Key Takeaways
- Claude Fable 5 is Anthropic’s first publicly available Mythos-class model, released on June 9, 2026. It sits above Opus in Anthropic’s model hierarchy and is built specifically for demanding long-horizon agentic work, complex coding, large codebase analysis, vision tasks, and senior-level reasoning. It uses the same underlying model as Claude Mythos 5 but with stricter safeguards for high-risk domains including cybersecurity, biology, and chemistry.
- On SWE-Bench Pro, Anthropic’s agentic coding benchmark, Claude Fable 5 scores 80.3%, ahead of Opus 4.8 at 69.2% and GPT-5.5 at 58.6%. On SWE-bench Verified it achieves 95%. On Terminal-Bench 2.1 it scores 83.4%, trailing GPT-5.6 Sol’s 88.8% but leading every other model in the category. These benchmarks collectively make Fable 5 the strongest publicly available model for long-horizon software engineering tasks.
- Claude Fable 5 supports a 1 million token context window with up to 128K output tokens per call. This is sufficient to hold an entire large codebase in context simultaneously. In a documented real-world deployment, Stripe ran a 50-million-line Ruby codebase migration using Fable 5 in one day, a task estimated to take a full engineering team two or more months.
- API pricing is $10 per million input tokens and $50 per million output tokens. This makes Fable 5 the most expensive model in Anthropic’s lineup and roughly double the cost of GPT-5.6 Sol ($5 input / $30 output) for equivalent token volumes. For production use cases where cost per task is a primary constraint, GPT-5.6 Sol or Gemini 3.5 Pro are materially cheaper alternatives.
- Claude Fable 5 supports vision, tool use, memory, compaction, and adaptive thinking. The vision capability extends to reconstructing web application source code from screenshots alone, which has direct applications in legacy UI modernization and design-to-code workflows. The model cleared Pokémon FireRed start to finish using only raw game screenshots, a frequently cited demonstration of its visual reasoning depth.
- Access to Claude Fable 5 was suspended on June 12, 2026, three days after launch, due to export control compliance requirements. Access was restored globally on July 1, 2026, after the applicable controls were lifted. This access gap affected teams that had begun integrating Fable 5 into production pipelines immediately after the June 9 launch.
- Claude Fable 5 hallucinates less than GPT-5.6 Sol and Gemini 3.5 Pro on comparable tasks, according to third-party head-to-head evaluations. For use cases where factual accuracy, citation quality, and instruction-following on complex prompts are more important than raw speed or cost efficiency, Fable 5’s lower hallucination rate is a meaningful practical advantage.
Claude Fable 5 launched on June 9, 2026, and immediately posted the highest SWE-bench Verified score of any publicly available model at 95%, alongside an 80.3% result on SWE-Bench Pro. Those numbers matter because they are grounded in real software engineering tasks: fixing bugs in real open-source repositories, making changes that pass existing test suites, and completing work that a human engineer would review and merge. The benchmark scores are not theoretical; they reflect Fable 5’s behavior on the specific kind of complex, multi-file, real-codebase work that engineering teams actually need AI to handle.
This review covers Fable 5’s full capability set, the pricing reality, where it outperforms its alternatives, and where its cost makes cheaper models the more practical choice.
What Is Claude Fable 5?
Claude Fable 5 is Anthropic’s Mythos-class AI model, the first in that tier to be released for public API access. Mythos is a new model category above Opus in Anthropic’s hierarchy, designed for tasks that require sustained reasoning over very long inputs, multi-step autonomous execution, and high-stakes output quality where errors are costly.
The model uses the same underlying weights as Claude Mythos 5 but applies stricter safeguards in high-risk domains: cybersecurity, biology, chemistry, and model distillation are subject to additional controls that limit outputs that could enable harm even when the intent is legitimate. For most software engineering and knowledge work applications, these safeguards are not a practical constraint.
