Claude Fable 5 for Enterprise Builds

TL;DR: Claude Fable 5 for enterprise builds is Anthropic's first Mythos-class model: 1M-token context, up to 128k output tokens, priced at $10/$50 per million tokens. It leads SWE-Bench Pro at 80.3% and FrontierCode at 29.3%, completed a 50M-line Ruby refactor at Stripe in a day, and (via its Mythos 5 twin) is contributing to frontier biological research. Three things change for enterprise builds today: long agent runs become feasible, analytics workflows can consolidate onto a single agent, and vision-in-the-loop is a real option.

Anthropic just opened the Mythos era with two new releases: Claude Fable 5 and Claude Mythos 5.

Fable 5 is the first widely available Mythos-class model — and Anthropic's most capable release to date. It targets complex reasoning, long-running agent workflows, and multimodal tasks. Mythos 5 is the same underlying model with fewer safety restrictions, available only to vetted partners in cybersecurity and life sciences.

Both models support a 1M-token context window, up to 128k output tokens per request, and pricing at $10 / $50 per million tokens. Compared to earlier Mythos Preview pricing, the drop cuts costs by roughly $20M on certain workload profiles.

Below: what Claude Fable 5 for enterprise builds actually changes, where the ceiling is, and what the model enables that Opus 4.8 didn't.

Where Fable 5 Sits in the Model Taxonomy

Fable 5 sits above the existing Claude family (Haiku, Sonnet, Opus) on overall capability. It's engineered for long-horizon agent tasks — autonomous coding runs, deep research pipelines, complex analysis over large document sets.

Mythos 5 uses the same base model and delivers the same core capabilities. The difference is access and safety. Fable 5 has classifiers that block or reroute certain high-risk requests. Mythos 5 gives approved partners direct access in areas like cybersecurity and advanced biology.

Anthropic reports that the two models behave nearly identically in 95%+ of user sessions. For nearly every enterprise builder, Fable 5 gives you the practical capability of Mythos 5.

Adaptive Thinking: One Mode, Effort as the Dial

Fable 5 and Mythos 5 ship with a single "adaptive thinking" mode on the API — replacing the multi-mode thinking configuration in earlier Claude generations.

Adaptive thinking is always on. If the thinking parameter is omitted, the model decides how much internal reasoning to spend. Explicitly disabling thinking (type: "disabled") is not supported.

The knob developers get is an effort parameter — it tunes how extensively the model plans, decomposes, and verifies internally. Higher effort means more reasoning depth, higher cost, better outcomes on hard tasks.

Neither model returns raw chain-of-thought. The API returns either omitted thinking blocks (display: "omitted") or summarised reasoning (display: "summarized"), with the expectation that clients pass thinking blocks back unchanged across multi-turn conversations. The design preserves the reasoning benefit while limiting exposure of internal traces that could be misused for model extraction or jailbreak engineering.

Core Capabilities

Long-Horizon Agentic Work

Fable 5 is built for workflows that plan and execute over extended periods with minimal supervision.

The Stripe deployment is the flag example: Fable 5 completed a migration across Stripe's 50 million-line Ruby codebase in about a day. The same task would normally take a team of engineers more than two months.

Evaluations from Cognition's Frontier Code, CursorBench, and FrontierBench rank Fable 5 among the top frontier models for long-horizon coding, particularly at medium or higher reasoning effort.

For enterprise builds where an agent runs unattended for hours or days — batch refactors, multi-doc research pipelines, autonomous ops — Fable 5 is now the model to beat.

Software Engineering and Coding

Fable 5 leads coding benchmarks per Anthropic's reporting, including Cognition's FrontierCode, where it posted the highest score among frontier models while maintaining efficiency at medium reasoning effort. Third-party evaluations (BenchLM and others) place it at or near the top across SWE-bench-style and software development tests.

Outside pure code, partners like IMC report strong performance on financial analysis workflows — factual research, conceptual reasoning, root-cause investigation, expected-value analysis. All in the same model, without hand-off between specialists.

Knowledge Work and Analytics

On long-form reasoning and analytical benchmarks, Fable 5 is currently at the top.

Hebbia's Finance Benchmark: highest score for document reasoning, including structured documents, charts, and tables. Complex-business-problem-solving evaluations: leader.

An analytics provider reported Fable 5 as the first model to score above 90% on core analytics benchmarks for complex multi-step tasks, outperforming Opus 4.8 by 10 points. It also completes analyses in fewer interactions and in under a third of the time of earlier generations.

For enterprise builds anchored in document-heavy analytics (finance, legal, life sciences, consulting), that's a step-function change in what a single agent can do end-to-end.

