GenGrades

Claude Fable 5.2 vs Opus 5.5: review, price and benchmarks

Published Updated

AnthropicClaude Fable 5.2 logoAPI availableClosed weightsLicense: Proprietary (Anthropic commercial and consumer terms)

What Claude Fable 5.2 is

In one or two sentences, before anything else.

Claude Fable 5.2 is Anthropic's most capable public model, released November 9, 2026 at 8 and 40 USD per million input and output tokens, 20% below Fable 5.1. Against Opus 5.5 it costs double per token and, at max effort, scores 65 to 59 on Artificial Analysis's index for 26.54 USD per task against 10.04; Anthropic still recommends Opus first.

Main features

What it can do, one sentence each, and whether we actually tested it.

  • Thinking that is always on, set by effort

    Reasons before every answer at one of five effort levels, low, medium, high, xhigh or max; high is the API default, and the Claude apps and Claude Code run it at medium.

    Per Anthropic's docs and launch post. Artificial Analysis scores it 56 at medium for 3.80 USD per task, 62 at high for 9.08 USD and 65 at max for 26.54 USD, so each step up costs far more than it adds.

  • 1M-token context and 128K output

    Reads up to 1M tokens of text and images per request and writes up to 128K tokens, thinking included, in one response.

    Per Anthropic's model page. Simon Willison and PlayCode both saw max effort spend all 128,000 tokens on thinking and return no drawing at all.

  • 3D scenes and games written in code

    Builds interactive three.js and WebGL scenes, browser games and animations from a written brief, checking its own output in a browser as it goes.

    Artificial Analysis's AA-3D, where a panel votes blind on 55 three.js briefs, ranks it first of 25 models at 1271 Elo; Anthropic published a quote from Unity praising its 3D scenes.

  • Long agent runs and teams of subagents

    Works for hours or days on one goal, splitting the work across subagents and acting as an orchestrator or advisor for cheaper Claude models.

    Per Anthropic's launch post, an engineer had it coordinate over 1,000 subagents for eight days to build a prototype operating system; Ethan Mollick ran 28 agent teams on one 3D project.

  • Reads screenshots, charts and design mockups

    Takes images as input, so it can rebuild a page from a screenshot or design file and read charts and figures.

    Anthropic reports 73.4% on Chartography without tools and 74.1% on SWE-bench Multimodal, whose issues include screenshots and design mockups, against 61.4% for Claude Opus 5.5.

  • Fast mode

    Generates tokens up to 2.5 times faster for twice the price: 16 USD per million input tokens and 80 USD per million output tokens.

    Per Anthropic's launch post and fast-mode docs, on the Claude API, Claude Code, Google Cloud and Microsoft Foundry; not with the Batch API.

  • Cyber and biology safeguards with fallback

    Blocks high-risk cybersecurity work and most research-biology tasks, and hands a blocked request to Claude Opus 4.8 for cyber or Claude Opus 5 for biology instead.

    Per Anthropic's launch post, which counts about 55% fewer blocks than Fable 5.1 in early use. On the API a blocked request returns a refusal unless the developer switches on fallback.

  • Batch and cache discounts

    Halves the price for batch jobs, to 4 and 20 USD per million tokens, and bills cache reads at 0.20 USD, 2.5% of the input price.

    Per Anthropic's pricing docs. PlayCode ran its largest drawings through the Batch API, which accepted longer outputs than a streamed request.

Who it is for

The people it helps and the problems it solves, and where it ranks.

Built for

  • Creative technologists and developers who build three.js or WebGL scenes, interactive demos and browser games in code
  • Teams that run long agent sessions in Claude Code or the API and find Claude Opus 5.5 at high effort falls short
  • Front-end developers who rebuild pages and components from screenshots and design files
  • Researchers and analysts with long, open-ended tasks that end in a document, dataset or working prototype
  • Claude Max and Ultra subscribers, whose plans include Fable models up to half the weekly limit

Problems it solves

  • Large creative-coding projects, such as games and 3D worlds, that cheaper models leave unfinished
  • Interactive scenes that need many build, render and check cycles before they look right
  • Turning a visual reference into working code that stays true to it
  • Coordinating many agents or cheaper models on one long job

What it is good at

Public tests other people ran: what they gave it and how it did.

  1. Case 1

    Drawing an SVG at every effort level

    Input

    Simon Willison's standing prompt, "Generate an SVG of a pelican riding a bicycle", run through the API at all five effort levels on November 9, 2026, launch day.

    What happened

    His best drawing came from xhigh, after 32,966 output tokens and almost six minutes; low and medium answered in about 15 seconds with no thinking. Max thought through all 128,000 tokens for 22 minutes and returned no image, a failure he had not seen from Fable 5 or 5.1.

  2. Case 2

    A three.js animation from one line

    Input

    Simon Willison asked Fable 5.2 at medium effort, on November 10, 2026, for a three.js animation in which Yonkers stands up as a giant mech and stomps around New York, using OpenStreetMap data; the prompt carried a phone-keyboard typo.

    What happened

    It read the typo as "stomps" and built an animated 3D diorama: the mech rises, walks through four boroughs, flattens buildings underfoot and shows a stomp counter. The same prompt in Google's Antigravity with Gemini 4 Carbon gave a result he called pretty great, after one more prompt.

  3. Case 3

    3D scenes judged blind

    Input

    Artificial Analysis's AA-3D: 55 written briefs such as a spinning black hole, a playable pinball table or the Colosseum, each built as one three.js page by an agent in an offline sandbox, then compared in blind pairwise votes by a recruited panel; read November 11, 2026.

