Claude Fable 5.2 and Mythos 5.2: what's new, the benchmarks, and what it means for your Mac

MacPip is an independent Mac app directory. It is not affiliated with, sponsored by or endorsed by Anthropic. Every number below links to where it was published, and was read on November 11, 2026.

What was released

On November 9, 2026, Anthropic put out two models in one announcement: Claude Fable 5.2 and Claude Mythos 5.2. Its own one-line description on the news page is "Our most capable models for daunting technical work".

Here is what changed, in the order I think matters:

• One model, two names. Anthropic says Fable 5.2 and Mythos 5.2 are the same underlying model. Fable 5.2 has cybersecurity and biology safeguards and is the one you can use. Mythos 5.2 is for vetted organizations only.
• It got cheaper. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Fable 5.1 was $10 and $50, so that is 20% less.
• It got faster. Anthropic says output comes up to 30% faster than the 5.1 models. A fast mode is up to 2.5 times faster again, at $16 and $80.
• It stops less often. Anthropic counted about 55% fewer blocks with Fable 5.2 in early testing: about 60% fewer in cybersecurity and 40% fewer in biology.
• It uses fewer tokens on hard tasks at medium effort, the default in the Claude apps and Claude Code. Anthropic gives no number for this.
• Higher scores. Fable 5.2 beats Fable 5.1 in every row of Anthropic's table, and has the top score in 10 of 12. The charts are below.

And one thing did not change: Anthropic still says that for most coding and knowledge tasks you should는 with Claude Opus 5.5, and move up to Fable only when the work needs it. CNBC quotes Anthropic describing Fable 5.2 as a model for power users부터 집계됐어요.

Nothing in the announcement is about the Mac in particular, and MacRumors found nothing Mac-specific either. So the Mac section below is about how you reach the new model from a Mac, not about new Mac features.

Fable 5.2 vs Mythos 5.2, side by side

Anthropic's Mythos page puts it this way: results are for Fable 5.2 unless noted, and where Mythos 5.2 is shown separately, the gap is from tasks where Fable's safeguards step in. For nearly everyone reading a Mac app directory, Mythos 5.2 is not a choice you get to make. The rest of this page is about Fable 5.2.

Sources: the announcement, the Mythos page and the Mythos 5.2 docs.

Fable 5.2 vs the rest of the Claude family

Select a column heading to sort. Prices are for the API, in US dollars per million tokens. Context and longest answer are in tokens. Speed is Anthropic's own word for each model, not a measurement.

Source: Claude Platform Docs: Models overview, Claude Platform Docs: Pricing, Claude Platform Docs: Claude Fable 5.1. Read on November 11, 2026.

What I take from that table:

• The whole current family has the same window. Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 all read 1M tokens and write up to 128K. So the choice between them is about quality, speed and price, not about how much fits.
• Fable 5.2 costs 2 times what Opus 5.5 does, 4 times Sonnet 5.5, and 80 times Haiku 5.5 on a short prompt.
• Haiku 5.5 has two price lists. A prompt over 100,000 tokens pays $0.50 in and $2.50 out, 5 times the short-prompt price. The other models charge the same rate at any length.
• Sonnet 5.5 and Haiku 5.5 are not in Anthropic's benchmark table for this release. So there is no like-for-like score to chart for them here.
• Fable 5.1 is still available, marked legacy, at the old price. Anthropic's docs say it will not be retired before September 1, 2027.

Tool: which Claude model should I use?

Four questions, one pick, and the reason for it. It runs on this page and sends nothing anywhere.

The benchmarks, as charts

These are Anthropic's numbers, from the table in its announcement, run by Anthropic. Scores for the Claude models are with safeguards on, and every score is at each model's highest available effort unless the label says otherwise. The GPT-6 Astra and Gemini 4 Carbon scores are what Anthropic reports for OpenAI's and Google's models. I have not checked them against those companies' own pages.

The first bar in each chart is Fable 5.2. Longer is better in all of them.

