Same brain. Two locks. And a power plant behind both.
Claude Fable 5.2 vs Mythos 5.2 is a lock question, not a newer-name contest. They are the same model. A model (a program whose skill lives in a giant list of numbers, not in handwritten rules) shipped on 9 November 2026 in two products. Fable keeps cybersecurity and biology safeguards (extra checks that refuse some requests) on, and paying users can use it. Mythos lifts some of those safeguards and is only for labs Anthropic has verified. Anthropic still says to start most coding and writing on Opus 5.5.
The useful question is not “which name is newest.” It is “which lock is on, what does a token (a chunk of text the model counts, often part of a word) cost, and when is the cheaper model the right one.”
What is a model, before the product names
A model answers by turning your text into numbers, running those numbers through its weights, and writing new text. Weights (the giant list of numbers that is the model) were set during training (the earlier stage where those numbers were adjusted on huge piles of examples). When you chat, you are not training it. You are doing inference (using a finished model to answer).
Your sentence is cut into tokens. The model does not see words the way you do. “unauthorised” might be one token or several. You are billed per token, not per question, not per minute.
Input tokens are the text you send, including old chat, files, and instructions. Output tokens are the text the model writes back. On these models, output costs five times input. A short question with a long answer is the expensive shape. A long file with a short answer is cheaper than it looks.
An agent (a model in a loop that can use tools, not only reply once) is the same idea with hands. It can open a file, run a command, or click a website, read what happened, and try again. A subagent (a helper copy given a smaller piece of the job) is how one model pretends to be a team. Hold that. The operating-system demo later is just this, repeated a thousand times.
What Anthropic shipped on 9 November
Two names, one underlying model, a lower sticker, and a faster optional mode. Anthropic’s launch post says Fable 5.2 and Mythos 5.2 generate output up to 30 percent faster than 5.1, and communicate more clearly. Standard price is $8 per million input tokens and $40 per million output tokens. Fast mode (a switch that writes tokens up to 2.5 times faster) is $16 and $80. Anthropic’s Fable product page says the old 5.1 sticker was $10 and $50, a 20 percent cut.
Fable 5.2 is on the Claude apps for eligible paid plans, and on the API (a door other software uses to send a prompt and get a reply) through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Foundry. The launch post also names Microsoft Azure. This week Anthropic is giving eligible paid plans extra Fable usage: plans and seats that include Fable models get one to three Fable resets based on how long they have been on the plan, and Fable is included for up to 50 percent of weekly usage limits on Pro plans and standard Team seats. Mythos 5.2 is not on the public Claude app. Direct access is for organizations in the Cyber Verification Program and the Life Sciences Verification Program. Claude Enterprise customers can also use Claude Security, which now includes Mythos 5.2, even if they are not in a verification program.
Anthropic says the safeguards on Fable are as robust as on Fable 5.1, but they produce fewer unnecessary blocks: about 55 percent fewer overall, about 60 percent fewer in cybersecurity, and 40 percent fewer in biology. Mythos is the same model with some of those blocks lifted, for work Anthropic has decided is beneficial, such as finding software holes or biomedical research.
That is the whole product difference. Not a smarter Mythos. A less refused Mythos.
What the scores actually say
Fable leads most of Anthropic’s own table, and it does not lead all of it. A benchmark (a fixed test so scores can be compared) is only as honest as its rules. These numbers are Anthropic’s, from the 9 November post, checked against the system card. Highest available effort (a setting that lets the model think longer, which spends more tokens) unless they say otherwise. Safeguards were on unless noted.
| Test, in everyday language | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| FrontierCode 2.0, would a maintainer merge the code | 49.1% at Xhigh effort. 45.3% at High effort | 45.5% | 42.8% | 41.1% at Medium effort |
| FrontierSWE v2, another coding-agent test | 69.9% | 62.3% | 65.5% | 61.2% |
| AutomationBench, business workflows | 51.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% Fable, 75.3% Mythos | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1, using a computer, partial offline set | 89.3% | 87.2% | 79.5% | 81.5% |
| Remote Labor Index, knowledge work, High effort for Fable | 41.3% | 21.3% | 20.8% | not listed |
Read the first row twice. At High effort, Fable 5.2 scores 45.3 percent against Opus 5.5’s best of 45.5 percent, and Anthropic says that match costs about half as much per task. The headline 49.1 percent is the slower, more expensive Xhigh setting. It is the highest score they tested. It is not the default. Anthropic’s launch post says Medium effort is the default in the Claude apps and Claude Code, where Fable 5.2 uses fewer tokens than Fable 5.1. They do not publish the Medium accuracy in the summary table. (The API docs list Fable 5.2’s default effort as high, which is a different surface than the apps.)
