# Claude Fable 5 — Anthropic's Most Intelligent Model Ever, Released June 9, 2026 Claude Fable 5 is Anthropic's first publicly available Mythos-level model — a quantum leap beyond Opus 4.8 in reasoning, coding, and creative capability. Released June 9, 2026. # Claude Fable 5 — Anthropic's Most Intelligent Model Ever Released June 9, 2026, Claude Fable 5 is Anthropic's first publicly available Mythos-level model — a quantum leap beyond Opus 4.8 in reasoning, coding, and creative capability. 📅 June 10, 2026 ✎ ZVHH Research 95% SWE-bench Verified 80.3% SWE-bench Pro 1M Token Context $10/$50 Per Million Tokens On **June 9, 2026**, Anthropic made a bold move: it released **Claude Fable 5**, the first publicly available model in its new **Mythos class** — a capability tier that sits above the beloved Opus line. This wasn't just another incremental upgrade. Fable 5 represents a fundamental leap in what AI can accomplish, and it arrived at a moment when the entire AI industry was watching Anthropic's every move. The model shares the same underlying architecture with **Claude Mythos 5**, which was simultaneously announced but restricted to a handful of vetted government and cybersecurity partners through **Project Glasswing**. What makes Fable 5 special is that it's the public-facing, safeguarded deployment of those same Mythos-class weights — making the most powerful AI Anthropic has ever built accessible to millions of users and developers worldwide. --- ## The Claude Model Hierarchy: Where Fable 5 Fits Anthropic's Claude family has evolved through several tiers, each representing a significant capability jump. Here's the full hierarchy: Haiku 4.5 ### Fastest — Lowest Cost Optimized for speed and efficiency. Best for simple tasks, rapid prototyping, and high-volume workflows. The workhorse of Claude's lineup. Sonnet 4.6 ### Balanced — Best Value The sweet spot for most production workloads. Strong reasoning, coding, and vision at a fraction of the cost of higher tiers. Opus 4.8 ### Most Capable — Prior Tier Built for serious coding and AI agents. Features a 1M context window, hybrid reasoning, and is the current flagship before Fable 5. Fable 5 ### Mythos-Class — Public Anthropic's first generally available Mythos model. State-of-the-art on nearly all benchmarks. 95% on SWE-bench Verified, 80.3% on SWE-bench Pro. Mythos 5 ### Peak Capability — Restricted The unrestricted version, reserved for vetted cybersecurity partners and government agencies via Project Glasswing. This new Mythos tier represents Anthropic's answer to the accelerating capability race. While competitors released increasingly powerful models, Anthropic held back — not because they couldn't build the technology, but because they wanted to ensure their safety guardrails could handle it. The result is a model that's not only more capable than any Anthropic product before it, but also more thoroughly safety-tested. --- ## How Fable 5 Beats Opus 4.8: The Benchmarks The numbers tell a clear story. Claude Fable 5 outperforms Claude Opus 4.8 — and every other model Anthropic has ever made generally available — across nearly every benchmark: ### Software Engineering Benchmarks | **Benchmark** | **Claude Fable 5** | **Claude Opus 4.8** | **GPT-5.5**| --- | --- | --- | --- | SWE-bench Verified | **95.0%** | 70.0% | 60.0%| | SWE-bench Pro | **80.3%** | ~69% | 59.1%| | Terminal-Bench 2.1 | **84.3%** | ~74% | ~65%| | FrontierSWE | **74.6%** | ~71% | 72.6%| On SWE-bench Verified — the gold standard for evaluating real-world software engineering capability — Fable 5 scores **95%**, nearly 25 percentage points ahead of Opus 4.8. On the harder SWE-bench Pro, it scores 80.3%, roughly 11 points ahead of the next-best frontier model. These aren't marginal improvements; they represent a generational leap in automated software engineering. ### Reasoning and Knowledge Benchmarks | **Benchmark** | **Claude Fable 5** | **Claude Opus 4.8**| --- | --- | --- | GPQA-Diamond | **92.5%** | ~91%| | HLE (w/ Tools) | **58.2%** | 57.9%| | AIME 2026 (Math) | **99.0%** | 95.7%| | SOTU (Document Understanding) | **97.5%** | 