Where Fable 5.2 is not first

Two rows in Anthropic's own table go to someone else:

• AutomationBench (business workflows): Gemini 4 Carbon scores 53.9%, Fable 5.2 51.4%.
• AA-Briefcase v1.1, Elo (knowledge work): Opus 5.5 scores 1800, Fable 5.2 1785. On the same test's rubric score, Fable 5.2 is ahead.

The biggest jumps

• Remote Labor Index: 41.3% at high effort, against 21.3% for Opus 5.5 and 17.9% for Fable 5.1. That is almost twice the Opus 5.5 score.
• RoboDojo Sim (run internally by Anthropic): 64.9%, against 39.4% for Opus 5.5. In points, this is the widest lead over the next best score in the table.
• Chartography (no tools): 73.4%, up from 44.8% for Fable 5.1. That is reading values off charts without tools.
• FrontierCode 2.0: 49.1% at xhigh effort, against 45.5% for Opus 5.5 and 32.4% for Fable 5.1.

The whole table, sortable

Select a column heading to sort by that model's score. The last column is how far Fable 5.2 is ahead of the next best score in the row, in points. A minus means it is behind.

Source: Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2. Read on November 11, 2026.

The rows where only Mythos 5.2 is shown

For cybersecurity and protein design, Anthropic shows Mythos 5.2 instead of Fable 5.2, because Fable's safeguards step in on those tasks.

Anthropic also lists GPT-6 Cyber at 61%, and notes that each lab's scores use different tasks. On De novo binder design (24 hr), Anthropic lists 85.1 for Mythos 5.2, 79.1 for Mythos 5.1 and 82.6 for Opus 5.5 with safeguards off.

What an independent tracker measured

Artificial Analysis tests models itself. For Fable 5.2 at maximum effort, read on November 11, 2026, it shows:

• Intelligence Index: 65, ranked 1 of 74. The median of comparable models is 40.
• Output speed: 101 tokens per second.
• Time to the first answer token: 233.95 seconds, against a median of 12.76. At maximum effort the model thinks for a long time before it answers.
• 45.2 minutes and $26.54 per task on its index, against a median of 12.6 minutes.

That is the slowest and most expensive way to run it. Medium effort, the default in the apps, is built to use fewer tokens, and Artificial Analysis had not published numbers for it when I read the page.

How to read all of this

• Most of these numbers come from the company selling the model. That is normal for a launch week, and it is why the independent numbers above are worth more per line.
• A benchmark is not your task. A score of 49.1% on FrontierCode 2.0 does not tell you whether it will fix your bug.
• Effort level changes everything. The same model at low and at max effort is two different products in speed and cost.
• I did not read the system card. It is a PDF of more than 10 MB that I could not open with my tools, so nothing on this page comes from it.

Prices, and what a task costs

The API prices, per million tokens, read on November 11, 2026 from Anthropic's pricing page:

What a task costs: four worked examples

These are sums, not measurements: the published price times a stated number of tokens. A real task also pays for thinking, which Anthropic bills as output tokens, so treat each as a floor.

A few things that move the bill:

• Thinking counts as output. Anthropic's docs say the billed output does not match the visible answer. Higher effort means more thinking.
• No surcharge for long prompts, except on Haiku 5.5. Anthropic's example: a 900,000-token request pays the same per-token rate as a 9,000-token one.
• Batch is half price if you can wait for the answer. It can't be combined with fast mode.
• Web search is $10 per 1,000 searches, on top of tokens.
• Price per token is not price per task. Anthropic reports that Fable 5.2 at high effort nearly matched Opus 5.5's top score on FrontierCode (45.3% against 45.5%) for roughly half the cost per task. It also reports Fable 5.2 matching Opus 5.5's top score on WANDR at medium effort for about 40% of the cost, and on Humanity's Last Exam at low effort for less than a tenth of the cost. Those are Anthropic's own measurements. If they hold for your work, the dearer model at a lower effort can be the cheaper one.

On a subscription

You don't pay per token on a Claude plan. What changes is whether Fable 5.2 is included, from Anthropic's plans page and its launch offer:

The announcement says Pro includes Fable, and the pricing page says Pro pays with usage credits. Both are right: the Pro inclusion is a launch offer that ends on November 16, 2026.

