Claude Fable 5 review, pricing and limits

Fable 5 is the pick for teams running autonomous, long-horizon coding agents that need the strongest available benchmark performance: it leads every generally available model on SWE-bench Pro at 80.3%. It replaces manual code review on large migrations, but its always-on reasoning adds latency, so cost-sensitive teams running many short queries should look to cheaper models instead.

Released June 9, 2026, Claude Fable 5 is a Mythos-class model, above the Opus tier in Anthropic's lineup, with a 1 million token context window and no long-context surcharge. It leads generally available models on autonomous coding benchmarks, built for long-horizon agentic work, research, and large-document reasoning.

Provider: Anthropic · Family: Claude Fable

More about Anthropic on HokAI

Context window: 1,000,000 tokens · Max output: 128,000

Input modalities: text, image, pdf, tool-calls · Output: text, tool-calls

About Claude Fable 5

Claude Fable 5 is Anthropic's first generally available Mythos-class AI model, released on June 9, 2026. It sits above the Opus tier in Anthropic's model lineup, designed for the most demanding reasoning tasks, autonomous software engineering, knowledge work, vision analysis, and long-horizon agentic execution. Fable 5 is the publicly accessible configuration of a shared Mythos-class architecture: the same underlying weights power Claude Mythos 5, available only to vetted partners through Project Glasswing. Anthropic has not disclosed the parameter count or detailed architecture type. Adaptive thinking is always enabled, meaning the model performs internal reasoning before every response, with depth controlled via an effort parameter rather than a discrete thinking toggle. On software engineering, Fable 5 scores 95.0% on SWE-bench Verified and 80.3% on SWE-bench Pro, ahead of Claude Opus 4.8 at 69.2%, GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2% on SWE-bench Pro. On FrontierCode Diamond, a harder multi-step coding benchmark, Fable 5 reaches 29.3% versus Opus 4.8 at 13.4% and GPT-5.5 at 5.7%. MMLU Pro sits at 91.5%, and HealthBench Professional reaches 66.0% versus Opus 4.8 at 56.9% and GPT-5.5 at 51.8%. On GDPpdf (visual document reasoning without tools), Fable 5 scores 29.8% versus GPT-5.5 at 24.9% and Opus 4.8 at 22.5%. GDPval-AA preference evaluation gives Fable 5 a score of 1932 versus Opus 4.8 at 1890 and GPT-5.5 at 1769. Anthropic did not publish GPQA Diamond, AIME 2025, or ARC-AGI 2 scores at launch. Claude Fable 5 ships with a 1 million token context window at standard pricing, with up to 128,000 output tokens per request. The entire 1M context is billed at the same per-token rate as shorter requests, with no long-context surcharge. Prompt caching applies across the full context window: 5-minute cache writes cost $12.50 per million tokens, 1-hour writes cost $20 per million, and cache reads cost $1 per million, making repeat-context agentic workflows substantially cheaper. Anthropic reports the model stays focused across millions of tokens in long-running agentic tasks, though independent needle-in-haystack evaluations for the full 1M range were not published at launch. The native memory tool is supported, enabling agents to store and retrieve information across conversation boundaries. Fable 5 accepts text, images, PDFs, and tool calls as inputs, and produces text and tool calls as outputs. Vision is fully live at launch: the model reads charts, analyzes screenshots, and extracts figures from PDFs directly, without a separate OCR step. Audio and video inputs are not available. Adaptive thinking replaces the optional extended-thinking parameter from earlier Claude models: every Fable 5 response includes internal reasoning, with depth configurable via the effort parameter. Raw thinking content is never returned; developers can request summarized thinking blocks by setting display to "summarized". Native capabilities include function calling, parallel tool use, the memory tool, task budgets (beta), context editing for long sessions, and compaction for managing context window consumption. Fable 5's list price sits at the top of the generally available frontier tier, roughly double Claude Opus 4.8's per-token rate (see pricing above). The Batch API halves both rates to $5 input and $25 output per million, with results delivered within 24 hours. Prompt caching pays off after a single cache read on the 5-minute TTL: a 100K-token system prompt costs $1.25 to write and $0.10 to read on each subsequent call. US-only inference routing (inference_geo: "us") adds a 1.1x multiplier across all token categories. A daily agentic coding session consuming 2 million input tokens and 400,000 output tokens costs roughly $40 at standard rates, or $20 via the Batch API. From June 9 to June 22, 2026, Fable 5 is included at no extra cost in Claude Pro, Max, Team, and Enterprise subscriptions; starting June 23, continued use requires usage credits billed at API rates. Claude Fable 5 is generally available on the Anthropic Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Azure via Microsoft Foundry, and Claude Platform on AWS. GitHub Copilot added Fable 5 on its launch day, June 9, 2026. Direct API authentication uses a standard Anthropic API key; Bedrock uses AWS IAM, Vertex AI uses GCP IAM, and Foundry uses Azure Active Directory. All Fable 5 sessions carry a mandatory 30-day data retention period: Fable 5 is a Covered Model and zero-data-retention is not available on any platform. Anthropic commits to not training on retained session data, retaining it solely for detecting complex jailbreak attempts and cross-session attack patterns, then deleting it after 30 days. Three classifier systems run on every Fable 5 request before any output is generated. The cybersecurity classifier blocks offensive exploitation, vulnerability discovery, and agentic hacking tasks. The biology and chemistry classifier blocks risky biological research including AAV vector design. The distillation classifier blocks attempts to extract model capabilities for reproduction. When a classifier fires, the API returns stop_reason: "refusal" as an HTTP 200 response with a field identifying which classifier triggered. Developers can pass the fallbacks parameter to retry automatically on Claude Opus 4.8, with fallback credit offsetting the prompt-cache cost of the switch. External bug bounty red-teaming found no universal jailbreaks after over 1,000 hours of effort. Classifiers trigger on fewer than 5% of real-world sessions. Fable 5 is the right choice for teams running autonomous long-horizon software engineering: its 95.0% SWE-bench Verified score leads all generally available models, and Stripe reported compressing months of Ruby codebase migration into a single day. Scientific research applications in drug design, protein hypothesis generation, and genomics benefit from the model's ability to reason across large bodies of literature. Legal and financial teams processing large document corpora benefit from the 1M context window. Teams that should not use Fable 5 include those requiring real-time response, since reasoning overhead adds latency to every request. Audio-first or voice applications have no path because there is no audio modality. Cost-sensitive teams running millions of short queries should use Sonnet 4.6 at $3/$15 per million or Haiku 4.5 at $1/$5 per million instead. Anthropic has not disclosed the training data cutoff date or training data composition for Fable 5. All API sessions carry mandatory 30-day data retention; Anthropic logs all human access to retained data and deletes it after 30 days in most cases. Zero data retention is not available on any platform for Fable 5. The system card published alongside Fable 5 runs 319 pages and covers Firefox zero-day discovery benchmarks, OSS-Fuzz, CyberGym, and CyScenarioBench evaluations. Responsible Scaling Policy (RSP) governs deployment. SOC 2, HIPAA, and GDPR compliance are available through enterprise arrangements on Anthropic-operated and partner cloud platforms. Fable 5 follows Claude Mythos Preview (April 2026, restricted access only) as Anthropic's next Mythos-class release, and the first one generally available to the public. Versus Claude Opus 4.8, Fable 5 improves SWE-bench Pro from 69.2% to 80.3%, FrontierCode Diamond from 13.4% to 29.3%, HealthBench Professional from 56.9% to 66.0%, and MMLU Pro to 91.5%. Adaptive thinking replaces the optional extended-thinking toggle: all responses include reasoning with no disable option. New API parameters include effort for reasoning depth control, task-budgets (beta) for token-aware agentic planning, and fallbacks for automatic classifier-triggered model switching. The API model ID is claude-fable-5.