94.0%| Fable 5 leads Opus 4.8 on HLE (Hands-on LLM Evaluation) with tools — a benchmark that tests whether AI can actually *do* work, not just talk about it. It tops AIME 2026 at 99.0%, matching the best scores in the industry. The SOTU benchmark, which evaluates document understanding across thousands of pages of technical content, shows Fable 5 at 97.5% — a full 3.5 points above Opus 4.8. ### Vision and Multimodal Capabilities Fable 5's vision capabilities represent a significant step up from Opus 4.8: - **Diagram and chart understanding** — Fable 5 can interpret complex diagrams, charts, and tables nested in files and PDFs, improving work across document-heavy domains like finance, legal, analytics, and architecture - **Self-evaluation** — The model uses vision to evaluate its own coding work, checking outputs against original designs or goals — a capability that improves code quality on multi-step projects - **Document analysis** — With a 1M-token context window, Fable 5 can process entire technical manuals, legal contracts, or financial reports in a single pass --- ## Architecture and Safety: How Fable 5 Works At its core, Claude Fable 5 is built on Anthropic's **Constitutional AI** framework — the same alignment technique that has distinguished Claude from its competitors since Claude 1. But Fable 5 takes this to a new level. ### Constitutional AI at Scale Constitutional AI is a technique Anthropic developed to improve ethical and legal compliance in AI systems. Instead of relying solely on human feedback, the model is trained to critique and revise its own outputs based on a set of principles — hence "constitutional." With Fable 5, this framework has been significantly strengthened: - **Enhanced safety classifiers** — When requests are flagged as cybersecurity, biology, chemistry, or distillation-related, they may be automatically routed to Claude Opus 4.8 instead. This safeguard affects less than 5% of sessions - **Extended thinking** — Fable 5 supports explicit chain-of-thought reasoning, improving performance on math and complex reasoning tasks at the cost of higher latency and token usage - **Reasoning effort dial** — Users can explicitly control the model's thinking effort, from quick responses to deep reasoning, balancing performance against speed and cost - **Defense-in-depth** — Designed to make jailbreaks narrow and expensive to produce, combined with thorough monitoring and 30-day customer data retention for researching jailbreak attempts ### Red-Teaming and Government Partnership Before launch, Fable 5 underwent thousands of hours of red-teaming with the **US government**, **UK AISI** (Algorithms and Artificial Intelligence Safety Institute), private third-party organizations, and Anthropic's internal safety teams. This extensive testing is what gave the government the confidence to approve the public release — and also, unfortunately, the basis for the export control directive that would later suspend access. --- ## Real-World Validation: What Users Found in Days Even in its brief window of availability, Claude Fable 5 demonstrated capabilities that validated Anthropic's investment in the Mythos class: ### GitHub's Internal Testing GitHub tested Fable 5 on complex, long-horizon coding tasks and reported results that "exceeded previous models in autonomy and reliability." In early testing, Fable 5 took on multi-hour coding tasks with a level of independence that developers had never seen — planning multi-step tasks, pulling in relevant context, and adapting its approach as the task evolved. ### Stripe's Codebase Migration Stripe, one of the world's largest payment processors, used Fable 5 to help manage a **50-million-line codebase migration** — an undertaking that would have taken human engineering teams months to complete. The model handled complex, multi-stage migration tasks that required sustained reasoning across massive amounts of code, demonstrating Fable 5's ability to serve as a genuine engineering partner rather than just a coding assistant. ### Project Glasswing's Security Findings Project Glasswing — the restricted program