Reference · verified 2026-11-13
Claude Fable 5.2, specified
Claude Fable 5.2 is Anthropic's frontier model for demanding reasoning and long-horizon agent work. It was released on 2026-11-09 with a 1M-token context, a 128K output ceiling, text and image input, five effort levels, an optional fast mode and a price of $8 in and $40 out per million tokens. This page is the long version: the model card, the full launch table, the independent Max-effort measurements and the limits that matter before you ship it.
Model card read 2026-11-13 Anthropic documentation → OpenRouter listing →
The facts on the card
| Field | Value |
|---|---|
| Maker | Anthropic Released 2026-11-09; active and latest. |
| Retirement | Not before 2026-11-09 Anthropic's published one-year lifecycle commitment. |
| Class | Frontier reasoning and agent model Adaptive thinking is always on; effort controls depth. |
| OpenRouter id | anthropic/claude-fable-5.2 Canonical OpenRouter route with the same base token rates. |
| Claude API id | claude-fable-5-2 Also listed on Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. |
| Context | 1,000,000 tokens 1M-token window. |
| Max output | 128,000 tokens 128K ceiling. |
| Input | Text and images Output is text. OpenRouter's catalogue also advertises file input. |
| Effort | Low · Medium · High · Xhigh · Max OpenRouter exposes reasoning_effort with all five supported values. |
| Default effort | High Anthropic's API model card; the launch post calls Medium the default in apps and Claude Code. |
| Latency class | Moderate Anthropic's comparative label, not a measured request latency. |
| Knowledge cutoff | June 2026 Reliable and training cutoffs are both listed as June 2026. |
| Data retention | 30 days Zero data retention requires express authorisation or Enterprise Frontier Safeguards. |
platform.claude.com/docs/en/models/fable-5-2/overview; openrouter.ai/api/v1/models Verified 2026-11-13 Release and platform list: AnthropicReasoning effort values: OpenRouter Pricing, line by line
The base rates are $8 per million input tokens and $40 per million output tokens. Cache reads are far cheaper than fresh input, which is the number that matters for an agent that keeps resending the same system prompt, tool definitions and repository context. Fast mode is a separate purchase: it doubles the token rates to $16 and $80.
| Item | Price | What it covers |
|---|---|---|
| Input | $8 / MTok | Fresh input tokens. |
| Output | $40 / MTok | Includes answer and reasoning output token charges. |
| 5-minute cache write | $10 / MTok | Writing a reusable prompt prefix. |
| 1-hour cache write | $16 / MTok | The longer-lived cache option. |
| Cache read | $0.2 / MTok | Reusing a prefix the provider has already seen. |
| Fast mode input | $16 / MTok | Double the standard input rate. |
| Fast mode output | $80 / MTok | Up to 2.5x generation speed, according to Anthropic. |
| Batch API | 50% off | Applied to input and output for work that can wait. |
Anthropic's pricing panel, read 2026-11-13 OpenRouter lists the same base input and output rates
What changed from Fable 5.1
Anthropic's launch puts the revision in three terms: faster output, a lower price and better scores. The table below keeps the claim next to the source. “Up to 30% faster” is a vendor comparison on output generation, not a promise that every task finishes 30% sooner.
| Field | Fable 5.2 | Basis |
|---|---|---|
| Output speed | Up to 30% faster than Fable 5.1 | Anthropic's launch claim |
| Price | 20% lower than Fable 5.1 | Anthropic's launch claim; current price is $8 / $40 |
| FrontierCode 2.0 | 49.1% at Xhigh | Fable 5.1 printed 32.4% on the same row |
| FrontierSWE v2 | 69.9% | Fable 5.1 printed 56.3% |
| AutomationBench | 51.4% | Fable 5.1 printed 31.4% |
| RoboDojo Sim | 64.9% | Fable 5.1 printed 27.6%; Anthropic's internal run |
| Chartography | 73.4% | Fable 5.1 printed 44.8% |
Anthropic's launch post and benchmark table, read 2026-11-13 No independent Fable 5.1 comparison was run for this page
What changed after launch
The model ID and list price stayed put; the evidence around the model did not. Artificial Analysis's current Max-effort run has a larger comparison set and a slower task profile than the launch-week reading. Anthropic also clarified the input contract and published the breaking API changes that matter to an existing agent. The demo supply grew even faster.
