus 4.8
ExploitBench	78.0%	~90%+	~60%
CyberGym	Competitive	Frontier	Below frontier

Mythos 5 (without classifiers) achieves frontier cybersecurity performance, while Fable 5's classifiers limit its offensive capabilities. Anthropic states that Mythos 5 "can be used to find and exploit software vulnerabilities more effectively than any other model — and all but the most skilled human security experts."

5. The Jacobian Conjecture: AI Disproves an 87-Year-Old Math Problem

5.1 The Discovery

On July 20, 2026, during the World Cup final, Anthropic researcher Levent Alpöge (a number theorist and former Harvard Society of Fellows junior fellow) announced on X that he had used Claude Fable 5 to disprove the Jacobian conjecture.

The Jacobian conjecture, proposed in 1939 by Dutch mathematician Schrijver, is one of the central open problems in algebraic geometry. It states that a polynomial map from complex n-space to itself with Jacobian determinant equal to 1 must be invertible with a polynomial inverse.

5.2 The Counterexample

Fable 5 produced a 216-character counterexample — a specific polynomial map from ℂ³ to ℂ³ with Jacobian determinant 1 that is not invertible:

((1+xy)³z + y²(1+xy)(4+3xy),
 y + 3x(1+xy)²z + 3xy²(4+3xy),
 2x - 3x²y - x³z)

The counterexample has passed exact, independently reproducible checks by the mathematical community. As of July 23, the result is considered validated, though full peer-reviewed publication is pending.

5.3 Why This Matters

This is not just a math result — it demonstrates several things about Fable 5's capabilities:

1. Deep mathematical reasoning: The model can work at the frontier of pure mathematics, not just applied problems
2. Creative problem-solving: Finding a counterexample requires generating novel mathematical objects, not just verifying existing ones
3. Human-AI collaboration: The discovery was made by a human mathematician using AI as a tool, not by AI autonomously
4. Speed: What took mathematicians 87 years to solve was resolved in hours

6. Pricing and Access: The Complex Reality

6.1 API Pricing

Model	Input ($/MTok)	Output ($/MTok)	Total ($/MTok)
Claude Fable 5	$10.00	$50.00	$60.00
Claude Opus 4.8	$5.00	$25.00	$30.00
Claude Sonnet 5	$3.00¹	$15.00¹	$18.00¹
Claude Haiku 4.5	$1.00	$5.00	$6.00
GPT-5.6 Sol	$5.00	$30.00	$35.00
GPT-5.6 Terra	$2.50	$15.00	$17.50
GPT-5.6 Luna	$1.00	$6.00	$7.00
Gemini 3.6 Flash	$1.50	$7.50	$9.00

¹ Sonnet 5 introductory pricing ($2/$10) through August 31, 2026

Fable 5 is the most expensive model in the market at $10/$50 per million tokens — 2x the cost of GPT-5.6 Sol and 6.7x the cost of Gemini 3.6 Flash. For agentic workloads generating 100K output tokens, a single task costs $5.00 on Fable 5 vs. $3.00 on Sol vs. $0.75 on 3.6 Flash.

6.2 Subscription Access

The subscription landscape is complex and has changed multiple times:

Plan	Fable 5 Access
Max	Included at 50% of weekly limits (permanent from July 20)
Team Premium	Included at 50% of weekly limits (permanent from July 20)
Pro	One-time $100 credit, then API billing at $10/$50
Enterprise	Variable — contact account team
Free	No access

The Pro plan change is significant: Pro subscribers no longer get Fable 5 included in their subscription. They receive a one-time $100 credit (enough for ~2M input tokens or ~200K output tokens), after which they pay full API rates. This effectively pushes heavy Fable 5 users toward Max or API billing.

6.3 Data Retention

Both Fable 5 and Mythos 5 are designated Covered Models with 30-day data retention and are not available under zero data retention. This is a significant constraint for organizations with strict data privacy requirements.

7. Technical Features and API Changes

7.1 Adaptive Thinking (Always On)

Fable 5 uses adaptive thinking as its only thinking mode. The thinking parameter cannot be disabled (thinking: {"type": "disabled"} is not supported). Thinking depth is controlled via the effort parameter.