Tool: Claude cost calculator

Type a size, or paste the text you would send, and see the price of one run and of a month on each model.

Tool: how much fits in the context window

Every current Claude model, Fable 5.2 included, reads up to 1,000,000 tokens at once. Anthropic's rule of thumb is that 1M tokens is roughly 555,000 words or 2,500,000 characters. Models before Claude Opus 4.7 fit about 750,000 words in the same number of tokens, because the newer tokenizer makes about 30% more tokens from the same text.

Mac에서 실행 중인 What it means

The model runs on Anthropic's servers, not on your Mac. An older MacBook Air gets the same Fable 5.2 as a new Mac Studio. Your Mac only needs to run the apps.

The Claude Mac app

• What it needs: macOS 13 (Ventura) or later, per Anthropic's install guide. One download covers Apple silicon and Intel.
• How to get Fable 5.2: open the model picker and choose it. It is never the default, on any plan.
• Which plans: the app itself installs on every plan, Free included. Fable 5.2 needs a paid plan, as the table above shows.
• What else is in the app: a Quick Entry shortcut, desktop extensions that reach local files and apps, Claude Code, and Cowork on paid plans, per the download page.

Claude Code

• What it needs: macOS 13.0 or later and 4 GB of memory or more, on Intel or Apple silicon, and a paid plan or a Console account. The Free plan does not include it. This is from the setup page.
• 설치: curl -fsSL https://claude.ai/install.sh | bash, or brew install --cask claude-code.
• Switch to Fable 5.2: type /model fable, or는 with claude --model fable. The docs say it needs Claude Code 2.1.321 or later부터 집계됐어요.
• The default effort is medium, not the API's high. That is the setting where Anthropic says Fable 5.2 uses fewer tokens than 5.1.
• Fast mode: type /fast. On a subscription it is billed to usage credits, not counted against your plan's limits.

Agents that drive a Mac

Claude can click, type and look at the screen of your Mac through a feature called computer use, in Cowork and Claude Code inside the desktop app. It is not new with this release. What the pages say:

• Plans: Pro, Max and Ultra. Not Team or Enterprise.
• Permissions: macOS asks for Accessibility, so Claude can click and type, and Screen Recording, so it can see the screen.
• Two of Anthropic's own pages disagree on the details. The Claude Code docs call it a research preview and say it needs macOS 14 or later. The Cowork help article calls it a beta and says working in the background needs macOS 15 or later. Either way, it is not finished.
• What Fable 5.2 adds is a higher score. On OSWorld 2.1 (offline set, partial), Anthropic reports 89.3% for Fable 5.2, 87.2% for Opus 5.5 and 83.5% for Fable 5.1. Anthropic's table does not say which operating systems that test covers, so I would not read it as a macOS score.

If you let any agent use your Mac, give it the narrowest access that gets the job done, and watch the first runs.

Tool: what can I use on my Mac?

Pick your plan and your macOS version. It tells you which Claude apps run, and whether Fable 5.2 is included.

Mac apps on MacPip that work with Claude

A new model also means more people running Claude Code all day, and a small crowd of Mac apps has grown up around that: meters for your usage limits, dashboards in the notch, session browsers, tools that keep the Mac awake while an agent works.

This list is read live from MacPip's database each time the page loads. None of these listings says the app has been tested with Fable 5.2, and I have not tested them with it. Check an app's own page before you rely on it with a new model. A blue tick means a person at MacPip tried the app on a date. It is not a safety guarantee. The order is newest first, by the day a listing was checked or added, and never by payment.

Listings that mention Claude (22)

These apps name Claude or Claude Code in their own MacPip listing: usage meters for the menu bar, session browsers, notch dashboards for agents and more. The words are the makers' own.

