Claude Fable 5.1

claude-fable-5-1

Claude Fable 5.1 expands on Claude Fable 5, keeping input and output costs unchanged, with cache read cost reduced to only one-quarter, and enhancing long-running agentic coding, multi-step research, and document, spreadsheet, and slide processing capabilities.

Context window1.0M

ProviderClaude

Released2026/09/02

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read	Cache write
10.00/M	50.00/M	0.25/M	12.50/M

Overview

Input

Text Image

Output

Text

Claude Fable 5.1 API: A Mythos-Level Model for Long-Horizon Coding and Research

Claude Fable 5.1 is a Mythos-level model released by Anthropic on September 1, 2026, and is the direct successor to Fable 5, released on June 9. It shares the same underlying model as Claude Mythos 5.1, released the same day, with the only difference being the safety filters: Fable 5.1 is open to all paid users and API accounts, while Mythos 5.1 is only provided to cybersecurity and life sciences organizations that pass Project Glasswing review.

The most striking change this generation is on research-oriented agentic tasks: Terminal-Bench-Science 0.1 jumped from 24.7% for Fable 5 to 52.6%, more than doubling; Terminal-Bench 4.0 rose from 42.0% to 55.8%; GDPval-AA v2 knowledge work benchmark rose from 1723 to 1853 points. Its Artificial Analysis Intelligence Index score of 66 is currently the top, ahead of Opus 5's 63 and GPT-5.6 Sol's 61.

Another thing that has a real impact on your bill: input and output unit prices are exactly the same as Fable 5, but cache read prices are cut to 1/4 of Fable 5's. Anthropic estimates that total cost for typical workloads drops about 25%, and by up to about 45% for long-horizon agentic tasks with heavy tool calls.

SeaWhale AI offers Claude Fable 5.1 through an OpenAI-compatible interface and Anthropic's native Messages API, with support for reasoning Effort control, tool calling, streaming output, and image and file inputs.

Get API Key · Model ID: claude-fable-5-1

Why Choose Claude Fable 5.1

Research agents doubled — Terminal-Bench-Science 0.1 scores 52.6%, vs 24.7% for Fable 5 and 22.4% for GPT-5.6 Sol
Top intelligence index — Artificial Analysis index 66, Opus 5 at 63, GPT-5.6 Sol and Grok 4.6 at 61
#1 in knowledge work benchmark — GDPval-AA v2 scores 1853, 130 points higher than Fable 5 and 142 points higher than GPT-5.6 Sol
Cache reads cut to 1/4 — prefixes hit repeatedly in long sessions are billed at one quarter of Fable 5's rate; agent loop costs drop by up to 45%
ARC-AGI-2 reaches 90.0% — ARC-AGI-1 is 97.5% at max, and even 78.3% at low
1M token context + 128K output — unified unit price across the full window, no tiered upcharge

Core Capabilities

01 Long-Horizon Autonomous Coding

Fable 5.1's improvements are concentrated in sessions that run for hours: multi-file features, large-scale refactoring and migration, cross-session debugging and code review. It scores 55.8% on Terminal-Bench 4.0 (Fable 5: 42.0%, Opus 5: 52.3%), 73.4% on CursorBench 3.2.0, and ranks first on LiveCodeBench at 90.52%. The higher the Effort level, the larger the gap over Fable 5.

One-shot completion of multi-file feature implementations and large-scale migrations
Cross-session defect localization and code review
Unattended, long-duration autonomous execution

02 Research & Deep Research

Terminal-Bench-Science 0.1 rises from 24.7% to 52.6%, the largest single jump of this generation; Humanity's Last Exam: 60.9% without tools and 65.0% with tools; ProofBench v1.1 scores a perfect 100%. It also outperforms Fable 5 on multi-step web research, proactively following leads found through search.

Terminal-level tasks for computational science and experimental workflows
Multi-step deep research across large numbers of papers and reports
Mathematical proof and formal reasoning

03 Document, Spreadsheet & Slide Workflows

GDPval-AA v2 real-world professional task benchmark: 1853 points; AutomationBench rises from 17.1% for Fable 5 to 31.4%. Starting from a single question, it directly produces complete documents, spreadsheets with live formulas, or presentations built from a blank page.

Financial models and multi-sheet workbooks
Analytical reports and research briefings
Presentations built from scratch

04 Vision, Long Context & Computer Use

Improved reading of dense charts, financial reports, and tables embedded in PDFs, with repeated verification using crop and zoom tools. On OSWorld 2.0 computer-use evaluation, it scores 77.9% partial and 41.7% strict, both higher than Fable 5's 72.9% and 36.1%, with more reliable recovery after failed steps. Its ability to correlate details across paragraphs within a 1M token window has also improved.