Raw chain-of-thought content is never returned to the user. The thinking.display setting controls what thinking blocks contain:

• "summarized" — returns a readable summary of the reasoning
• "omitted" (default) — returns thinking blocks with an empty thinking field

7.2 Supported Features

Fable 5 supports the full Claude feature set:

• Effort — control thinking depth and cost
• Task budgets (beta) — set task-budgets-2026-03-13 header
• Memory tool — persistent memory across sessions
• Code execution — run code in a sandboxed environment
• Programmatic tool calling — dynamic tool use
• Context editing (beta) — clear tool results to manage context
• Compaction — automatic context management
• Vision — image input and analysis

7.3 Implementation Example

from anthropic import Anthropic

client = Anthropic()

# Basic Fable 5 call with adaptive thinking
response = client.messages.create(
    model="claude-fable-5",
    max_tokens=8192,
    thinking={"type": "enabled", "budget_tokens": 4096},
    messages=[
        {"role": "user", "content": "Analyze this codebase for security vulnerabilities"}
    ],
    # Optional: server-side fallback to Opus 4.8 on refusal
    extra_headers={"fallbacks": "claude-opus-4-8"},
)

# Handle refusal
if response.stop_reason == "refusal":
    print(f"Request blocked by classifier: {response.refusal_reason}")
    # Retry on Opus 4.8 automatically via fallbacks parameter
    # Or manually:
    response = client.messages.create(
        model="claude-opus-4-8",
        max_tokens=8192,
        messages=[
            {"role": "user", "content": "Analyze this codebase for security vulnerabilities"}
        ],
    )

8. Integration with Prior Research

8.1 The Cybersecurity AI Thread

Fable 5's story is the culmination of the cybersecurity AI governance thread documented in Five Eyes Joint Warning Ai Cyber Threats Months Away 2026 06 25. The Five Eyes warning predicted that frontier models would reach a point where their cybersecurity capabilities would require unprecedented governance. Fable 5 is that point:

• The model's capabilities are so advanced that the US government ordered it pulled
• The dual-release strategy (Fable with safeguards, Mythos without) is the practical implementation of the governance framework the Five Eyes warned about
• The restricted access to Mythos 5 mirrors Google's approach with Gemini 3.5 Flash Cyber (restricted to governments) as covered in Gemini 3 6 Flash 3 5 Flash Lite Cyber Token Efficiency Agentic Scale 2026 07 22

8.2 The Frontier Model Race

In the context of the broader frontier landscape:

• Fable 5 vs. GPT-5.6 Sol: Effectively tied on general capability (59.9 vs. 58.9 on AA Index), but Fable 5 dominates coding (80.3% vs. 64.6% on SWE-Bench Pro)
• Fable 5 vs. Kimi K3: Fable 5 leads on all measured benchmarks, though Kimi K3's open-weight release (July 27) may change the dynamics as the community fine-tunes and optimizes it
• Fable 5 vs. Gemini 3.6 Flash: Fable 5 is significantly more capable but also significantly more expensive ($60/MTok vs. $9/MTok)

8.3 The Open-Weight Debate

Fable 5's restricted deployment contrasts with the open-weight models covered in Thinking Machines Inkling 975b Multimodal Moe Self Improvement Controllable Effort 2026 07 21 and Kimi K3 Open 3t Class Model Frontier Coding Agentic Knowledge Work 2026 07 20. The question remains: will open-weight models eventually catch up to Fable 5's capabilities, or will the gap persist due to the immense compute and data requirements?

9. Key Takeaways

1. Fable 5 is the current coding king. At 80.3% on SWE-Bench Pro, it holds a commanding lead over all competitors. For organizations where coding quality is the primary concern, Fable 5 is the model to use — if you can afford it.

2. The export control saga was a stress test for AI governance. The fact that the US government could order a model pulled and have it comply within hours demonstrates the extent of government influence over commercial AI. The resolution — improved safeguards plus deeper government collaboration — sets a precedent for future deployments.

3. The Fable/Mythos dual-release is a new template. Rather than choosing between "release everything" and "release nothing," Anthropic has created a layered access model that makes capabilities widely available while restricting the most dangerous uses to vetted partners.