| Reading | Launch-week note | 2026-11-15 | What changed |
|---|---|---|---|
| AA Intelligence Index | 65 · #1 of 74 | 64.6 · #1 of 78 | The public page rounds the score to 65. The ranking pool expanded; Fable still leads it. |
| AA task profile | 44.4 min · 230 s first answer · 103 tok/s | 46.8 min · 267.23 s · 97.48 tok/s | A rerun, not a model change. Cost per task is still $26.54. |
| Input contract | Text, images and files | Text and images | Anthropic's capability table is narrower; OpenRouter still advertises file input. |
| Public demos | 57 live + 61 video-only | 165 live + 119 video-only | The community index grew from 57 to 165 live works by 2026-11-15. |
Model facts read 2026-11-13; the CC0 community demo index read 2026-11-15 The launch-week values are kept as the comparison baseline
The launch table, row for row
Anthropic selected the tasks and chose the models beside Fable. The table is still useful: it shows the shape of the improvement, and several rows have large gaps rather than marginal wins. Read “Xhigh”, “High” and “Max” as part of the result, because a score from a higher effort is not a like-for-like comparison with a lower one.
| Category | Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | Scale |
|---|---|---|---|---|---|
| Agentic coding | FrontierCode 2.0 | 49.1% · Xhigh | 32.4% · Xhigh | 45.5% | GPT-6 Astra 42.8% · Gemini 4 Carbon 41.1% · Medium |
| Agentic coding | FrontierSWE v2 | 69.9% | 56.3% | 62.3% | GPT-6 Astra 65.5% · Gemini 4 Carbon 61.2% |
| Business workflows | AutomationBench | 51.4% | 31.4% | 42.5% | GPT-6 Astra 41.4% · Gemini 4 Carbon 53.9% |
| Scientific research | Terminal-Bench-Science 0.1 · Mythos 5.2 prints 75.3% | 74.6% | 52.6% | 58.7% | GPT-6 Astra 68.1% · Gemini 4 Carbon 64.8% |
| Agentic search | WANDR · relaxed score | 82.1% | 68.7% | 72.3% | |
| Computer use | OSWorld 2.1 · offline, partial | 89.3% | 83.5% | 87.2% | GPT-6 Astra 79.5% · Gemini 4 Carbon 81.5% |
| Robotics | RoboDojo Sim · Anthropic's internal run | 64.9% | 27.6% | 39.4% | GPT-6 Astra 29.8% |
| Knowledge work | AA-Briefcase v1.1 · The Elo and rubric score do not agree on the winner | 66.1% · 1,785 Elo | 61.5% · 1,669 Elo | 61.2% · 1,800 Elo | GPT-6 Astra 52.0% · 1,569 Elo · Gemini 4 Carbon 61.8% · 1,611 Elo |
| Knowledge work | Remote Labor Index | 41.3% · High | 17.9% | 21.3% | GPT-6 Astra 20.8% |
| Reasoning | Humanity's Last Exam · with tools | 74.2% | 65.6% | 67.7% | GPT-6 Astra 57.2% |
| Visual recognition | Chartography · no tools | 73.4% | 44.8% | 64.4% | GPT-6 Astra 71.0% · Gemini 4 Carbon 72.5% |
| Hardware engineering | BenchCAD 2.1 · with tools | 83.5 · Max | 44.4 | 65.3 |
Anthropic's launch table, read 2026-11-13 Rows with no comparison model were left blank rather than inferred
Cybersecurity and protein-design rows are not in this table because Anthropic reports the unrestricted Mythos 5.2 results there, not the safeguarded Fable results. That distinction is not a footnote: a team that needs those capabilities must apply for Mythos access rather than assume that Fable will perform the same work.