    What happened

    Fable 5.2 at max effort ranks first of 25 models with 1271 Elo, 67 points above Claude Sonnet 5.5 (1204) and 69 above Claude Opus 5.5 (1202). Its median scene took 105 turns, with about 40 browser checks and 40 image views per task.

  4. Case 4

    Finishing a 12-level game overnight

    Input

    A developer with early access gave Claude Code with Fable 5.2 one long prompt and a folder of a child's drawings to finish a four-player beat 'em up, letting it call an image model and ElevenLabs for art and sound (NotesByLex, November 10, 2026).

    What happened

    By morning it had rewritten most of the code, generated about 175 images and 140 audio clips, and written a bot that played all twelve stages in a headless browser. Over two days the project used about 600 requests; the author put the Claude tokens at roughly 300 USD at API prices.

  5. Case 5

    Hundreds of 3D worlds drawn by code

    Input

    Ethan Mollick described an endless flight through 3D worlds linked by portals in one paragraph and left the building to agents running Fable 5.2 before launch (published November 9, 2026).

    What happened

    Twenty hours later 204 worlds were live; the published code holds 471 worlds in about 1.4 million lines of JavaScript, with no 3D models, textures or sound files. The agents' own grading put 40 worlds in the top band and 28 in the bottom, and Mollick saw repeated themes and a rule book that slowed the work.

  6. Case 6

    An animated video and a game from a brand kit

    Input

    Claire Vo of the How I AI podcast spent two days building with Fable 5.2 before its launch episode on November 9, 2026: a marketing video from a single tweet and a brand book, and a bug-defense game with a trailer from brand sprites.

    What happened

    She describes the video as one that does not look like slides, and also lists web design as where it falls down, saying she still would not let it design a website on its own.

Where it falls short

What it does badly, with the evidence, and what we could not settle.

Weak

Weak

  • Expensive per finished task, especially at high effort.

    Evidence: Artificial Analysis counts 26.54 USD per index task at max effort against 10.04 for Claude Opus 5.5 and 7.76 for Gemini 4 Carbon; PlayCode's MacBook drawing at high cost 7.7 times Opus 5.5's. Decrypt reported Matthew Berman's largest builds, two games and an Excel clone, at about 11,700 to 14,710 USD each in estimated API cost.

  • Can spend its whole output budget thinking and return nothing.

    Evidence: Simon Willison's max-effort pelican and PlayCode's xhigh and max MacBooks each used all 128,000 output tokens on thinking and came back empty, 5.12 USD apiece at list price.

  • Slow at the settings that score best.

    Evidence: At max effort Artificial Analysis measured 247.85 seconds to the first answer token and 45 minutes per index task against a 12.6-minute median; Simon Willison's xhigh pelican took 5 minutes 54 seconds.

  • Not the voters' favourite for websites and games.

    Evidence: On DesignArena on November 11, 2026, it ranked fourth overall behind Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra, ninth on websites and seventh on single-file games; Opus 5.5 also leads Artificial Analysis's Code2Video-AA, where Fable 5.2 has no score yet.

  • Its taste needs a human to steer it.

    Evidence: Ethan Mollick found a first course site accurate and joyless and a second attempt childish, and says the model did not notice either problem; Anthropic's system card scores Mythos 5.2, the same model, below Mythos 5.1 on creative mastery.

  • Most cybersecurity work is blocked.

    Evidence: Anthropic says Fable 5.2's safeguards will block most cybersecurity-related tasks, including penetration testing and exploit writing; that work goes to Claude Mythos 5.2, which only verified organisations can use.

  • No zero data retention on the API by default.

    Evidence: Anthropic lists Fable 5.2 as a Covered Model that requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorizes it.

  • Not for everyday chat or routine coding, for budget creative work, for teams that need zero data retention, or for security researchers.

    Evidence: Anthropic itself tells most users to start with Claude Opus 5.5; Claude Sonnet 5.5 and Opus 5.5 rank second and third on AA-3D, about 70 Elo behind, at a quarter and half the token price; the retention rule and the cyber blocks are written into Anthropic's docs.

Uncertain

Uncertain

  • Whether its 3D lead holds as votes come in.

    Why unsure: AA-3D ranks it clearly first, but on DesignArena's 3D board it sits 5 Elo behind GPT-6 Astra with a margin of 5.1 points after only 266 battles, three days after launch.

  • How much of the gain over Opus 5.5 comes from spending more tokens.

    Why unsure: PlayCode found that, matched by tokens, Fable 5.2 at high drew about what Opus 5.5 drew at xhigh, while Artificial Analysis scores Fable 5.2 at high above Opus 5.5 at max for slightly less per task. One drawing per setting cannot settle it.

  • How it compares with Gemini 4 Carbon on creative coding.

    Why unsure: Carbon opened to all paying API users a day after Fable 5.2 and had only 221 DesignArena battles and no AA-3D score on November 11; the only head-to-head creative prompt found was Simon Willison's Yonkers animation, which he did not rank.

  • How well it drives Blender.

    Why unsure: DesignArena's Blender board, where agents write Blender Python, had no Fable 5.2 result, and the evidence is one Hacker News user's report plus Anthropic's own CAD benchmark.

Price, limits and formats

The numbers, each with its source and the date we checked it.

SpecValueNote and source
ReleasedNovember 9, 2026

Successor to Claude Fable 5.1, released in September 2026, per The Verge.

Source: platform.claude.comchecked

MakerAnthropic

Source: anthropic.comchecked

Model IDclaude-fable-5-2 (Amazon Bedrock: anthropic.claude-fable-5-2)

Source: platform.claude.comchecked

SiblingClaude Mythos 5.2: the same model with looser cyber and biology safeguards, for verified organisations only

Source: platform.claude.comchecked

Context window1M tokens

Source: platform.claude.comchecked

Maximum output128K tokens, thinking and answer together

Synchronous Messages API limit. PlayCode's Batch API runs finished with up to 250,082 tokens.