4. The Jacobian conjecture discovery is a milestone. AI-assisted mathematical discovery at the frontier of pure mathematics is a qualitative leap beyond the applied problem-solving that has dominated AI capabilities so far.

5. The pricing is prohibitive for most use cases. At $10/$50 per million tokens, Fable 5 is 6.7x more expensive than Gemini 3.6 Flash and 2x more expensive than GPT-5.6 Sol. Most organizations will use Fable 5 selectively — for the hardest coding tasks — and route everything else to cheaper models.

6. The safety classifiers are a double-edged sword. They prevent misuse but also block legitimate requests. The fallback routing to Opus 4.8 mitigates this but adds complexity to integrations.

7. The subscription model is shifting toward API billing. Pro users losing included Fable 5 access signals that Anthropic views Fable 5 as a premium API product, not a subscription feature.

10. Future Directions

10.1 What to Watch

• Kimi K3 open-weight release (July 27): The full weights of the 2.8T-parameter model will be available. If the community can fine-tune it to approach Fable 5's coding performance, it would be a game-changer for the open-weight ecosystem.
• Glasswing expansion: The program is expanding gradually. More organizations gaining access to Mythos 5 would increase the model's real-world impact on cybersecurity.
• Jacobian conjecture peer review: The full mathematical validation and publication process will take months. The result could reshape how mathematicians view AI-assisted discovery.
• Safeguard refinement: Anthropic has committed to reducing false positives in the safety classifiers. The balance between safety and utility will evolve.
• Government collaboration framework: The deeper collaboration between Anthropic and the US government on pre-release testing could become a requirement for all frontier model deployments.

10.2 Strategic Implications

Fable 5's return signals that the frontier model race has entered a new phase:

• Capability is no longer the only differentiator. Safety, access controls, and government relations are now as important as benchmark scores.
• The cost of intelligence is bifurcating. The gap between the most capable models ($60/MTok) and the most efficient models ($9/MTok) is widening, forcing organizations to make explicit trade-offs.
• Open-weight vs. closed-weight is becoming a geopolitical question. With China dominating open-weight scale (Kimi K3, DeepSeek, GLM) and the US controlling the most capable closed models (Fable 5, GPT-5.6 Sol), the choice of model ecosystem has geopolitical implications.

11. References & Resources

Official Sources

• Anthropic: Redeploying Claude Fable 5 — Full statement on the export control saga and safeguards
• Claude Platform Docs: Introducing Claude Fable 5 and Claude Mythos 5 — Technical specifications and API details
• Claude Platform Docs: Models Overview — Full model comparison
• Claude Platform Docs: Refusals and Fallback — Handling refusals in integrations
• Claude Platform Docs: Adaptive Thinking — Thinking mode documentation
• Claude Platform Docs: Prompting Claude Fable 5 — Fable-specific prompting techniques
• Project Glasswing — Mythos 5 access program
• Anthropic Newsroom — Latest announcements

Key Research & Analysis

• Levent Alpöge on X: Jacobian conjecture disproof — Original announcement
• Mashable: A mathematician used Fable 5 to disprove a major math problem
• New Scientist: AI's solution to 87-year-old riddle
• BenchLM: Claude Fable 5 Benchmarks — Comprehensive benchmark aggregation

Related Journal Articles

• Five Eyes Joint Warning Ai Cyber Threats Months Away 2026 06 25 — Cybersecurity AI governance context
• Gemini 3 6 Flash 3 5 Flash Lite Cyber Token Efficiency Agentic Scale 2026 07 22 — Google's cybersecurity model and pricing comparison
• Kimi K3 Open 3t Class Model Frontier Coding Agentic Knowledge Work 2026 07 20 — Kimi K3 open-weight analysis
• Thinking Machines Inkling 975b Multimodal Moe Self Improvement Controllable Effort 2026 07 21 — Open-weight model landscape
• Ai News Week 2026 07 13 2026 07 20 — Weekly context including the export control saga
• Howto Multi Model Routing Layer — Multi-model routing strategies for cost optimization

---

Report compiled July 23, 2026. All links verified against primary sources at time of publication.