Parsing dense charts and tables embedded in PDFs
Autonomous operation of browsers and desktop applications
Detail-correlation reasoning across the full window of long documents

Best Use Cases

Use Case	Description
Overnight autonomous coding	Multi-file refactoring, migration, and cross-session debugging; Terminal-Bench 4.0 leads Fable 5 by nearly 14 percentage points
Research agents	Terminal-Bench-Science doubled; suitable for automating computational science and experimental workflows
Deep research	HLE with tools 65.0%; multi-step web research proactively follows leads
Enterprise-grade deliverables	#1 in GDPval-AA v2; end-to-end production of documents, spreadsheets, and slides
Computer-use agents	OSWorld 2.0 leads Fable 5 on both metrics; recoverable from failed steps
Cache-heavy long sessions	Cache reads billed at 1/4 of Fable 5's rate; longer prefixes save more

Claude Fable 5.1 vs Fable 5 vs Opus 5

Capability	Claude Fable 5.1	Claude Fable 5	Claude Opus 5
Model ID	`claude-fable-5-1`	`claude-fable-5`	`claude-opus-5`
Release date	September 1, 2026	June 9, 2026	—
AA Intelligence Index	66	—	63
Terminal-Bench 4.0	55.8%	42.0%	52.3%
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%
GDPval-AA v2	1853	1723	1824
Context window	1M tokens	1M tokens	1M tokens
Max output	128K tokens	128K tokens	128K tokens
Thinking mode	Always on, cannot be disabled	Always on, cannot be disabled	On by default; can be disabled below `high`
Forced tool use `tool_choice: any/tool`	Not supported, returns 400	Supported	Supported
Cache read relative unit price	0.025× base input price	0.1× base input price	0.1× base input price
Positioning	Strongest publicly released model	Previous-generation Mythos-level	Default starting point for most workloads

Actual billing is subject to the real-time price card at the top of the page.

FAQ

When was Claude Fable 5.1 released? September 1, 2026, the same day as Claude Mythos 5.1. It went live on Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry the day it was released.

What are the upgrades over Fable 5? Three areas. First, capability: terminal-level research tasks doubled, and coding, knowledge work, vision, and computer use all improved across the board; the higher the Effort, the larger the gap. Second, cost: unit prices unchanged, cache reads cut to 1/4, reducing total cost of long sessions by 25% to 45%. Third, safety filtering: false refusals on cybersecurity questions dropped by 60%, false refusals on benign basic biology and medicine questions dropped by 85%, and defensive vulnerability discovery is now allowed. Multilingual capability is on par with Fable 5.

Do I need to change code to migrate from Fable 5? There are three breaking changes. The any and tool types of tool_choice return 400; switch to auto with a prompt instruction, or use strict: true and structured outputs to guarantee JSON format. Thinking blocks are bound to the model that produced them and are silently dropped when you switch back to an earlier model. Editing earlier turns invalidates subsequent thinking blocks; accounts created after August 31, 2026 will get a 400 error directly, so sessions must remain append-only.

What are the context and output limits? Context is 1M tokens, which is also the default maximum; the full window is billed at a unified unit price with no tiered upcharge. Max output is 128K tokens. The tokenizer is the same as Fable 5, producing about 30% more tokens than models before Opus 4.7.

Are there any areas where it lags? Yes. It scores 85.02% on Terminal-Bench 2.1, slightly below GPT-5.6 Sol's 85.77%. Output speed is about 66 token/s, below the median for reasoning models, and first-token latency is notably long at high Effort. Artificial Analysis also points out that at max Effort it outputs about 1.7× as many tokens as Fable 5, making per-task cost 20% higher; Anthropic's claimed cost savings only hold at low/medium Effort levels and cache-heavy scenarios. In addition, it still frequently refuses on biology- and cybersecurity-related tasks, which is why it ranks only 18th on the Harvey legal-agent benchmark.

How should I choose the Effort level? There are five levels: low, medium, high, xhigh, and max, with high as the default. Anthropic says that at low and medium it achieves results comparable to or better than Fable 5 at lower cost, so everyday tasks don't need the higher levels. This generation adds the ability to switch Effort mid-session: you can raise it for a difficult step and lower it for routine steps without invalidating the cache. Note that at low it tends to answer from memory and call search tools less, so raise the level for turns that need fresh information.

What behavioral differences are there compared to Fable 5? Several things to watch: parallel tool calls are less stable, and in long loops it may send only one call per turn, using more turns but without affecting quality; it writes fewer progress notes between tool calls, so you need display: "updates" to retrieve them; when editing text files it is more likely to rewrite the whole file rather than make local edits; and when summarizing documents it is more likely to restate the original text without quotation marks. All of these can be fixed with prompt adjustments.

Why Choose SeaWhale AI for Claude Fable 5.1 API

Native Messages API complete passthrough — Effort control, mid-session Effort switching, thinking blocks, caching, and beta features can all be used directly
Direct connection in China — no overseas account or self-hosted proxy needed; stability and latency are guaranteed on the platform side
Pay-as-you-go — pay only for actual usage, with no minimum spending threshold
One key, many models — use Fable 5.1 for hard tasks, switch to Opus 5 or Sonnet 5 for everyday work, all on one unified bill under the same account

API

API integration

Model IDUse this value as the model in inference requests

claude-fable-5-1

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

claude-fable-5-1 usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "claude-fable-5-1",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="claude-fable-5-1",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'claude-fable-5-1',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}