Pricing

$10/$50 per 1M input/output tokens. Batch API: $5/$25 per 1M (50% off, async within 24h). Cache read: $1/MTok (0.1x input rate). 5-min cache write: $12.50/MTok; 1-hr cache write: $20/MTok. US-only inference (inference_geo: 'us') adds 1.1x multiplier on all token categories. Free on Claude Pro/Max/Team/Enterprise plans June 9-22, 2026; requires usage credits from June 23 onward.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Adaptive Thinking Always On: Internal reasoning is applied to every request by default, with depth controlled via the effort parameter. Unlike earlier Claude models, thinking cannot be disabled on Fable 5, and raw chain-of-thought is never returned.
• 1M Token Context Window: Handles up to 1 million tokens in a single context at standard per-token pricing with no long-context surcharge. Supports processing entire large codebases, legal document sets, or research corpora in one API call.
• Safety Classifier Fallback System: Three built-in classifiers (cybersecurity, biology/chemistry, distillation) detect high-risk requests and route them to Claude Opus 4.8 automatically, returning HTTP 200 with stop_reason: 'refusal' and the triggering classifier identified.
• Vision and Document Understanding: Reads images, charts, screenshots, and PDFs natively at launch. Demonstrated rebuilding web application interfaces from screenshots and playing Pokemon FireRed using only visual input with no text.
• Multi-Platform Day-One Availability: Generally available on Anthropic API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry, Claude Platform on AWS, and GitHub Copilot simultaneously on June 9, 2026, with no waitlist for API access.

Pros

• Leads every generally available model on coding benchmarks, beating GPT-5.5 by a 21.7-point margin on the harder SWE-bench Pro split, which matters most for teams running long autonomous coding sessions with minimal human review.
• 1M context window at standard pricing with no surcharge, plus adaptive thinking that maintains coherence across full-million-token agentic sessions.
• Day-one availability across five major platforms (Anthropic API, Bedrock, Vertex, Foundry, GitHub Copilot) covering every major enterprise cloud stack without waitlists.

Cons

• Mandatory 30-day data retention with no zero-retention option on any platform, blocking adoption in sensitive sectors with strict data handling requirements.
• The priciest generally available frontier model on the market today, a poor fit for high-volume short-query workloads that need a cheaper per-token rate.
• No audio input or output; adaptive thinking adds latency to every request regardless of complexity, ruling out real-time voice-first and latency-critical applications.

Benchmarks

• Gdppdf: 29.8 vendor-reported · 09 Jun 2026
• MMLU-Pro: 91.5% vendor-reported · 09 Jun 2026 — A harder version of the 57-subject knowledge exam, % correct.
• Gdpval Aa: 1,932 vendor-reported · 09 Jun 2026
• SWE-bench Pro: 80.3% vendor-reported · 09 Jun 2026 — Harder, longer real-repository coding tasks, % solved.
• SWE-bench Verified: 95% vendor-reported · 09 Jun 2026 — Real GitHub issues fixed end to end, % solved.
• Frontier Code Diamond: 29.3 vendor-reported · 09 Jun 2026
• Health Bench Professional: 66 vendor-reported · 09 Jun 2026

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Claude Fable 5 Official Page