Claude Fable 5.2: 20% Cheaper, 30% Faster, and What Changes for Developers
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, 2026: $8 input and $40 output per million tokens (down from $10/$50), output up to 30% faster than 5.1, 55% fewer safety-classifier blocks, and big jumps on agentic benchmarks. Here's the pricing math, the benchmarks against Opus 5.5, the four breaking API changes, working TypeScript, and when Fable 5.2 is actually worth it over Opus 5.5.
- Published
- Reading time
- 11 min read

On this page
TL;DR: On November 9, 2026, Anthropic released Claude Fable 5.2 (claude-fable-5-2) and its restricted twin Claude Mythos 5.2. Fable 5.2 is Anthropic's most capable widely released model. It costs $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1 ($10/$50), generates output up to 30% faster, and keeps the 1M-token context window and 128K max output. Its safety classifiers block about 55% fewer legitimate requests than 5.1's. On Anthropic's benchmarks it beats Fable 5.1 by wide margins and beats Opus 5.5 on most agentic and coding evals, though not every one. Anthropic still calls Opus 5.5 "our strongest daily driver" for most coding and knowledge work, so Fable 5.2 is the model you reach for on the hardest, longest tasks, not the default. Below: pricing, benchmarks, the API changes that will break 5.1 code, working TypeScript, and how I'd decide between Fable 5.2 and Opus 5.5.

What Anthropic Announced
Fable and Mythos are the same underlying model. The difference is the safeguards: Fable 5.2 ships with stronger cyber and biology classifiers and is available to everyone; Mythos 5.2 has those safeguards relaxed and is limited to organizations in Anthropic's Cyber Verification Program and Life Sciences Verification Program.
The headline changes from Fable 5.1:
- 01Lower price: $8 / $40 per million tokens, down from $10 / $50. Cache reads are $0.20 per million.
- 02Faster output: up to 30% faster generation than Fable 5.1 and Mythos 5.1.
- 03Fast mode: an optional mode that generates tokens up to 2.5x faster at $16 / $80 per million tokens.
- 04Fewer false refusals: about 55% fewer classifier blocks overall (about 60% fewer in cyber, 40% fewer in bio).
- 05Medium effort by default in the Claude apps and Claude Code.
- 06Better coding, reasoning and vision, plus stronger judgement on ambiguous requests and the ability to orchestrate teams of agents.
Fable 5.2 is available on the Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Foundry, and in Claude Code and the Claude apps. Pro plans and standard Team seats get a launch promotion: Fable usage of up to 50% of weekly limits, plus one to three Fable resets depending on subscription length.
Claude Fable 5.2 Specs at a Glance
| Spec | Claude Fable 5.2 |
|---|---|
| API model ID | claude-fable-5-2 |
| Release date | November 9, 2026 |
| Input price | $8 per million tokens |
| Output price | $40 per million tokens |
| Cache reads | $0.20 per million tokens |
| Fast mode | $16 / $80 per million tokens, up to 2.5x faster output |
| Context window | 1M tokens (the default and the maximum) |
| Max output | 128K tokens |
| Thinking | Always on (adaptive); depth set with effort |
| Effort levels | low, medium, high, xhigh, max |
| Platforms | Claude Platform, Amazon Bedrock, Google Cloud, Microsoft Foundry |
| Data retention | 30-day retention required (ZDR only via Enterprise Frontier Safeguards, private beta) |
Pricing: Fable 5.2 vs Fable 5.1 vs Opus 5.5
| Model | Input / MTok | Output / MTok | Cache read / MTok |
|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 |
| Claude Fable 5.1 | $10 | $50 | — |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 |
Per token, Fable 5.2 is still 2x the price of Opus 5.5. But per-token price isn't the number that matters for agent work; cost per completed task is. Anthropic's own figure makes the point: on FrontierCode at high effort, Fable 5.2 scores 45.3%, about the same as Opus 5.5's 45.5%, at roughly half Opus 5.5's cost per task. A model that finishes in fewer turns, with fewer retries, can cost less even at a higher list price. Measure on your own workload before you assume either way. I walk through how to do that in my LLM cost audit guide.