Fable 5 is available through Anthropic’s API with the model ID claude-fable-5 and through third-party providers including OpenRouter. Access was suspended from June 12 to July 1, 2026, due to export control compliance requirements; as of July 2026 it is available globally.
Claude Fable 5 Features
Long-Horizon Agentic Coding
Fable 5’s most significant capability is sustained, multi-step coding work across large codebases. The 1 million token context window allows the model to ingest an entire large repository in a single call, reason across files, identify interdependencies, plan changes, execute them, and validate against existing tests without losing context between steps. This is qualitatively different from models that handle individual files or short-context code completion: Fable 5 can reason about how a change in one module affects behavior in another module five directories away.
The Stripe codebase migration is the most striking documented example. Fable 5 completed a migration of a 50-million-line Ruby codebase in one day, a task Stripe estimated would require a full engineering team two or more months. The ability to hold the entire codebase context and execute coordinated changes across it simultaneously is what made this possible.
Benchmark Performance
On SWE-bench Verified (the canonical benchmark for AI software engineering), Fable 5 scores 95%, the highest reported score for a publicly available model. On SWE-Bench Pro, Anthropic’s internal agentic coding benchmark covering more complex real-world engineering tasks, it scores 80.3%, ahead of Opus 4.8 at 69.2% and GPT-5.5 at 58.6%. On Terminal-Bench 2.1, a benchmark specifically measuring autonomous terminal-based coding agents, Fable 5 scores 83.4%, trailing only GPT-5.6 Sol at 88.8%.
These scores place Fable 5 at or near the top of every major coding benchmark where publicly available results exist. The gap between Fable 5 and the previous Anthropic flagship (Opus 4.8) is particularly notable on SWE-Bench Pro: 80.3% versus 69.2% is an 11-point improvement on a benchmark that measures real engineering work rather than text generation.
Vision and Multimodal Capabilities
Fable 5 supports image input with a level of visual reasoning that extends beyond basic image description. The model can reconstruct web application source code from screenshots of the running application, identifying UI components, layout structure, and interactive elements from visual input alone. This has direct applications in legacy UI modernization (recreating code for interfaces where source has been lost), design-to-code workflows (building from Figma exports or design mockups), and visual debugging (identifying layout issues from rendered screenshots).
The Pokémon FireRed completion using only raw game screenshots is frequently cited as a demonstration of Fable 5’s visual reasoning depth. The model navigated the full game using only what it could see on the screen, without access to maps, walkthroughs, or structured game state data. The practical implication is that Fable 5 can operate in visual environments where structured programmatic access to state is unavailable.
1M Token Context and 128K Output
The 1 million token context window and 128K maximum output per call are the enabling infrastructure for Fable 5’s long-horizon capabilities. A 1M token context accommodates roughly 750,000 words of text, which is sufficient to hold a large enterprise codebase, a full legal document set, or an extensive research corpus in a single call. The 128K output limit allows the model to produce a complete, large deliverable (a refactored module, a detailed technical document, a comprehensive analysis) in a single response rather than requiring multiple calls to assemble.
Adaptive Thinking
Fable 5 includes adaptive thinking, a reasoning mode that extends the model’s internal chain-of-thought for complex problems before producing output. For software engineering tasks, this means the model can work through a multi-file refactoring plan, identify edge cases, and validate its approach before writing a single line of output code. The depth of thinking scales with problem complexity rather than applying a fixed reasoning budget to every task.
Claude Fable 5 Pricing
| Token type | Price per million tokens |
|---|---|
| Input | $10.00 |
| Output | $50.00 |
At $10 input and $50 output per million tokens, Fable 5 is Anthropic’s most expensive API model and one of the most expensive frontier models available. GPT-5.6 Sol costs $5 input and $30 output per million tokens. Gemini 3.5 Pro costs $2 input and $12 output. For teams running high-volume production workloads where token cost is a primary operational constraint, these price differences are significant over millions of calls.