Source: platform.claude.comchecked

Input and outputText and images in; text out

Source: platform.claude.comchecked

ThinkingAdaptive, always on; effort low, medium, high, xhigh or max

API default high; the Claude apps and Claude Code default to medium, per the launch post.

Source: platform.claude.comchecked

Knowledge cutoffReliable through June 2026

Source: platform.claude.comchecked

API price8.00 USD per million input tokens, 40.00 output, 0.20 cache reads

Cache writes cost 10.00 USD per million tokens for 5 minutes and 16.00 for 1 hour.

Source: platform.claude.comchecked

Batch price4.00 USD input, 20.00 output per million tokens

Source: platform.claude.comchecked

Fast modeUp to 2.5 times faster at 16.00 USD input and 80.00 output per million tokens

Source: anthropic.comchecked

US only inference1.1 times the standard token price

Source: claude.comchecked

PlatformsClaude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry

Source: platform.claude.comchecked

Claude plansMax (100 USD a month) and Ultra (200 USD) include Fable up to half the weekly limit; Pro (20 USD) pays with usage credits; Free has none

Source: claude.comchecked

Not accepted by the APItemperature, top_p, top_k, budget_tokens, thinking switched off, forced tool use, assistant prefill

Source: platform.claude.comchecked

Data retention30 days required; zero data retention only if Anthropic expressly authorizes it

Source: platform.claude.comchecked

RetirementNot sooner than November 9, 2027

Source: platform.claude.comchecked

Artificial Analysis Intelligence Index65 at max effort, first of 74 models (v5.0.1)

62 at high effort and 56 at medium.

Source: artificialanalysis.aichecked

AA 3D Elo1271, first of 25 models

Source: artificialanalysis.aichecked

Common questions

Answer first, then why, an example, the limits and what to do next.

Claude Fable 5.2 vs Claude Opus 5.5: which should I use?

Start with Claude Opus 5.5, which Anthropic itself recommends for most work at half the token price. Move to Fable 5.2 for long, open-ended jobs and 3D scenes where Opus falls short: Anthropic reports 49.1% against 45.5% on FrontierCode 2.0, and Artificial Analysis scores Fable 62 at high effort against Opus's 59 at max for slightly less per task. Opus still ranks higher on DesignArena's overall, website and game boards and leads its Blender board and Code2Video-AA.

What is Claude Fable 5.2?

Claude Fable 5.2 is Anthropic's most capable model for general use, released on November 9, 2026 for demanding reasoning and long agent work. Claude Mythos 5.2 is the same underlying model with looser cybersecurity and biology safeguards, open only to organisations Anthropic has verified. Fable 5.2 reads text and images, writes text, holds 1M tokens of context and always thinks before it answers.

What changed from Claude Fable 5.1 to Fable 5.2?

Fable 5.2 is cheaper, faster and stronger than Fable 5.1, and its API has four breaking changes. Tokens cost 8 and 40 USD per million instead of 10 and 50, Anthropic says outputs come up to 30% faster with about 55% fewer blocked requests, and its FrontierCode 2.0 score rose from 32.4% to 49.1%. AA-3D rates it 1271 against 1139 for Fable 5.1. Code that edits earlier turns, uses the older computer-use tool or pairs the advisor tool with Fable 5.1 has to change.

How much does Claude Fable 5.2 cost?

On the API, Claude Fable 5.2 costs 8 USD per million input tokens and 40 USD per million output tokens, with cache reads at 0.20 USD, batch jobs at half price and fast mode at double. Its thinking is billed as output, so Artificial Analysis counts 9.08 USD per index task at high effort and 26.54 USD at max. In the Claude apps, Max at 100 USD a month and Ultra at 200 USD include it up to half the weekly limit, while Pro at 20 USD pays for it with usage credits.

Is Claude Fable 5.2 free?

Claude Fable 5.2 is not free: the Free plan leaves Fable models out, and on Pro they run on paid usage credits. During launch week, until November 16, 2026, Pro plans and standard Team seats had Fable models included up to half their weekly limit, and Anthropic warns that Fable models use up limits faster than other Claude models.

Claude Fable 5.2 vs Gemini 4 Carbon: which is better?

Fable 5.2 scores higher on most published tests, and Gemini 4 Carbon costs far less. Artificial Analysis rates them 65 and 60, Anthropic's own table puts Fable ahead on coding and science but behind Carbon on AutomationBench (51.4% against 53.9%), and Carbon costs 2 and 10 USD per million tokens until November 30, 2026, then 4 and 20, against 8 and 40. Carbon also takes video, audio and PDF input; on creative coding it is too new for the vote-based boards to compare.

How does Claude Fable 5.2 compare with GPT-6 Astra?

Fable 5.2 leads GPT-6 Astra on Anthropic's launch table and Artificial Analysis's index, while Astra costs far less per task and wins more DesignArena boards. Artificial Analysis rates them 65 and 57 at 26.54 and 6.24 USD per index task; Anthropic reports 49.1% against 42.8% on FrontierCode 2.0 and 41.3% against 20.8% on the Remote Labor Index. DesignArena voters put Astra first on 3D, single-file games and HTML code and one place above Fable overall, while AA-3D puts Fable ahead, 1271 to 1168.

What is the difference between Claude Fable 5.2 and Claude Mythos 5.2?

Only the safeguards and who may use them differ. Both are one model with the same specs and prices; Fable 5.2 blocks most cybersecurity and research-biology tasks and is open to everyone, while Mythos 5.2 lets verified organisations in those fields do that work. On Anthropic's CVE Rediscovery test Mythos 5.2 found 89.6% of the published vulnerabilities it was set.