Since Fable 5.2 uses the same tokenizer as Opus 4.7 and 4.8, token counts don't change if you're coming from those models or from Fable 5.1, so the 20% cut is a straight 20% saving on the same traffic.
Benchmarks: Fable 5.2 vs Fable 5.1 vs Opus 5.5
These are Anthropic's published numbers. They're vendor-reported, so treat them as a starting point, not a verdict.
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 (xhigh) | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1 (offline, partial) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| Remote Labor Index (high) | 41.3% | 17.9% | 21.3% | 20.8% | — |
| Humanity's Last Exam (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | — |
| RoboDojo Sim | 64.9% | 27.6% | 39.4% | 29.8% | — |
| AA-Briefcase v1.1 | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| Chartography (no tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
It's not a clean sweep: Gemini 4 Carbon beats Fable 5.2 on AutomationBench and comes within a point on Chartography.
What stands out:
- Long, agentic work is where the gap is biggest. Remote Labor Index more than doubles (17.9% to 41.3%) and is nearly 2x Opus 5.5. RoboDojo Sim goes from 27.6% to 64.9%. These are tasks that run for a long time and need the model to keep its plan together.
- Coding gains are real but smaller over Opus 5.5. FrontierSWE v2 is +7.6 points over Opus 5.5, FrontierCode +3.6. Good, not a different league.
- Computer use is near the ceiling. OSWorld 2.1 at 89.3% vs Opus 5.5's 87.2%.
On the Mythos side, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 on Anthropic's internal CVE Rediscovery evaluation (89.6%). Fable 5.2's safeguards block most cybersecurity tasks, so don't plan on using Fable for offensive security work.
Safety: Fewer False Refusals, Better Alignment
Fable models have had a reputation for over-refusing, which hurt in production. Fable 5.2 improves on that:
- About 55% fewer classifier blocks than Fable 5.1 (about 60% fewer in cyber, 40% fewer in bio).
- When the classifier does block, the request can be answered by a fallback model: Opus 4.8 for blocked cyber queries and Opus 5 for blocked biology queries.
- On Anthropic's automated behavioural audit (about 1,900 scenarios, lower is better), Mythos 5.2 scores 2.67, better than Opus 5.5 (2.79), Sonnet 5.5 (3.10) and Mythos 5.1 (3.28).
- In the sandbox-escape evaluation, Mythos 5.2 used outside credentials or files in 3.1% of scenarios, down from 21% for Mythos 5.1 and on par with Opus 5.5 (3.3%).
External testing came from the UK AI Security Institute, METR and Frontier Design.
The API Changes That Will Break Fable 5.1 Code
If you're already on claude-fable-5-1, the surface is mostly the same, but four things change:
- 01The history-editing check now applies to every account. Fable 5.2 runs preserved thinking: thinking blocks are tied to the model and conversation that produced them. If your harness edits earlier turns (trims, rewrites, reorders), replayed thinking blocks get dropped. Make your conversation history append-only.
- 02Older models can't read Fable 5.2's thinking blocks, Fable 5.1 included, and they only work in the account that produced them (or a linked one). A fallback to an older model runs without them.
- 03
computer_20251124returns a 400 on every platform except Amazon Bedrock. Usecomputer_toolset_20260801. - 04Advisor tool pairing: a Fable 5.2 executor accepts only
claude-fable-5-2orclaude-mythos-5-2as its advisor.
And if you're coming from Opus 4.x or Sonnet, these also apply:
- Thinking is always on. Omit
thinkingor send{ type: "adaptive" }. Both{ type: "disabled" }andbudget_tokensreturn a 400. Control depth withoutput_config.effort. - The raw chain of thought is never returned. By default the thinking text is empty; set
display: "summarized"to get a readable summary. - No assistant prefill. Use structured outputs or system prompt instructions.