Vision and Multimodal

Anthropic positions Fable 5 as setting "a new standard for visual reasoning." The model can extract precise numerical information from dense scientific figures and generate working code from screenshots of web applications without access to underlying source.

The model also completed Pokémon FireRed with minimal tool support — a task earlier Claude models couldn't finish even with additional scaffolding. Novelty aside, it's a real signal about spatial reasoning and long-horizon multimodal planning.

Both Fable 5 and Mythos 5 accept text and image inputs across the Claude Platform and supported cloud providers.

Memory, Long-Context, and Persistent Workflows

1M-token context window. 128k output tokens per request.

Fable 5 maintains focus across millions of tokens and can improve outputs across long-running tasks by reading and reusing its own notes.

An internal evaluation on Slay the Spire found Fable 5 with persistent file-based memory improved 3× more than Opus 4.8 and reached the game's final act 3× as often. The signal: Fable 5 is materially better at exploiting long-term memory than the previous generation.

For enterprise builds with multi-day agent runs, extensive codebases, or large document collections, this is the capability that unlocks patterns that weren't previously reliable.

Life Sciences and Scientific Research (Mythos 5)

The most advanced scientific capabilities land through Mythos 5, where safety constraints are relaxed for vetted researchers in areas like protein design, molecular biology, and genomics.

• Protein design. Mythos 5 autonomously runs an end-to-end protein design workflow — selecting binding sites, choosing and operating protein design and bioinformatics tools, recovering from failures. Matches or exceeds skilled human operators on 9 of 14 evaluated protein targets.
• Molecular biology hypothesis generation. Mythos 5 consistently generates novel, compelling hypotheses preferred by Anthropic scientists over Opus-class outputs in ~80% of blinded comparisons. At least one hypothesis has been independently corroborated by an external lab working on the same protein.
• Genomics. Mythos 5 assembled single-cell RNA datasets across 138 animal species, designed and trained a custom model to align cell types across species, and reportedly exceeded the performance of a recent Science paper using a model 100× smaller — operating largely autonomously over more than a week.

Anthropic says these results will be submitted for peer-reviewed publication. Mythos 5 is being positioned as an active contributor to frontier biological research, not just an assistant to it.

Benchmarks and Performance

Fable 5 launches with strong results across coding, knowledge work, vision, agentic tasks, and safety-sensitive domains. It outperforms Claude Mythos Preview, GPT 5.5, and Gemini 3.1 Pro on several core evaluations — particularly long-horizon agentic work and real-world knowledge tasks.

Agentic Coding

• SWE-Bench Pro: Fable 5 at 80.3%, Mythos Preview at 77.8%, GPT 5.5 at 58.6%, Gemini 3.1 Pro at 54.2%
• FrontierCode: Fable 5 at 29.3%, GPT 5.5 at 13.4%, Gemini 3.1 Pro at 5.7%

Knowledge Work

• GDPval-AA: Fable 5 at 1932, GPT 5.5 at 1890, Gemini 3.1 Pro at 1314
• Multidisciplinary reasoning: 59.0%
• Humanity's Last Exam style: 64.5%

Multimodal and Computer-Use

• ODIN (knowledge work vision): 29.8%
• Blueprint Bench 2 (spatial reasoning): 38.6%
• Automation bench (tool use): 17.4%
• OSWorld (computer use): 85.0%

Domain-Specific

• Biology: 46.1% and 83.9% on biology benchmarks
• Cybersecurity: 78.0%
• Health-related tasks: 66.0%

The domain-specific numbers explain why Anthropic treats Mythos-class capability carefully. The same model family that dominates on coding and research also reaches a level where safety and access controls become necessary. That's the trade the Fable / Mythos split is designed to manage.

What Claude Fable 5 Changes for Enterprise Builds

Three things change immediately for teams building on Claude.

Long Agent Runs Become Feasible

Multi-hour and multi-day agentic workflows now have a model that maintains focus and self-verifies over that time horizon. Enterprise builds that were bounded by short attention windows can be re-scoped.

Analytics Workflows Can Consolidate

Fable 5's document reasoning and multi-step analytics performance mean pipelines with multiple specialist steps can often be collapsed into a single agent — fewer moving parts, fewer integration bugs, less orchestration overhead.

Vision-in-the-Loop Is a Real Option

The multimodal benchmarks aren't just headline numbers. Fable 5's ability to work from figures, screenshots, and dense scientific images opens builds that previously needed a separate OCR-plus-parsing stack.

Fable 5 sets a new bar for agentic coding, multimodal reasoning, and domain expertise. If your team is building systems that need to plan, act, and deliver across complex long-horizon workflows, the routing math on new enterprise builds just changed.