• Cost DoctorMenu bar token tracker for Claude Code and Codex개발자 도구
• WhoIsCookingSee which AI coding agent is heating your Mac, then cool it downCoding Agents, by @andrwwilsn
• DuplexRun a second copy of Claude, SlackProductivity, by Aetrix Foundry
• LimitlyA free macOS menu bar app that shows how much ClaudeMenu Bar, by @amitdesignn
• CoralProductivity for Claude Code and Codex agentsProductivity, by @liorpesoa
• ToolnotchAI agent alerts, clipboard and windows in your Mac's notchNotch, by @hijooh_
• TokenBarUsage limits and weekly reset countdowns for Claude, Codex, Cursor, Gemini, CopilotDeveloper Tools, by Aura Technologies
• LancettaA native macOS menu-bar app that tells you when and how to use Codex and Claude CodeCoding Agents, by @gfazioli
• ElevenTake apart any app and ask your agent what it does.Security and Privacy, by @chiziaruhoma
• SylloSyllo is a desktop productivity app that adds a Dynamic Island and dynamic notchProductivity, by @rayftsx
• SwitchyMac controls, deploys, 2FA codes and Claude/Codex limits in your menu bar.Menu Bar, by @vladdesign
• CCAssistBrowse, search and resume every Claude Code, Codex and Cursor sessionCoding Agents, by @yashagl
• Trailback모든 것을 기억하는 생산성 앱Productivity, by @ItsPavanGK
• AutoRouterRoutes Codex and Claude Code tasks to the right model, carries context between agents, andCoding Agents, by @auto_router
• NotchNook ProThe original notch app for Mac, taken to the next level.Notch, by @getnotchnook
• Keep Mac AwakeKeeps your MacBook awake with the lid closed, so agents and long jobs keep goingMenu Bar, by @keepmacawake
• PulseFloating usage rings Claude Code, Codex, Antigravity, and Grok.Coding Agents, by @tommyent
• Caddie작업, 메모, 캘린더, AI 에이전트를 하나의 Mac 앱에Productivity, by Harsha Garlapati
• UsageBeaconSee your remaining Cursor, Claude, and Codex usage budgets right in the Mac menu bar.Menu Bar, by David Cohen
• PodiumPrompt your software factory alongside your ideaCoding Agents, by @podium_ade
• QuotchiTrack Claude, Codex and Gemini limits from your menu barDeveloper Tools, by @thomaskanze
• Vexilo CodeSee what each Claude Code skill and agent costs per turnCoding Agents, by @lilhawk0312

More coding agent apps (3)

Listed in 코딩 에이전트, without naming Claude in the listing.

• GitKrakenMeet GitKraken, the creator of legendary Git tools for developers and teamsDeveloper Tools, by @gitkraken
• MacSpacesMacSpaces turns your notch into a workstation.노치
• Raydo Agent OSThe AI agent operating system for real workAI, by @raydoai

More AI apps (12)

Listed in AI, without naming Claude in the listing.

• CoverHide your identity from Ai in documents.Security and Privacy, by @cover_redact
• ScreenVerbThe screenshot app built around AI. Capture, mark up, and paste AI answers in three keys.Productivity, by @davidtranjs
• Interview RooNative Mac workspace for coding interview preparationEducation, by Qelumo LLC
• ActionClipTurn selected text into actions and create your own Mac workflowsClipboard and Text, by @AdvaitVaidya2
• MacVisorApple silicon Mac용 가상화 플랫폼Productivity, by @scaleninja
• Phosphene AIFree, open-source AI video with sound, images and songs on Apple Silicon Macs.AI, by @PhospheneAI
• yovoiceOpen-source, local-first AI voice cloning, speech, and music creation studio for MacAI, by @leemysw
• OrkasTurn goals and project files into reports, slides, videos and code with a desktop AI team.AI, by Orkas
• Lectern AI슬라이드를 강사용 발표 원고, 실습, 퀴즈, 유인물로 바꿔요AI, by Ian Trimble
• CortexCortex is The Activity Monitor for the AI era.AI, by @eduard_pantazi
• KepoKepo turns any website, RSS feedAI, by Kepo
• EspressoEspresso is the Everything-style file search app for MacFiles and Storage, by @gianni135mc

For Mac app makers

If you build Mac apps, here is where I would and would not spend Fable 5.2.