For tasks where Fable 5’s quality advantage is decisive (large codebase migrations, complex multi-file refactoring, high-stakes document analysis), the cost premium is typically justified by the reduction in engineering time required to review and correct outputs. For routine tasks where a cheaper model produces acceptable results (short code completions, FAQ drafting, simple data extraction), using Fable 5 is cost-inefficient.
Claude Fable 5 Pros and Cons
Pros:
- Highest SWE-bench Verified score of any publicly available model at 95%
- 80.3% on SWE-Bench Pro, 11 points ahead of Opus 4.8
- 1 million token context window handles entire large codebases in a single call
- Vision capability extends to reconstructing web app source from screenshots
- Lower hallucination rate than GPT-5.6 Sol and Gemini 3.5 Pro in head-to-head evaluations
- 128K output tokens per call enables large complete deliverables without multi-call assembly
- Documented real-world codebase migration at scale (Stripe 50M-line Ruby migration)
Cons:
- $50 output per million tokens is roughly double GPT-5.6 Sol’s output rate
- Access was suspended June 12 to July 1, 2026, due to export control issues
- Terminal-Bench 2.1 score (83.4%) trails GPT-5.6 Sol (88.8%) on agentic terminal tasks
- Overkill and cost-inefficient for routine tasks where Sonnet 5 or GPT-5.6 Luna are adequate
- Stricter safeguards in cybersecurity and biology domains may limit legitimate research use cases
Claude Fable 5 vs Alternatives
Claude Fable 5 vs GPT-5.6 Sol: GPT-5.6 Sol leads on Terminal-Bench 2.1 at 88.8% versus Fable 5’s 83.4%, and costs roughly half as much per token ($5/$30 versus $10/$50). For agentic terminal tasks and cost-sensitive production pipelines, Sol is the more practical choice. Fable 5 leads on SWE-Bench Pro (80.3% versus 58.6% for GPT-5.5) and hallucinates less on complex reasoning tasks. For codebase-scale work and vision-heavy workflows, Fable 5 is the stronger choice despite the cost premium.
Claude Fable 5 vs Gemini 3.5 Pro: Gemini 3.5 Pro offers a 2 million token context window (double Fable 5’s 1M), native multimodal input across text, image, video, and audio, and substantially lower pricing at $2 input / $12 output per million tokens. For extremely large document analysis and mixed-media inputs, Gemini 3.5 Pro’s context and cost advantages are significant. Fable 5 leads on coding benchmarks (SWE-Bench Pro 80.3% versus Gemini 3.1 Pro’s 54.2%) and on factual accuracy for text-heavy reasoning tasks.
Claude Fable 5 vs Claude Opus 4.8: Fable 5 outperforms Opus 4.8 on every published coding benchmark, including an 11-point gap on SWE-Bench Pro (80.3% versus 69.2%). For teams already using Opus 4.8 for complex agentic coding tasks, Fable 5 is the natural upgrade. For non-coding use cases (writing, analysis, conversation), Opus 4.8 remains cost-effective at a lower price point, and the quality gap is narrower outside the coding domain.
Who Is Claude Fable 5 Best For?
Claude Fable 5 is best suited for engineering teams working on large-scale codebase migrations, modernization projects, complex multi-file refactoring, and agentic coding pipelines where the task requires sustained reasoning across a full repository. The Stripe migration example is the clearest signal: for work that would otherwise require multiple engineers over weeks or months, Fable 5’s context size and coding accuracy can compress timelines dramatically.
It is also well-suited for legal and financial knowledge work requiring senior-level analysis over very large document sets, and for vision-heavy workflows where source code reconstruction from visual input has value. Researchers and teams in regulated industries should note the stricter safeguards in cybersecurity, biology, and chemistry domains before building pipelines that touch those areas.
Claude Fable 5 is not cost-efficient for high-volume, low-complexity tasks. Teams using it for routine completions, short code snippets, or FAQ generation are paying the Mythos-class price for work that Claude Sonnet 5 or GPT-5.6 Luna handles adequately at a fraction of the cost. The right use pattern is reserving Fable 5 for the tasks where its specific advantages (codebase-scale context, 95% SWE-bench accuracy, vision reasoning) justify the premium.