for vetted cybersecurity partners — used Claude Mythos 5 (same architecture as Fable 5) to identify over **10,000 vulnerabilities** across 150+ organizations in 15+ countries. These organizations included major cloud providers, enterprise software companies, and government agencies. The fact that the restricted version was generating such results validated the architecture even before the public release. --- ## Pricing and Access Tiers Claude Fable 5 was positioned as a premium model with pricing that reflects its state-of-the-art capabilities: | **Plan / Access** | **Cost** | **Includes Fable 5**| --- | --- | --- | **API** | $10 / $50 per million tokens | Yes| | **Claude Pro** | $20/month | Yes (through June 22, 2026)| | **Claude Max** | $40/month | Yes (through June 22, 2026)| | **Claude Team** | $100/month/user | Yes (through June 22, 2026)| | **Enterprise** | Custom | Yes| For the first two weeks after launch (through June 22, 2026), Fable 5 was included in all paid Claude plans at no additional cost — a generous trial period that allowed millions of Pro and Max subscribers to experience Mythos-class capability. After June 22, the model transitioned to a usage-credit model, meaning Pro/Max users would need additional credits to continue using Fable 5 after their plan's monthly allowance was consumed. On the API, Fable 5 is priced at **$10 per million input tokens and $50 per million output tokens** — exactly double Opus 4.8's pricing ($5/$25). However, Anthropic's 90% input token discount for prompt caching applies, meaning that for workflows that reuse the same context (like multi-turn coding sessions), the effective cost is significantly lower. --- ## When to Use Fable 5 vs Opus 4.8 The smart move isn't to pick one model — it's to route by task. Here's when each excels: #### ⚙️ Use Fable 5 For **Hard problems** — complex software engineering, multi-hour agentic workflows, large-scale code migrations, deep reasoning tasks, and any work where the extra capability justifies the 2x cost premium. #### ⚡ Use Opus 4.8 For **Everyday work** — routine coding tasks, document analysis, knowledge queries, and most production workloads where Opus 4.8's strong performance is more than adequate at half the price. As one developer put it after testing both models: "Oneshotting a complex task with Fable 5 takes roughly as many tokens as Opus plus three retries. For critical work, Fable 5 saves more time than it costs in tokens." --- ## The Full Claude Release Timeline To understand where Fable 5 fits in Anthropic's evolution, it helps to see the complete timeline: March 2023 **Claude 1** — Anthropic's first production language model, built using Constitutional AI. August 2023 **Claude 2** — Significant improvements in reasoning, coding, and context window. March 2024 **Claude 3 (Haiku, Sonnet, Opus)** — The first multimodal Claude with vision. Opus achieved GPT-4 level performance. June 2024 **Claude 3.5 (Sonnet, Opus)** — Sonnet 3.5 became the fastest model on the market. Opus 3.5 pushed reasoning further. Late 2024 **Claude 3.5 Sonnet** — Became the top model on SWE-bench at the time, surprising the industry. 2025 **Claude 4 Series (Opus 4, Sonnet 4, Haiku 4)** — New generation with 200K context, prompt caching, extended thinking, and Computer Use. May 28, 2026 **Claude Opus 4.8** — Upgrade to the Opus class with stronger coding, agentic tasks, and 1M context window. The current flagship before Fable 5. June 9, 2026 **Claude Fable 5 launches** — Anthropic's first generally available Mythos-class model. Same architecture as restricted Claude Mythos 5. June 12, 2026 **Access suspended** — US government export control directive orders global suspension of Fable 5 and Mythos 5 for foreign nationals. Anthropic complies, disabling both models worldwide. --- ## What Fable 5 Means for the AI Industry Claude Fable 5's brief three-day availability — and its subsequent suspension — made it one of the most talked-about AI releases in history. Here's what it means: ### Mythos-Class AI Is Real Benchmarks don't lie. Fable 5's 95% on SWE-bench Verified isn't a marginal