Is Claude Fable 5.2 good at design and creative coding?

For 3D scenes, public tests say yes: it ranks first on Artificial Analysis's AA-3D and second on DesignArena's 3D board, and Simon Willison and Ethan Mollick published striking three.js work made with it. For websites and single-file games, DesignArena voters prefer Claude Opus 5.5 and GPT-6 Astra, and two early users said its taste in web design or course design needed a human to steer it.

Which effort level should I use with Claude Fable 5.2?

Anthropic recommends starting at high, the API default, stepping up to xhigh only for hard coding and agent work where you measure a gain, and keeping max for rare cases. Artificial Analysis shows why: high scores 62 for 9.08 USD per task, max adds 3 points for nearly three times the cost. Below high, Anthropic says turns end before the work is done more often.

Why did Claude Fable 5.2 return an empty answer?

Most likely it spent its whole 128,000-token output limit on thinking, which counts against the same limit as the answer. Simon Willison and PlayCode both saw this at max effort, and PlayCode also at xhigh, each time paying about 5.12 USD for nothing. Use high effort, ask for shorter work, or send very long jobs through the Batch API, where PlayCode's runs finished.

Does Anthropic train on what I send to Claude Fable 5.2?

Not on the API: Anthropic's docs say retained data is never used for model training without your express permission, though Fable 5.2 requires 30-day retention for safety monitoring. In the Claude apps, chats are used for training only if you allow model improvement, in which case de-identified data can be kept up to five years; incognito chats are never used.

What the public tests show

This review is built from what could be checked on November 11, 2026: Anthropic's launch post, system card, model documentation, prices and data rules; independent results from Artificial Analysis, including its AA-3D benchmark of three.js scenes, from DesignArena's vote-based boards and from PlayCode's SVG benchmark; public experiments by Simon Willison, Ethan Mollick, the How I AI podcast and a developer's game write-up; Decrypt's report on Matthew Berman's builds; user reports on Hacker News; and the price pages of rival models. Benchmark figures are Anthropic's own claims unless another source is named.

What Claude Fable 5.2 is

Claude Fable 5.2 is the most capable model Anthropic sells to all customers. Anthropic released it on November 9, 2026 as the successor to Claude Fable 5.1 (model page), and The Verge and CNBC covered the launch the same day (The Verge, CNBC). Claude Fable and Claude Mythos are the same underlying model: Fable carries strong cybersecurity and biology safeguards so that anyone can use it, while Mythos 5.2 goes only to organisations verified through Anthropic's programmes (launch post).

It reads text and images, writes text, holds 1M tokens of context and writes up to 128K tokens per response. It always thinks before answering; a request sets the effort to low, medium, high, xhigh or max, high is the API default, and the Claude apps and Claude Code run it at medium (what's new, launch post).

Claude Fable 5.2 vs Claude Opus 5.5

Claude Opus 5.5, released in September 2026, is Anthropic's everyday model, and Fable 5.2 costs twice as much per token. Both have a 1M-token context and 128K output. Side by side on November 11, 2026:

Claude Fable 5.2Claude Opus 5.5
API price, USD per million tokens in and out8.00 and 40.004.00 and 20.00
Default effort on the APIHighMedium
FrontierCode 2.0, Anthropic49.1% at xhigh; 45.3% at high45.5% at its best
FrontierSWE v2, Anthropic69.9%62.3%
Terminal-Bench-Science 0.1, Anthropic74.6%58.7%
Remote Labor Index, Anthropic41.3%21.3%
Intelligence Index, Artificial Analysis62 at high; 65 at max59 at max
Cost per index task, Artificial Analysis9.08 USD at high; 26.54 USD at max10.04 USD at max
Minutes per index task, Artificial Analysis16.9 at high; 45.0 at max25.0 at max
AA-3D, three.js scenes1271 Elo, 1st1202 Elo, 3rd
DesignArena overall agentic1358, 4th1441, 1st (high effort)
DesignArena 3D design1513, 2nd1461, 3rd
DesignArena websites1318, 9th1324, 7th
DesignArena single-file games1399, 7th1428, 2nd
DesignArena BlenderNo result yet1412, 1st (max effort)
Code2Video-AA, videos written in codeNo result yet1536, 1st
PlayCode MacBook SVG2nd of 341st

Sources: Anthropic pricing, models overview, launch post; Artificial Analysis on Fable 5.2 high, Fable 5.2 max, Opus 5.5, AA-3D and Code2Video-AA; DesignArena overall, 3D, websites, games and Blender; PlayCode; Opus 5.5's release month from CNBC.

Anthropic's headline claim is about cost as much as score: at high effort, it says, Fable 5.2 scored about level with Opus 5.5's best FrontierCode 2.0 result for roughly half the cost per task. Artificial Analysis's figures in the table point the same way; at max effort the trade turns against Fable.

Fable 5.2 is worth twice the token price on long, open-ended work, where Anthropic reports big gains such as the Remote Labor Index and Terminal-Bench-Science, and on three.js scenes, where AA-3D puts it 69 Elo ahead. Run it at high rather than max.

Opus 5.5 is the better pick for everyday coding, websites, single-file games, Blender models and videos written in code, where voters and Code2Video-AA rank it higher or Fable has no result yet. It also writes fewer tokens at the same effort label: on PlayCode's MacBook drawing at high, Opus wrote 19,767 tokens for 0.40 USD and Fable 76,158 for 3.05 USD, and matched by tokens Fable at high drew about what Opus drew at xhigh. Even Anthropic's own results have Opus ahead on two Elo ratings of office work, AA-Briefcase (1800 to 1785) and GDPval-AA (1866 to 1831) (system card), and one Hacker News user found the two hard to tell apart on ordinary coding (Hacker News).

