Anthropic released Claude Fable 5.1 on September 1, 2026, as the successor to Claude Fable 5. The pitch is unusual for a frontier launch: the model got better at long-horizon reasoning while the cost of running it went down. This review covers what actually changed, what it costs in practice, and when a developer should reach for it.
What changed in Claude Fable 5.1
Fable 5.1 is positioned as Anthropic's most capable generally available model for coding, knowledge work, and long-running problem-solving. Three changes matter most for developers.
Built for long-horizon agentic work
The headline capability change is stamina. Fable 5.1 ships with adaptive thinking that is always on and defaults to a high effort level, meaning it reasons longer before answering on genuinely hard tasks. Anthropic reports a large jump on agentic and long-horizon work: on Terminal-Bench-Science, a benchmark of autonomous scientific coding, Fable 5.1 scores 52.6% at maximum effort, more than double the 24.7% that Fable 5 achieved.
In practical terms, this is the model you hand a coding session that runs for hours — multi-file refactors, repository-wide migrations, research with many dependent steps — rather than single-shot questions.
1M-token context window, 128K output
Fable 5.1 keeps 1 million tokens in view at once, which Anthropic equates to roughly 555,000 words. That is enough to hold an entire mid-sized codebase, a stack of documentation, and your conversation history in a single session. The maximum output is 128K tokens, so the model can emit finished documents, large spreadsheets, or complete slide decks in one response.
For developers, the big context window changes how you work: instead of carefully selecting which files to show the model, you can dump the relevant directory structure and let it find what it needs. The trade-off is cost, which brings us to pricing.
Cheaper to run than Fable 5
Per-token pricing is unchanged from Fable 5: $10 per million input tokens and $50 per million output tokens. The savings come from cache reads — when the model reuses context it has already processed, which agentic coding loops do constantly — falling 75% to $0.25 per million tokens. Anthropic says typical workloads end up about 25% cheaper than on Fable 5, and heavily agentic workloads up to 45% cheaper.
Cache reads usually dominate the bill on agentic tasks, so this is a bigger deal than it sounds. If you were priced out of long agent runs on Fable 5, the math is meaningfully better now.
Pricing: what it actually costs developers
Here is the pricing picture in one place:
- Input: $10 per million tokens
- Output: $50 per million tokens
- Cache reads: $0.25 per million tokens (75% cheaper than Fable 5)
- Typical workload cost vs Fable 5: roughly 25% less
- Heavily agentic workload cost vs Fable 5: up to 45% less
A few concrete examples help. A short coding question — 2,000 input tokens, 800 output tokens, no caching — costs about $0.06. A long Claude Code session that re-reads 500K cached tokens and generates 40K output tokens costs roughly $2.12 in cache reads plus $2.00 in output, versus about $2.50 just for the cache reads on Fable 5. The model is still priced for professional use, but the caching improvement removes the penalty for the exact workflow the model is designed for.
Where you can access it: Fable 5.1 is available in the Claude apps on paid plans, through the Anthropic API, and via Amazon Bedrock, Google Cloud, and Microsoft Foundry. Third-party API aggregators have also added it under model IDs like claude-fable-5-1.
The safeguard changes that matter for coding
Anthropic says Fable 5.1 delivers higher benchmark scores while reducing how often its safeguard filters interrupt legitimate work — a real friction point for developers on Fable 5. The numbers: cyber-safeguard interventions are down 60% per Claude Code session, and interventions on benign elementary-biology and medical requests are down 85%.
Two concrete policy shifts are worth knowing. Fable 5.1 is allowed to find software vulnerabilities, though it cannot build exploits. And when its automated classifiers do flag a request in the Claude apps, an older, less capable Opus model answers instead of a hard refusal — you get a working, if weaker, response rather than a dead end.
Anthropic also introduced Enterprise Frontier Safeguards in response to enterprise feedback: the 30-day data retention requirement of Fable 5 had limited adoption by regulated companies, so the new program stores data in customer-controlled infrastructure (AWS, Google Cloud, or Azure) and offers zero-retention options, rolling out in phases through fall 2026.
When to use Fable 5.1 — and when not to
Anthropic's own guidance is refreshingly direct: start with the cheaper Opus 5 for most workloads, and move to Fable 5.1 when the work genuinely benefits from deeper or longer-running reasoning. Treat that as the decision rule.
Reach for Fable 5.1 when:
- The task spans multiple files and hours: large refactors, migrations, codebase-wide changes
- You need autonomous scientific or analytical coding with many dependent steps
- You are generating long finished artifacts: technical documents, data pipelines, slide decks
- Your workflow already benefits from prompt caching, so the cost structure favors you
Stay on Opus 5 when:
- The task is a single well-scoped question, code snippet, or quick debugging session
- Latency matters more than depth — Fable 5.1's always-on adaptive thinking adds reasoning time
- You are prototyping and want the cheapest iteration loop
The default high effort level is the right default for Fable 5.1's intended use, but it does mean the model thinks longer even when the task is simple. If you are calling the API, be aware that simpler requests will cost more tokens of thinking than they strictly need — routing simple work to Opus avoids that tax.
Fable 5.1 vs Mythos 5.1: which one you get
Anthropic shipped two models on September 1: Fable 5.1 and Mythos 5.1. They are the same underlying model at the same price, differing only in their safeguard layer. Fable 5.1 is the generally available one with standard safeguards. Mythos 5.1 has a more permissive safeguard layer — it can do exploit work and advanced biology — and is restricted to vetted organisations, initially in the life-sciences sector with broader cybersecurity access planned later.
Unless you are a vetted partner in a restricted domain, you will use Fable 5.1. The split means the general model can be tuned for human-facing work without being forced to serve restricted-domain tasks, which is part of why the false-positive rate dropped.
Getting started: practical tips for developers
Use prompt caching aggressively. Cache reads are the line item that got 75% cheaper, so structure your prompts with stable, reusable prefixes — system prompts, repository maps, shared context — that the API can cache across calls. This is where the 25–45% savings come from.
Lean into the context window, but watch the bill. Dumping a whole directory tree into the context is now reasonable, but 1M tokens of fresh input still costs $10 per million. Prefer cached reads over re-sending the same context.
Let the model plan before it acts. Fable 5.1's strength is long-horizon work, and it performs best when it can decompose a task first. In Claude Code and similar tools, let it investigate the codebase and lay out a plan before making changes — the deeper default reasoning pays off there.
Keep simple work on the cheaper tier. Use Opus-class models for quick questions and single-file edits, and reserve Fable 5.1 for sessions where its reasoning depth and output capacity earn their cost.
Check your provider's model ID. Depending on the platform, look for claude-fable-5-1 (Anthropic API, aggregators), or the corresponding model entry in Amazon Bedrock, Google Cloud's Vertex AI, or Microsoft Foundry.