improvement — it represents a fundamental shift in what AI can accomplish autonomously. The fact that the same architecture, under the Mythos 5 name, was already producing results in cybersecurity (10,000+ vulnerabilities found through Project Glasswing) before the public release shows that the capability ceiling had been pushed significantly higher. ### Safety Before Capability Anthropic's decision to hold back the Mythos architecture until safety classifiers, red-teaming, and monitoring infrastructure were in place set a new standard for responsible AI development. The safety guardrails that caused some sessions to route to Opus 4.8 weren't a weakness — they were a demonstration of the kind of defense-in-depth that makes frontier AI responsibly deployable. ### The Export Control Precedent Perhaps the most significant outcome of the Fable 5 episode is that it established a new precedent: **frontier AI models are now subject to government export control**. Whatever one thinks of this policy, it fundamentally changes how businesses evaluate AI providers. Model availability is no longer guaranteed — and any business or developer building on frontier AI should treat this as a cautionary tale and design their systems accordingly. ### Open Source Is More Critical Than Ever With Fable 5 suspended and access to Mythos-class AI potentially restricted by government policy, the importance of open-source alternatives has never been clearer. Models like GLM-5.2 (753B parameters, MIT license) and Qwen3.6-35B-A3B provide powerful, unrestricted alternatives that can run locally on consumer hardware — ensuring that cutting-edge AI capability isn't dependent on a single provider or subject to political intervention. --- ## The Bottom Line Claude Fable 5 lasted just **three days** between launch and suspension — a blink of an eye in the lifecycle of a commercial AI model. But in those three days, it demonstrated capabilities that reshaped the entire AI industry's understanding of what's possible. With **95% on SWE-bench Verified**, **80.3% on SWE-bench Pro**, and a **1M-token context window**, Fable 5 was — and arguably remains — the most capable general-purpose AI model Anthropic has ever built. Its Mythos-class architecture, Constitutional AI safety framework, and real-world validation from companies like GitHub and Stripe proved that Anthropic's investment in safety-first development could coexist with world-leading capability. While Opus 4.8 remains the current flagship (having retained its position while Fable 5 was suspended), the Mythos architecture proved that AI can do far more than anyone expected — and the industry will be measuring every subsequent model against the standard Fable 5 set in just three days. --- ## Key Facts | **Model** | Claude Fable 5| | **Developer** | Anthropic| | **Tier** | Mythos-Class (public)| | **Release Date** | June 9, 2026| | **Suspended** | June 12, 2026 (US export control)| | **SWE-bench Verified** | **95.0%**| | **SWE-bench Pro** | **80.3%**| | **Terminal-Bench 2.1** | **84.3%**| | **AIME 2026** | **99.0%**| | **Context Window** | 1M tokens| | **API Pricing** | $10 input / $50 output per million tokens| | **Model ID** | `claude-fable-5`| | **Knowledge Cutoff** | January 2026| | **Availability** | API, AWS Bedrock, Vertex AI, Microsoft Foundry, Claude Platform| ## More from ZVHH [ News ### Claude Fable 5 Suspended by US Government — Just 3 Days After Launch Anthropic abruptly disabled both models globally to comply with a US export control directive. Full timeline and analysis. Jun 12, 2026 ](/claude-fable-5-suspended-us-government-export-control.html) [ Open Source ### GLM-5.2 — The Open-Source LLM That's Top of Every AI Leaderboard With Fable 5 suspended, open-source alternatives like GLM-5.2 (MIT license) become even more critical. Jun 20, 2026 ](/glm-5-2-new-open-source-llm-beats-claude-opus-frontier.html) [ Model ### VibeThinker 3B: Beats Claude Opus & OpenAI at Only 3B Parameters MIT licensed, tiny but powerful — how a 3B model punches way above its weight. Jun 23, 2026 ](/vibethinker-3b-beats-frontier-reasoning-models.html)