Anthropic says the same. Its docs tell developers to start with Opus 5.5 and move to Fable 5.2 when Opus at higher effort still falls short (models overview), its launch post calls Opus 5.5 its strongest daily driver, and its head of product management for research told CNBC that Fable 5.2 is mainly for "power users or power use cases" (CNBC).

What changed from Claude Fable 5.1

Claude Fable 5.1Claude Fable 5.2
API price, USD per million tokens in and out10.00 and 50.008.00 and 40.00
Cache reads, USD per million tokens0.250.20
FrontierCode 2.0 at xhigh, Anthropic32.4%49.1%
Terminal-Bench 4.0, Anthropic55.8%67.6%
AA-3D, Elo11391271
DesignArena overall agentic, Elo13271358

Sources: pricing, launch post, system card, AA-3D, DesignArena.

  • Price: input and output are 20% cheaper, and Anthropic says outputs come up to 30% faster (launch post).
  • Safeguards: in early use across claude.ai, Claude Code and Cowork, Anthropic counted about 55% fewer blocked requests.
  • API changes: edits to earlier turns now invalidate thinking blocks on every account, only Fable 5.2 and Mythos 5.2 can read Fable 5.2's thinking, the older computer-use tool is rejected outside Amazon Bedrock, and the advisor tool needs a Fable 5.2 or Mythos 5.2 advisor (what's new).

What it costs

On the API, Fable 5.2 is billed per token at one price across its 1M context. In USD per million tokens on November 11, 2026 (pricing):

ItemClaude Fable 5.2
Input8.00
Output, thinking included40.00
Cache read0.20
Cache write, 5-minute lifetime10.00
Cache write, 1-hour lifetime16.00
Batch input and output4.00 and 20.00
Fast mode input and output16.00 and 80.00

US-only inference costs 1.1 times these prices (Claude pricing). Because thinking is billed as output, effort decides the bill more than the price list does. Artificial Analysis ran its index at three settings (medium):

EffortIntelligence IndexCost per taskMinutes per task
Medium563.80 USD7.7
High629.08 USD16.9
Max6526.54 USD45.0

Max adds three points over high for nearly three times the cost.

In the Claude apps (Claude plans):

PlanPriceFable 5.2
Free0 USDNot included
Pro20 USD a month, or 200 USD a yearPaid with usage credits
Max100 USD a monthIncluded up to half the weekly limit
Ultra200 USD a monthIncluded up to half the weekly limit; no 5-hour limit for verified accounts

For launch week, until November 16, 2026, Pro plans and standard Team seats had Fable models included up to half the weekly limit plus one reset; Max, Ultra and premium seats got one to three resets that expire on March 8, 2027. Anthropic's help article warns that Fable models use limits faster than other Claude models (launch offers). One Hacker News user's code review used a 5-hour allowance in half an hour (Hacker News).

Big projects get expensive. Decrypt reported that Matthew Berman's Gears of War-style level ran for 43 hours 48 minutes and 8.63 billion tokens, about 14,710 USD at API prices, and that his Excel clone came to about 11,700 USD (Decrypt).

Creative coding: what public tests show

Most public evidence on Fable 5.2's taste comes from things built in code to be looked at:

SVG drawing. Simon Willison asked it for his standard "pelican riding a bicycle" SVG at all five effort levels on launch day (Simon Willison). His run log shows low and medium answering in 14 and 15 seconds with no thinking, high in 59 seconds, and xhigh after 5 minutes 54 seconds and 32,966 output tokens (run log). He judged xhigh best, with a good frame and pouch. At max it thought for 22 minutes, hit the 128,000-token limit and returned nothing. PlayCode, which sells an AI app builder that offers these models, found no fault in Fable 5.2's medium MacBook drawing, which took 3.91 USD and 16 minutes, while its xhigh and max runs came back empty.

Three.js and WebGL scenes. This is where the independent results are strongest. Artificial Analysis's AA-3D, in which a recruited panel votes blind between three.js scenes built from the same brief, ranks Fable 5.2 first of 25 models, ahead of Claude Sonnet 5.5 (1204) and GPT-6 Astra (1168). DesignArena's 3D board, where visitors try two single-file 3D pages and pick one, puts it five points behind GPT-6 Astra after 266 battles. Simon Willison's one-line prompt produced Yonkers Rises, an animated diorama in which a mech stomps through New York (Yonkers Rises). Ethan Mollick's Portals project, built by teams of agents, reached 471 worlds and about 1.4 million lines of JavaScript, each world built and its sound synthesized in the browser, with no 3D model, texture or sound files (Portals code). Anthropic also published a Unity creative technologist's praise for its 3D scenes.

Games. A developer with early access let Fable 5.2 finish a 12-level, four-player beat 'em up overnight from one prompt and a child's drawings; it wrote a bot that played every stage in a headless browser to check its work, while the sprites came from an image model (NotesByLex). One Hacker News reader found an exploit within seconds of playing (Hacker News). Saltline, a small first-person shooter built with Fable 5.2, drew comments that the graphics were pretty and the gameplay mid, with reports of broken strafing and of the game not running in Firefox (Hacker News).

Videos made in code. Ethan Mollick asked for the most impressive thing it could make for his readers and got, in one shot, a video built entirely in code that he places at the level of a human expert (One Useful Thing). Claire Vo's How I AI episode shows an animated marketing video made from a single tweet and a brand book (How I AI). Artificial Analysis's Code2Video-AA had no Fable 5.2 score yet; its predecessor Fable 5.1 sits fifth there at 1325.