• Keep Opus 5.5 for the daily work. That is Anthropic's own advice, and it costs half as much.
• Bring in Fable 5.2 for the long jobs. A migration across a whole codebase, a bug that has survived three attempts, a review of everything before a release. Those are my examples, not Anthropic's, but they are the kind of long, hard task its numbers are about.
• Lower the effort before you lower the model. Anthropic's cost claims above are all about Fable 5.2 at low, medium or high effort, not max.

If your Mac app calls the Claude API itself, read the what's new page before you change the model id. A few things return an error on Fable 5.2 that worked on older models:

• Thinking is always on. Sending thinking: disabled or a thinking budget returns a 400 error.
• No prefilled answers, no temperature. Prefilling the assistant message returns a 400, and the sampling settings should be left out.
• No forced tool choice. tool_choice set to any or to a named tool returns a 400.
• A refusal is a normal answer. A blocked request comes back with HTTP 200 and stop_reason: "refusal", so handle it and decide on a fallback.
• 30-day data retention is required. An organization set to zero retention gets an error unless Anthropic has cleared it.

And the MacPip side, all free:

• Prompts for Mac makers: copy-ready prompts for writing your listing, checking what your app needs, release notes and launch day. They work in any Claude model.
• MacPip for AI agents: the public MCP server and the Markdown copies of every page, so your agent can read the directory.
• Submit with your AI agent: one private link, and you can say "list my app on MacPip" to Claude. A person still checks every app.
• 무료 도구: the launch checklist, the notarization explainer and the rest.
• Mac 앱 제출 방법, if you have not launched yet. Listing on MacPip is free.

Should you switch?

My honest read, from the published numbers only:

Switch now if you are on Fable 5.1. It is 20% cheaper, faster, blocks less, and scores higher in every row. The only work is on the API, where a few requests that used to pass now return an error.

Try it if your work is long and hard. Multi-hour agent runs, large refactors, research that needs many steps. That is the kind of work the Remote Labor Index tries to measure, and there Anthropic reports 41.3% for Fable 5.2 against 21.3% for Opus 5.5.

Stay on Opus 5.5 if your work is everyday coding and writing. Anthropic says so itself. Opus 5.5 is half the price, and on a subscription Fable 5.2 can use only up to half of your weekly limit.

Stay on Sonnet 5.5 or Haiku 5.5 if speed or cost is the point. Nothing in this release changes them.

Wait if any of these apply:

• You are on the Free plan. Fable 5.2 is not there.
• Your company needs zero data retention. Fable 5.2 requires 30 days unless Anthropic has cleared you.
• Your code depends on prefilled answers, forced tool choice or temperature.
• You want independent numbers first. Right now there is one tracker's result at one effort level, and the rest is Anthropic's.

What I can't tell you: how it does on your code. The cheapest honest test is to take one task Opus 5.5 struggled with last week and run it again on Fable 5.2 at medium effort.

질문

How this page was made

• Primary sources first. Anthropic's announcement, its model and pricing docs, its help center and the Claude Code docs. The benchmark, price and model tables were read from the page source, not from a summary.
• Read on November 11, 2026. Prices and plan details change. The date is on every tool.
• Nothing is estimated. Where a number was not published, the page says so. The worked examples are sums on published prices, and say so.
• Where sources disagree, both are shown. Two of them are on this page: the computer use details, and Anthropic's announcement saying "Microsoft Azure" where its docs say Microsoft Foundry.
• What I could not check: the system card, the competitor scores against their makers' own pages, any score for Sonnet 5.5 or Haiku 5.5 on these tests, and whether any listed Mac app has been tested with Fable 5.2.
• One data file. Every number on this page and in the four tools comes from a single file in MacPip's code, so a price change is one edit.

MacPip is an independent Mac app directory, and is not affiliated with Anthropic. Claude, Claude Fable, Claude Mythos, Claude Opus, Claude Sonnet, Claude Haiku and Claude Code are Anthropic's product names. If something here is wrong, tell me and I will fix it and note the change.