- No forced tool use.
tool_choiceofanyortoolreturns a 400. Useautowithstrict: truetools and say which tool to use in the prompt. - No
temperature,top_portop_k. - Handle
stop_reason: "refusal"before readingcontent. - No Priority Tier, and 30-day data retention is required.
Calling Claude Fable 5.2 from TypeScript
A basic request with effort set explicitly and refusal fallback turned on:
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();
const response = await client.beta.messages.create({
model: "claude-fable-5-2",
max_tokens: 16000,
output_config: { effort: "high" },
// If a safety classifier declines, the API re-runs the request on a
// fallback model inside the same call, routed by refusal category.
betas: ["server-side-fallback-2026-07-01"],
fallbacks: "default",
messages: [
{ role: "user", content: "Review this migration plan and list the risks..." },
],
});
if (response.stop_reason === "refusal") {
// The whole fallback chain declined.
console.warn("Refused:", response.stop_details?.category);
} else {
for (const block of response.content) {
if (block.type === "text") console.log(block.text);
}
}Hard tasks can run for many minutes on Fable 5.2, so stream anything with a large max_tokens:
const stream = client.messages.stream({
model: "claude-fable-5-2",
max_tokens: 64000,
thinking: { type: "adaptive", display: "summarized" },
output_config: { effort: "xhigh" },
messages: [{ role: "user", content: longTaskSpec }],
});
stream.on("text", (delta) => process.stdout.write(delta));
const final = await stream.finalMessage();
console.log(final.usage);And fast mode, when latency matters more than cost:
const fast = await client.messages.create({
model: "claude-fable-5-2",
max_tokens: 8000,
speed: "fast", // $16 / $80 per MTok, up to 2.5x faster output
messages: [{ role: "user", content: "Summarize this incident timeline..." }],
});
console.log(fast.usage.speed);Choosing an Effort Level
Effort is the main dial on Fable 5.2, and it matters more than on older models:
| Effort | Use it for |
|---|---|
low | Routine edits, classification, subagent steps |
medium | Default in Claude apps and Claude Code; most everyday work |
high | Start here on the API for hard reasoning and agentic coding |
xhigh | Only where you've measured a gain over high |
max | Measuring the model's ceiling, or rare tasks where time and tokens don't matter |
Anthropic's guidance is to start at high: gains above it are uneven. Effort barely changes how long the final answer is, so if you want shorter replies, say so in the prompt. And because Fable 5.2 is less in need of hand-holding, prompts written for older models are often too prescriptive and make results worse. Give it the full task, the constraints and the definition of done, then get out of the way.
Fable 5.2 or Opus 5.5? How I'd Decide
Use Opus 5.5 by default. It's half the price per token, it's what Anthropic recommends as the daily driver, and for most coding and knowledge tasks the quality difference won't show up.
Reach for Fable 5.2 when:
- The task is long-horizon: multi-hour agent runs, large refactors, research that needs to hold a plan together. That's where Remote Labor Index and RoboDojo show 2x gaps.
- The task is ambiguous and needs judgement about what to build, not just how.
- You're orchestrating a team of agents and need a strong planner on top.
- A cheaper model is failing and retrying. Fable at higher price may still cost less per finished task.
A pattern that works well: Fable 5.2 as the planner or advisor, and Opus 5.5, Sonnet 5.5 or Haiku 5.5 doing the bulk of the execution.
Migration Checklist: Fable 5.1 to Fable 5.2
- 01Change the model ID to
claude-fable-5-2. - 02Make conversation history append-only; stop rewriting or trimming earlier turns.
- 03Swap
computer_20251124forcomputer_toolset_20260801(unless you're on Bedrock). - 04If you use the advisor tool, set the advisor to
claude-fable-5-2orclaude-mythos-5-2. - 05Re-run your effort sweep, starting at
high, and includelowandmediumfor routine routes. - 06Keep refusal handling and fallbacks; expect fewer refusals than on 5.1.