Blender and 3D models. One Hacker News user calls Fable 5.2 really good at using Blender, while adding that most agent models can now make low-poly assets (Hacker News). Anthropic's own evidence is CAD rather than Blender: 83.5 on its BenchCAD 2.1 against 65.3 for Opus 5.5, and three devices whose cases it modelled in CAD, built with Diode, a company Anthropic has a commercial relationship with (Anthropic video).

Websites and front-end design. A Hacker News user who rebuilds a hard design from images found it truer to the source than Opus 5.5, with better SVG curves and icons, but preferred Opus 5.5's motion (Hacker News). Claire Vo lists web design as where it falls down.

Taste. Another early user calls Fable "exceptional at drawing" with JavaScript; others reported a grammar slip (Ask HN) and a design discussion in which it drifted into wrong answers that Opus 5.5 avoided (Hacker News). The counterweight comes from Ethan Mollick, whose course site came out accurate and joyless, then childish when he asked for beauty, and whose 3D worlds repeated themes across agent teams; he writes that knowing when its own work is not good enough is the part Fable does least reliably. Anthropic's system card adds that Mythos 5.2, the same model, scores below Mythos 5.1 on creative mastery as judged by another Claude model (system card).

Benchmarks: Fable 5.2, Gemini 4 Carbon and GPT-6 Astra

TestWho ran itFable 5.2Gemini 4 CarbonGPT-6 Astra
FrontierCode 2.0Anthropic49.1% (xhigh)41.1% (medium)42.8%
FrontierSWE v2Anthropic69.9%61.2%65.5%
AutomationBenchAnthropic51.4%53.9%41.4%
Terminal-Bench-Science 0.1Anthropic74.6%64.8%68.1%
Chartography, no toolsAnthropic73.4%72.5%71.0%
Intelligence Index v5.0.1Artificial Analysis65 (max)60 (high)57 (max)
Cost per index taskArtificial Analysis26.54 USD7.76 USD6.24 USD
AA-3D, EloArtificial Analysis1271, 1st1168
Overall agentic, EloDesignArena voters1358, 4th1270, 10th1386, 3rd

Sources: launch post; Artificial Analysis on Carbon and Astra; read November 11, 2026.

Anthropic's table puts Fable 5.2 ahead of both except on AutomationBench, where Carbon leads. Independently, both rivals cost under a third as much per Artificial Analysis task, and DesignArena ranks Astra one place higher overall, while AA-3D puts Fable 103 Elo ahead of it.

Claude Fable 5.2 vs Gemini 4 Carbon

Claude Fable 5.2Gemini 4 Carbon
MakerAnthropicGoogle
API price, USD per million tokens in and out8.00 and 40.002.00 and 10.00 until November 30, 2026; 4.00 and 20.00 from December 1
Context window1M tokens1,048,576 tokens
Maximum output128K tokens65,536 by default, up to 1,048,576 with long decode
InputText and imagesText, images, video, audio and PDF
Default reasoningHigh effortMedium thinking
DesignArena battles by November 111,064221

Sources: Anthropic model page, Google model page, Google pricing, what's new in Carbon.

Google pitches Carbon at deeper build-and-check loops, including front-end UI design and 3D scenes, and opened it to all paying API users a day after Fable 5.2's launch, as Simon Willison noted (Simon Willison). He ran the Yonkers prompt on Carbon in Google's Antigravity and called the result pretty great, without ranking it against Fable's (Hacker News). Hacker News users disputed which of the two costs more per task (Hacker News). Fable 5.2 has the stronger independent record on 3D work today; Carbon is the cheaper bet, at least until its price doubles on December 1.

Using it through the API

Fable 5.2 takes the same Messages API as recent Claude models, with stricter rules (model page, migration guide):

  • Thinking cannot be turned off; sending a thinking budget or disabling thinking returns an error.
  • temperature, top_p and top_k must be left out, and assistant prefill is rejected.
  • Forced tool use returns an error; strict tool schemas or JSON outputs are the replacement.
  • The 128K output limit covers thinking and answer together, so long work needs a large max_tokens, a lower effort or the Batch API.
  • Requests that trip a safety classifier return a refusal; server-side fallback can rerun them on Claude Opus 4.8 or Claude Opus 5 (what's new).

Anthropic's prompting guide notes behaviour worth planning for: below high effort the model sometimes stops with a list of what remains, it follows style rules very literally, and telling it the code is its own makes it far more willing to restructure (prompting guide).

Privacy, data and safety

On the API, Fable 5.2 is a Covered Model: it requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorizes it. Anthropic's docs state that retained data is never used for model training without the customer's express permission (data retention). Eligible enterprise customers can keep zero retention through Enterprise Frontier Safeguards, a private beta that runs the safety monitoring inside their own infrastructure (launch post).

In the Claude apps, chats are used for training only if you allow model improvement; then de-identified data can be kept for up to five years. Deleted chats leave Anthropic's back-end within 30 days, incognito chats are never used, and chats flagged for policy violations can be kept for two years (Anthropic Privacy Center).

Fable 5.2's classifiers block high-risk cyber work, such as penetration testing and exploit writing, and most research-biology tasks. Anthropic also capped its AI-research abilities, such as accelerator kernel optimisation, at about Opus 5.5's level. In its behavioural audit the shared model misled users less than Opus 5.5 on most measures, but voiced suspicion of being evaluated more often than any recent Claude model (launch post).

Who should not use Claude Fable 5.2

  • Anyone whose work is chat, writing or routine coding: Claude Opus 5.5 does it for half the token price.
  • Teams on a budget for creative coding: Claude Sonnet 5.5 and Opus 5.5 rank second and third on AA-3D, about 70 Elo behind, at a quarter and half the token price.
  • Anyone who wants a website designed without supervision: DesignArena voters rank it ninth on websites, and Claire Vo would not hand it that job.
  • Users who need fast replies: on Artificial Analysis's runs the first answer token took over four minutes at max effort and 30 seconds at high.
  • Organisations that require zero data retention and have no authorization from Anthropic.
  • Security researchers and biology labs, whose work is mostly blocked without Mythos 5.2 access.
  • Anyone running max effort unattended: it can think through its whole output limit and return nothing.