Our Verdict
Claude Fable 5 is the most capable publicly available model for large-scale software engineering tasks. Its benchmark results are not incremental improvements on the previous generation: 95% on SWE-bench Verified and 80.3% on SWE-Bench Pro represent a meaningful step up in what AI can reliably complete on real engineering work. The Stripe migration example gives those numbers practical grounding.
The cost is real and the constraint is real. At $50 per million output tokens, Fable 5 is not a default model for every task. It is a specialized tool for work where quality and scope exceed what cheaper models can reliably deliver. For engineering teams facing large-scale migrations, complex refactoring, or codebase-wide architectural changes, Fable 5’s capability advantage typically justifies its cost. For everything else, a cheaper model in the same generation is the more practical choice.
Frequently Asked Questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s first publicly available Mythos-class AI model, released on June 9, 2026. It is the most capable model in Anthropic’s lineup, positioned above Opus in the model hierarchy. It is purpose-built for long-horizon agentic coding, large codebase analysis, complex reasoning, and vision tasks. It achieves 95% on SWE-bench Verified and 80.3% on SWE-Bench Pro, the highest scores for any publicly available model on those benchmarks.
How much does Claude Fable 5 cost?
Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens on Anthropic’s API. It is Anthropic’s most expensive model and roughly double the cost of GPT-5.6 Sol ($5/$30 per million tokens) for equivalent token volumes. Pricing is available through Anthropic’s API directly and through third-party providers including OpenRouter.
What is the context window for Claude Fable 5?
Claude Fable 5 supports a 1 million token context window with up to 128K output tokens per call. A 1M token context accommodates approximately 750,000 words, which is sufficient to hold a large enterprise codebase or extensive document set in a single call. The 128K output limit allows large complete deliverables to be produced in a single response.
Is Claude Fable 5 better than GPT-5.6 for coding?
For codebase-scale software engineering, Claude Fable 5 leads on SWE-Bench Pro (80.3% versus GPT-5.5’s 58.6%) and on SWE-bench Verified (95%). For agentic terminal-based coding tasks specifically, GPT-5.6 Sol leads on Terminal-Bench 2.1 at 88.8% versus Fable 5’s 83.4%. Fable 5 also hallucinates less than GPT-5.6 Sol on complex reasoning tasks and is better suited for work requiring very large context (1M tokens). GPT-5.6 Sol is faster and roughly half the cost per token, making it the more practical choice for high-volume coding pipelines where task complexity is moderate.
Why was Claude Fable 5 access suspended after launch?
Access to Claude Fable 5 was suspended on June 12, 2026, three days after its June 9 launch, due to export control compliance requirements. The suspension affected global API access while Anthropic resolved the applicable regulatory requirements. Access was restored globally on July 1, 2026. Teams that had begun integrating Fable 5 into production pipelines between June 9 and June 12 needed to handle this access gap in their deployment plans.
What are the safeguards in Claude Fable 5?
Claude Fable 5 applies stricter output controls than standard Claude models in several high-risk domains: cybersecurity (exploits, vulnerability research), biology (pathogen-related work), chemistry (hazardous synthesis), and model distillation (reproducing model weights or training data). These safeguards are more restrictive than in Opus 4.8 or Sonnet 5. For most software engineering, business analysis, legal research, and knowledge work use cases, these restrictions do not affect practical operation.
Can Claude Fable 5 understand images and screenshots?
Yes. Claude Fable 5 supports image input with advanced visual reasoning. It can reconstruct web application source code from screenshots of the running application, identifying UI components, layout, and interactive elements from visual input. It completed the game Pokémon FireRed using only raw game screenshots with no maps or navigation aids, demonstrating sustained visual reasoning over a long-horizon task. Practical applications include legacy UI modernization, design-to-code workflows from Figma exports, and visual debugging of rendered interfaces.