The independent Max-effort check
Artificial Analysis placed Fable at the top of its Intelligence Index with a score of 64.6; its public page rounds that to 65. That run is worth reading closely because it is not a vendor table: the model was evaluated at Max effort, averaged 46.8 minutes per task and cost $26.54 per task. It also measured 267.23 seconds to the first answer token and 97.48 output tokens per second.
| Independent figure | Value | How to read it |
|---|---|---|
| Intelligence Index | 64.6 | Rank 1 of 78 on the 2026-11-13 index; the public page displays 65. |
| Time to first answer token | 267.23 s | Max effort; this includes thinking time. |
| Output speed | 97.48 tokens/s | An average, not an interactive SLA. |
| Time per task | 46.8 min | Weighted over the index tasks. |
| Cost per task | $26.54 | Token charges at the public rates; not a monthly subscription. |
Artificial Analysis, read 2026-11-13 Max effort only; lower effort settings are not represented in these figures
A second, vote-based signal
OpenRouter's model record also carries Design Arena category results. They measure pairwise public preference by task, so they answer a different question from the Intelligence Index: not “how hard a problem can it solve?” but “which result do people choose?” Fable's strongest category is 3D; its weakest listed rank is fifth in game development and UI components.
| Design Arena category | Rank | Elo | Win rate |
|---|---|---|---|
| 3D | #1 | 1,516 | 76.8% |
| Mobile apps | #2 | 1,323 | 65.5% |
| #3 | 1,306 | 68.2% | |
| Code categories | #3 | 1,374 | 62.7% |
| Website | #3 | 1,353 | 60.6% |
OpenRouter model record, Design Arena field, read 2026-11-13 Category ranks are point-in-time and move as voters submit more comparisons
The 1,000-agent formalization run
The most concrete long-horizon result in the launch material is a multi-agent run, not a single prompt. Anthropic gave 1,000 agents running an early Fable 5.2 build a shared SQL database, separate network-isolated sandboxes, Git branches and the Lean toolchain. It also supplied 60 statements, a one-sentence sketch of the classical route, the definitions those statements use, and 47 already-proved lemmas. The agents chose how to organise the rest.
| Result | Fable 5.2 team | What it measures |
|---|---|---|
| Proof completed | 28.7 hours | Wall time for the full formalization, including coordination. |
| Output tokens | One-third of the earlier run | Token efficiency against Anthropic's September formalization. |
| Proof length | One-seventh of the earlier proof | A shorter Lean artifact, not merely a faster transcript. |
| Opus 5.5 control | 26% in 18 hours | Same 1,000-agent setup; Anthropic says it barely progressed after 30 hours. |
Every shell command and file edit passed a classifier before execution, and an offline monitor reviewed the team's actions. Treat this as a harness result, not a claim that any long task will complete unattended. The useful lesson is that the model was given shared state, narrow tools and a checkable finish line — the same ingredients a production agent needs.
Anthropic's multi-agent write-up, read 2026-11-13 Harness, safety controls and results →
Model IDs and where it runs
| Platform | Model id |
|---|---|
| Claude API | claude-fable-5-2 |
| Amazon Bedrock | global.anthropic.claude-fable-5-2 |
| Claude Models on Google Cloud · Google managed | claude-fable-5-2 |
| Claude Models on Google Cloud · Anthropic managed | claude-fable-5-2 |
| Microsoft Foundry · Anthropic hosted | claude-fable-5-2 |
| Claude Platform on AWS | claude-fable-5-2 |
| OpenRouter | anthropic/claude-fable-5.2 |
Anthropic's model card, read 2026-11-13 OpenRouter id from the catalogue
What the launch material does not make easy
- It is not the default model. Opus 5.5 has the same window and output ceiling at half the price. Anthropic itself recommends Opus as the daily driver.