Alternatives

Prices from each maker's own price page on November 11, 2026, in USD per million tokens, with Artificial Analysis's index score. Claude Opus 5.5 has its own section above.

ModelMakerInputOutputAA indexAA-3D Elo
Claude Fable 5.2Anthropic8.0040.00651271
Claude Sonnet 5.5Anthropic2.0010.00551204
Gemini 4 CarbonGoogle2.00 until November 30, then 4.0010.00, then 20.0060
Gemini 4 ArgonGoogle2.00 until November 30, then 4.0010.00, then 20.00501147
Kimi K3.1 ProMoonshot AI2.0010.00461112

Sources: Anthropic, Google, Moonshot; Artificial Analysis on Sonnet 5.5 and Kimi K3.1 Pro; AA-3D.

Claude Sonnet 5.5 is the budget pick for 3D and video work: second on AA-3D and on Code2Video-AA at a quarter of Fable's token price.

Gemini 4 Carbon is the one to choose for video, audio or PDF input, or for heavy use before its price doubles.

GPT-6 Astra from OpenAI is the pick when DesignArena's voters matter most: it leads the 3D, single-file game and HTML code boards.

Kimi K3.1 Pro suits buyers who want open weights at 2 and 10 USD per million tokens, though it has the lowest scores in the table.

To try Fable 5.2, a short API run at high effort costs a few dollars, and Max and Ultra plans include it; for long creative builds, budget the tokens first.

Sources

Sources

  1. Introducing Claude Fable 5.2 and Claude Mythos 5.2: benchmarks, prices, safeguards (Anthropic, November 9, 2026) (anthropic.com), accessed
  2. Claude Fable 5.2 model page: released November 9, 2026, specs and prices (Claude Platform docs) (platform.claude.com), accessed
  3. What's new in Claude Fable 5.2: breaking changes, effort, pricing, availability (Claude Platform docs) (platform.claude.com), accessed
  4. Claude Fable 5.2 migration guide: prefill and sampling parameters (Claude Platform docs) (platform.claude.com), accessed
  5. Prompting Claude Fable 5.2 (Claude Platform docs) (platform.claude.com), accessed
  6. Effort: recommended levels for Claude Fable 5.2 (Claude Platform docs) (platform.claude.com), accessed
  7. Models overview: Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5 (Claude Platform docs) (platform.claude.com), accessed
  8. Pricing: model, cache, batch and fast-mode prices (Claude Platform docs) (platform.claude.com), accessed
  9. Claude plans and API pricing (claude.com) (claude.com), accessed
  10. Claude Fable 5.2 launch offers (Claude Help Center) (support.claude.com), accessed
  11. Claude Fable 5.2 and Claude Mythos 5.2 System Card (Anthropic, November 9, 2026) (anthropic.com), accessed
  12. API and data retention: Covered Models need 30-day retention (Claude Platform docs) (platform.claude.com), accessed
  13. Fast mode: availability and limits (Claude Platform docs) (platform.claude.com), accessed
  14. How long do you store my data? (Anthropic Privacy Center) (privacy.claude.com), accessed
  15. Claude Fable 5.2 designs a working voice-assistant hub (Claude, YouTube, November 9, 2026) (youtube.com), accessed
  16. Anthropic launches new Claude Fable and Mythos 5.2 models (The Verge, November 9, 2026) (theverge.com), accessed
  17. Anthropic releases first frontier models since Amodei's call for slowdown (CNBC, November 9, 2026) (cnbc.com), accessed
  18. Anthropic's Claude Fable 5.2 is impressive, but be prepared to pay (Decrypt, November 9, 2026) (decrypt.co), accessed
  19. Claude Fable 5.2 (Max): intelligence, speed and price (Artificial Analysis) (artificialanalysis.ai), accessed
  20. Claude Fable 5.2 (High) (Artificial Analysis) (artificialanalysis.ai), accessed
  21. Claude Fable 5.2 (Medium) (Artificial Analysis) (artificialanalysis.ai), accessed
  22. Claude Opus 5.5 (Max) (Artificial Analysis) (artificialanalysis.ai), accessed
  23. Claude Sonnet 5.5 (Max) (Artificial Analysis) (artificialanalysis.ai), accessed
  24. Gemini 4 Carbon (High) (Artificial Analysis) (artificialanalysis.ai), accessed
  25. GPT-6 Astra (Max) (Artificial Analysis) (artificialanalysis.ai), accessed
  26. Kimi K3.1 Pro (Artificial Analysis) (artificialanalysis.ai), accessed
  27. AA-3D: interactive three.js scenes ranked by blind votes (Artificial Analysis) (artificialanalysis.ai), accessed
  28. Code2Video-AA: product videos written in HTML and GSAP (Artificial Analysis) (artificialanalysis.ai), accessed
  29. Overall agentic leaderboard, updated November 11, 2026 (DesignArena) (designarena.ai), accessed
  30. 3D design leaderboard, updated November 11, 2026 (DesignArena) (designarena.ai), accessed
  31. Blender leaderboard, updated November 11, 2026 (DesignArena) (designarena.ai), accessed
  32. Game dev leaderboard, updated November 11, 2026 (DesignArena) (designarena.ai), accessed
  33. Website leaderboard, updated November 11, 2026 (DesignArena) (designarena.ai), accessed
  34. MacBook SVG benchmark: 34 AI models ranked (PlayCode, updated November 9, 2026) (playcode.io), accessed
  35. llm-anthropic 0.31: Fable 5.2 pelicans at five effort levels (Simon Willison, November 9, 2026) (simonwillison.net), accessed
  36. Fable 5.2 pelican runs with token counts and durations (Simon Willison, GitHub gist) (gist.github.com), accessed
  37. Yonkers Rises: a three.js animation by Fable 5.2 (Simon Willison, November 10, 2026) (simonwillison.net), accessed
  38. llm-gemini 0.36: Gemini 4 Carbon pelicans and the same three.js prompt (Simon Willison, November 10, 2026) (simonwillison.net), accessed
  39. Fable 5.2 and Jagged AGI (Ethan Mollick, One Useful Thing, November 9, 2026) (oneusefulthing.org), accessed
  40. Portals: 471 code-generated 3D worlds, README (emollick/ai-portals, GitHub) (github.com), accessed
  41. Fable 5.2 one-shotted my nephew's dream game (NotesByLex, November 10, 2026) (notesbylex.com), accessed
  42. I was a Claude hater. Fable 5.2 changed my mind. (How I AI, YouTube, November 9, 2026) (youtube.com), accessed
  43. User report: 'Fable is exceptional at drawing' (Hacker News comment) (news.ycombinator.com), accessed
  44. User report: image-to-HTML comparison with Opus 5.5 (Hacker News comment) (news.ycombinator.com), accessed
  45. User report: a code review used the 5-hour budget in half an hour (Hacker News comment) (news.ycombinator.com), accessed
  46. User report: drifted and gave wrong answers in a design discussion (Hacker News comment) (news.ycombinator.com), accessed
  47. Game author's cost estimate at API prices (Hacker News comment) (news.ycombinator.com), accessed
  48. Fable 5.2 one-shotted my nephew's dream game (Hacker News discussion) (news.ycombinator.com), accessed
  49. Saltline: first-person shooter built by Fable 5.2 (Hacker News discussion) (news.ycombinator.com), accessed
  50. User report: Fable 5.2 with Blender (Hacker News comment) (news.ycombinator.com), accessed
  51. Ask HN: anyone else notice Fable 5.2 making grammar errors? (Hacker News) (news.ycombinator.com), accessed
  52. Gemini 4 Carbon is now available (Hacker News discussion) (news.ycombinator.com), accessed
  53. Gemini 4 Carbon pelicans and three.js result (Hacker News comment) (news.ycombinator.com), accessed
  54. What's new in Gemini 4 Carbon (Google AI for Developers) (ai.google.dev), accessed
  55. Gemini 4 Carbon model page: inputs, context, output (Google AI for Developers) (ai.google.dev), accessed
  56. Gemini Developer API pricing: Gemini 4 Carbon and Argon (Google) (ai.google.dev), accessed
  57. Model inference pricing: Kimi K3.1 Pro (Kimi API Platform) (platform.kimi.ai), accessed