- 07Re-check your cost dashboard: same tokens, 20% lower price.
My Take
The 5.2 release fixes the two things that kept Fable 5.1 out of a lot of production systems: price and false refusals. A 20% price cut plus 30% faster output plus 55% fewer classifier blocks makes it much easier to justify for the hardest parts of a pipeline.
It still isn't a default model. For the AI product work I do at Modelia, most calls are high-volume and narrow: attribute tagging, catalogue summaries, routing, structured checks. Those belong on Haiku 5.5 or Sonnet 5.5. Fable 5.2 earns its price on the few tasks that are long, ambiguous and expensive to get wrong, like planning a multi-step generation workflow or reviewing a large change across a codebase. Use it there, measure cost per finished task, and keep everything else on cheaper models.
The breaking change I'd watch is the history-editing check. If your agent harness compacts or edits history in place, you'll silently lose thinking blocks on 5.2. Audit that before you switch.
For the rest of the lineup, see my Claude Haiku 5.5 guide, the original Claude Fable 5 relaunch story, and my guides on Claude agentic workflows in production and the Model Context Protocol.
Sources: Anthropic: Claude Fable 5.2 and Mythos 5.2, Claude models overview, SD Times, 9to5Mac, Decrypt and The New Stack.
Harsh Rastogi is an AI Product Engineer and Business AI Head at Modelia, building enterprise-grade generative AI products for the fashion ecosystem, and the creator of carcode. He writes about AI systems, Claude and shipping AI to production at harshrastogi.tech.
Frequently asked questions
What is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's most capable widely released model, released on November 9, 2026 alongside Claude Mythos 5.2. It is built for the hardest reasoning, coding and long-running agentic work, with a 1M-token context window and 128K max output.
How much does Claude Fable 5.2 cost?
Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1's $10/$50. Cache reads are $0.20 per million tokens. Fast mode costs $16/$80 per million tokens.
What is the Claude Fable 5.2 model ID?
The Claude API model ID is claude-fable-5-2. The restricted Mythos counterpart is claude-mythos-5-2, available only to verified organizations.
Is Claude Fable 5.2 better than Opus 5.5?
On Anthropic's published benchmarks, Fable 5.2 beats Opus 5.5 on most coding and agentic evaluations, for example FrontierSWE v2 (69.9% vs 62.3%) and Remote Labor Index (41.3% vs 21.3%). But it costs twice as much per token, and Anthropic still recommends Opus 5.5 as the daily driver for most coding and knowledge work.
How is Claude Fable 5.2 different from Mythos 5.2?
They are the same underlying model. Fable 5.2 has stronger cyber and biology safeguards and is available to everyone. Mythos 5.2 relaxes those safeguards and is limited to organizations in Anthropic's Cyber Verification and Life Sciences Verification programs.
Can I disable thinking on Claude Fable 5.2?
No. Thinking is always on. Sending thinking type disabled or a budget_tokens value returns a 400 error. Control how much the model thinks with output_config.effort, from low to max.
What breaks when migrating from Fable 5.1 to Fable 5.2?
Four things: the preserved-thinking history-editing check applies to every account, so history must be append-only; earlier models can't read Fable 5.2 thinking blocks; computer_20251124 returns a 400 outside Amazon Bedrock, so use computer_toolset_20260801; and the advisor tool only accepts claude-fable-5-2 or claude-mythos-5-2 as advisor.
Is Claude Fable 5.2 faster than Fable 5.1?
Yes. Anthropic says Fable 5.2 generates output up to 30% faster than Fable 5.1. Optional fast mode increases output speed by up to 2.5x at double the price.
- Claude
- Anthropic
- Claude Fable 5.2
- LLM Pricing
- AI Agents
- Developer Tools
Written by Harsh Rastogi, AI Product Engineer and Business AI Head at Modelia. More on AI products, agents and production engineering on LinkedIn.