What is Claude Fable 5 and when was it released?

Claude Fable 5 is Anthropic's most capable publicly available AI model, released on June 9, 2026. It is the first model from Anthropic's "Mythos-class" tier to be made generally available to developers and enterprises. The model is designed for long-horizon agentic work, complex software engineering, and demanding knowledge tasks that previous models could not sustain over extended periods. The name "Fable" comes from the Latin fabula, meaning "that which is told" — a deliberate linguistic cousin of the Greek mythos, reflecting the model's shared intelligence with the restricted Mythos 5 model.

What is the difference between Claude Fable 5 and Claude Mythos 5?

Fable 5 and Mythos 5 share the exact same underlying neural network weights and core intelligence. The difference is entirely based on access control and safety guardrails. Fable 5 is the publicly available version with dynamic safety classifiers that block high-risk queries (cybersecurity exploits, biological hazards, model distillation) and route them to Claude Opus 4.8. Mythos 5 is a restricted model with some of those safety classifiers disabled. It is only available to vetted partners through Project Glasswing — primarily US government cyber-defenders and approved biology researchers. For all standard enterprise and developer tasks, Fable 5 performs effectively identically to Mythos 5.

How much does Claude Fable 5 cost? What is the API pricing?

Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. This is exactly double the cost of Claude Opus 4.8 ($5/$25). A 90% prompt caching discount applies to cached input tokens, reducing the effective input cost to $1/M on cache hits. For US-only inference (data residency requirements), pricing is available at a 1.1x multiplier. On Claude.ai subscription plans, Fable 5 counts as 2x usage. Anthropic also offered a free promotional window for Pro, Max, Team, and Enterprise subscribers from June 9–22, 2026.

Is Claude Fable 5 better than GPT-5.5?

Based on Anthropic's published benchmarks and extensive third-party testing, Claude Fable 5 significantly outperforms GPT-5.5 on complex reasoning and software engineering tasks. On SWE-Bench Pro, Fable 5 scored 80.3% compared to GPT-5.5's 58.6% — a gap of over 21 percentage points. On FrontierCode Diamond, Fable 5 scored 29.3% versus GPT-5.5's 5.7%. GPT-5.5 remains competitive in specific niches: it performs well on Terminal-Bench 2.1 via its Codex CLI harness (83.4%) and is competitive on spatial reasoning (Blueprint-Bench 2: 36.2% vs Fable 5's 38.6%). GPT-5.5 also offers a lower price point ($5/$30 vs $10/$50). For long-horizon autonomous coding and complex knowledge work, however, Fable 5 is the clear leader.

What is Claude Fable 5's context window size?

Claude Fable 5 supports a 1,000,000 token (1M token) context window by default. It also supports generating up to 128,000 output tokens per single request. This massive context window enables Fable 5 to process entire codebases, lengthy legal documents, or extensive research papers in a single session without losing context.

Why did my Fable 5 request get answered by Opus 4.8?

This is the intended behavior of Fable 5's safeguard system. If your prompt contained keywords or concepts related to offensive cybersecurity (e.g., building exploits or malware), creating biological hazards, or model distillation, Fable 5's safety classifiers intercepted the request and automatically routed it to Claude Opus 4.8. Importantly, you are not charged Fable 5 rates when this fallback occurs — you only pay for the Opus 4.8 usage. The API will return a successful HTTP 200 response with stop_reason: "refusal" in the metadata to indicate the fallback occurred. Anthropic reports this happens in fewer than 5% of average sessions.

When should I use Fable 5 instead of Opus 4.8?

You should reach for Claude Fable 5 when dealing with tasks that require sustained autonomy, complex multi-step reasoning, or massive code refactoring. Ideal use cases include: large-scale codebase migrations, multi-day autonomous coding sessions, complex financial document analysis, multi-stage research synthesis, and any task where the AI needs to plan, execute, and self-verify over an extended period. You should stick with Claude Opus 4.8 for routine, high-volume, or latency-sensitive work: classification, summarization, drafting, interactive chat, and most day-to-day agentic tasks. At half the price, Opus 4.8 is the sensible default for the majority of enterprise workloads. The optimal strategy for most organizations is intelligent routing — using Fable 5 for the hardest 20% of tasks and Opus 4.8 for the rest.

Where can I access Claude Fable 5?

Claude Fable 5 is available across multiple platforms as of June 9, 2026. For developers, it is accessible via the Claude API directly (model string: claude-fable-5), as well as through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. For enterprise customers, it is available on the consumption-based Enterprise plan. For individual users, Fable 5 is accessible through the Claude.ai web interface and mobile apps on paid plans (Pro, Max, Team, and Enterprise). Note that Fable 5 counts as 2x usage on subscription plans due to its higher compute requirements.

Initial Impressions of Claude Fable 5

After ~5.5 hours of testing on launch day, Simon Willison calls Claude Fable 5 "something of a beast" — slow, expensive, and happily churning through ambitious tasks. Key traits: 1M context, 128k max output, January 2026 knowledge cutoff, $10/$50 pricing (2× Opus), and conservative safety classifiers with new API fallback mechanisms.

The big model smell

Fable feels big in speed, cost, and how much it knows. Asked to list Simon Willison's open source projects with dates (no search), Fable produced a far richer list than Opus 4.8 — from files-to-prompt and LLM through Datasette, Django, and dozens of utilities. Willison treats knowledge depth as a proxy for model size; Anthropic has not disclosed parameter counts.

Using Fable in Claude.ai

Willison upgraded micropython-wasm to full CPython-in-WASM: Fable cloned the repo, identified Brett Cannon's cpython-wasi-build, and after he uploaded the zips, produced a working 13.9MB wheel runnable via uv. A strong first-day coding demo.

Datasette Agent and LLM library

Fable implemented human-in-the-loop tool approval for Datasette Agent, then unraveled hacks into proper LLM 0.32a3 features: llm_tool_call parameter, guaranteed unique tool_call_ids, PauseChain exceptions, resume from pending tool calls, and async executor fixes — work that felt like several days in one session.

How much it cost

Using AgentsView, Willison tracked ~$110.42 of Fable usage in one day on a $100/month Max subscription — mostly a Datasette Agent review session at $99.26 (78.2M tokens). Fable is not cheap for long agent runs.

Effort levels and pelicans

Generating "SVG of a pelican riding a bicycle" across effort levels: low ~1,929 out tokens (~10¢), medium ~2,290 (~11¢), high ~2,057 (~10¢), xhigh ~5,992 (~30¢), max ~14,430 (~72¢). High sometimes beat medium on token count for the same prompt.

“The challenge is finding tasks that it can't do.”