Changelog

Changelog

  1. Page created from Anthropic's own pages, system card, documentation, prices and data rules, results from Artificial Analysis, DesignArena and PlayCode, public experiments by Simon Willison, Ethan Mollick, How I AI and NotesByLex, Decrypt's report, user reports on Hacker News, and the price pages of rival models.

Recommended from GenGrades

  • Tool review

    Devin

    Devin is Cognition's autonomous coding agent: it works on its own cloud computer and opens pull requests, and ships with Devin Desktop (formerly Windsurf) and a CLI. In November 2026 Cognition added the cheaper SWE-2 Flash model and Devin Review 2.0. Plans cost US$0 to US$200 a month; the best-known outside test, Answer.AI's, saw 3 successes in 20 tasks.

  • Model review

    Kimi K3.1 Pro

    Kimi K3.1 Pro is Moonshot AI's open-weight 2.8-trillion-parameter model with a 1M-token context, released October 23, 2026, weights out November 4. Its API costs 2 and 10 USD per million tokens, half Claude Opus 5.5's. Vals AI puts it near Opus on Terminal-Bench 4.0, but independent aggregates rank it lower, it is slow, and an audit questions its benchmark lead.

  • Model review

    Inkling 2 Small

    Inkling 2 Small is Thinking Machines Lab's 594-billion-parameter mixture-of-experts model (17 billion active) for text and images. It launched on Tinker on November 5, 2026 with closed weights, promised under Apache 2.0 after safety testing, at 0.30 USD per million input tokens and 1.00 output. Artificial Analysis scores it 31, behind DeepSeek V4.1 Flash, and finds it slow.

  • Model review

    Mistral Large 4

    Mistral Large 4 is Mistral AI's open-weight mixture-of-experts model, 1 trillion parameters with 52 billion active, previewed on October 6, 2026, with weights out October 29. Its API costs 0.70 USD per million input tokens and 2.10 output. Artificial Analysis scores it 26, behind GLM-5.3 and DeepSeek V4.1 Flash at 32, and self-hosting its full checkpoint takes eight data-centre GPUs.

  • Model review

    LongCat-2.5

    LongCat-2.5 is Meituan's 1.6-trillion-parameter model for long coding and agent tasks, released on November 2, 2026; unlike LongCat-2.0, it also reads images and video. Its API costs 0.30 USD per million input tokens and 1.20 USD per million output at a launch discount. The weights are not out yet, and every benchmark so far is Meituan's own.

  • Tool review

    Venice AI

    Venice AI is a chat, image, video and API service from Venice.ai, founded in 2024 by Erik Voorhees, that adds no content filter and keeps chat history on your device. It raised $65 million at a $1 billion valuation in July 2026. Pro costs $18 a month, but prompts sent to Claude, GPT or Gemini still reach their makers.