- It is not a cheap Max button. A Max-effort task can take 46.8 minutes and cost $26.54 in the current independent evaluation. Medium and High exist for a reason.
- It is not zero-retention by default. Anthropic requires 30-day data retention unless zero data retention is expressly authorised or Enterprise Frontier Safeguards are in use.
- Some API shapes fail. Forced tool use with
tool_choiceset toanyortoolreturns a 400, and temperature, top_p and top_k should be omitted. - The safety story is part of the product. Fable's cyber and biology classifiers block some legitimate work; Mythos is the verified access path for those fields.
- Benchmarks are selected. The launch table is a vendor comparison, and the independent score is one Max-effort run on a fixed index. Neither replaces an evaluation on your own tasks.
What this page cannot tell you
There is no private benchmark here, no production latency test and no cost forecast for your token mix. The model card does not publish a per-effort latency table, and Artificial Analysis only measured Max. The safest reading is the one Anthropic itself gives: use Opus for the daily work, switch up to Fable when the task needs a harder attempt, and test the effort setting against the work you actually do.
FAQ
What is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's frontier model released on 9 November 2026 for demanding reasoning and long-horizon agentic work. Anthropic's capability table lists text and image input and text output; OpenRouter also advertises file input. It has a 1M-token context window and a 128K output ceiling. Anthropic positions it above Opus 5.5 for the hardest open-ended work, not as the default for every task.
How much does Claude Fable 5.2 cost?
The list price is $8 per million input tokens and $40 per million output tokens. Cache reads are $0.20, a five-minute cache write is $10 and a one-hour cache write is $16. Fast mode doubles the token rates to $16 input and $80 output. Batch API requests receive a 50% discount on input and output.
What is the difference between Fable 5.2 and Mythos 5.2?
Anthropic says they share the same underlying model, specifications and price. Fable is the generally available version with cyber and biology safeguards. Mythos is offered only to organisations verified through Anthropic's cyber or life-sciences access programmes, where those safeguards would block legitimate work.
Is Fable 5.2 the same as Claude Opus 5.5?
No. Both have a 1M-token context and 128K output, but Fable lists at $8 / $40 per million against Opus at $4 / $20 and is positioned for harder, longer-running work. Anthropic recommends starting most coding and knowledge tasks on Opus 5.5 and moving up to Fable when the scope or difficulty justifies it.
What effort setting should I use?
Start with High for API work and Medium for interactive Claude apps or Claude Code if that is the client's default. Move to Xhigh only when a hard task has survived a cheaper attempt. Max is for benchmarking or unattended work where the result is worth a multi-dollar, tens-of-minutes run. The effort level changes thinking-token use; it does not change the list price per token.
Why is one task sometimes reported to cost $26.54?
That figure comes from Artificial Analysis running the model at Max effort across its Intelligence Index. The current run averages 46.8 minutes and $26.54 per task because a hard reasoning task can produce a large number of reasoning and answer tokens. It is not the price of every call, and it is not the price of an interactive chat.
What is fast mode?
Fast mode is an optional Anthropic API setting that can generate tokens up to 2.5 times faster. It costs $16 per million input tokens and $80 per million output tokens, double the standard rates. It is available in Claude Code and through Anthropic's API and supported cloud platforms.
Can I use Fable 5.2 through OpenRouter?
Yes. The catalogue lists anthropic/claude-fable-5.2 with a 1M-token context, 128K maximum output, text, image and file input, and the same $8 / $40 base rates. OpenRouter also exposes the reasoning-effort parameter and lists max, xhigh, high, medium and low as supported efforts. Anthropic's own model card lists text and images; use the provider's contract you actually call.
Medium vs High vs Max Fable vs Opus The demo wall Anthropic's launch