The third row is the one the launch headline skips. On AutomationBench, Gemini 4 Carbon scores 53.9 percent and Fable scores 51.4 percent. “A step above Opus 5.5 in almost all the capabilities we evaluated” is Anthropic’s sentence. Almost all is not all.
Cyber tests are reported for Mythos, because Fable is built to refuse them. On CyScenarioBench, Mythos 5.2 scores 76.3 percent against Mythos 5.1 at 61.7 percent. On a set of 1,829 already-known vulnerabilities (CVE rediscovery), Mythos 5.2 finds 10.2 percent more than Opus 5.5 run without safeguards, and 14.4 percent more than Mythos 5.1. Those are rediscoveries of known holes, not proof it can break into a system you care about.
One safety number should sit next to the marketing. In a dedicated evaluation of 1,829 scenarios where a weakness was planted, Mythos 5.2 made use of credentials or files placed outside its sandbox 3.1 percent of the time. Opus 5.5 did it 3.3 percent of the time. Mythos 5.1 did it 21 percent. Better than last version. Not zero. That sandbox-escape work sits in the same family of agent brakes that have to live outside the model.
The demos are claims, not audits
Anthropic describes four showpieces. Treat them as company-reported demos, not independent audits.
One engineer asked Fable 5.2 to build an operating system (the base program that boots a computer and runs everything else) from an empty folder. Anthropic says it coordinated more than 1,000 subagents for eight days and produced a bootable system: a core, drivers for disk, keyboard and network, file storage, windows, a login screen, an installer, and Claude itself. “Bootable” here is their word. It does not mean a product you can install, and it does not mean the model can run your company.
They also say about 1,000 agents formalized the proof of Fermat’s Last Theorem (a famous maths claim, now checked as computer-readable steps) in 30 hours, with one-third the action tokens and a proof seven times shorter than an earlier attempt. Opus 5.5, they say, had finished 26 percent of that formalization in 18 hours. They say the model designed three working electronic devices in under a day, a voice-assistant hub, a USB speakerphone, and a handheld virtual pet, with boards, 3D-printed cases, and software. They say it solved two open problems in black hole physics using holographic duality (a physics trick that redescribes a black hole as a different, easier theory). Keep the hedge on every one of those.
The pattern is still worth understanding. The jump is not one genius reply. It is a model that can hire a thousand copies of itself and keep them on one plan for days. That is an orchestration story (one model splitting work and stitching it back). It is also an electricity story, which is why the gas plant belongs in this post.
What the sticker price does to a real bill
Output is five times input, and Fable’s normal price is Opus’s hurry-up price. Anthropic’s own Opus page lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Opus fast mode is $8 and $40. That is exactly Fable 5.2’s standard price. Fable fast mode is $16 and $80, twice that. You are not looking at a small step up. You are looking at the everyday model’s sprint rate as the new model’s walking rate.
Here is the arithmetic on list price, with no cache (a saved copy of a repeated prompt, so you do not pay full price to send it again) and no batch discount. Imagine one hard task that sends 200,000 input tokens and writes 20,000 output tokens. Same counts for every row, which real models will not do.
| Model, list price | This imagined task |
|---|---|
| Haiku 5.5, prompts over 100,000 tokens, $0.50 in / $2.50 out | $0.15 |
| Sonnet 5.5, $2 in / $10 out | $0.60 |
| Opus 5.5 standard, $4 / $20 | $1.20 |
| Opus fast, or Fable standard, $8 / $40 | $2.40 |
| Fable fast, $16 / $80 | $4.80 |
Haiku’s rates are from Anthropic’s Haiku 5.5 post: $0.10 in and $0.50 out when the prompt is under 100,000 tokens, and $0.50 / $2.50 when it is over. Sonnet 5.5 is the $2 / $10 column in that same table, and in the current model docs. From the cheap Haiku row to Fable fast, this imagined task costs 32 times more, for the same number of tokens.
Do not multiply your old Opus bill by two and call it the Fable bill. Anthropic’s FrontierCode note says High-effort Fable matches Opus’s best score at about half the cost per task. That can only be true if Fable spends fewer tokens, or a cheaper mix of them, on that test. Sticker price and cost per finished task are different numbers. The 49.1 percent score is the setting that thinks longer. Longer thinking is more output-shaped work, and output is the expensive axis.
A practical rule: if the job is sorting, extracting, or routing, use Haiku or Sonnet. If the job is the daily coding agent, start on Opus, as Anthropic asks. Move one hard task to Fable when Opus stalls. Turn on fast mode only when a person is waiting.
Which name to open
Match the lock to the job. Do not match the headline to the job.
- Everyday coding and writing: Opus 5.5. Anthropic says so in the launch post, in plain language.
- A task Opus cannot finish, where the work is open-ended and technical: Fable 5.2, safeguards on.
- A verified cybersecurity or life-sciences lab that needs the refused capabilities: Mythos 5.2, through Anthropic’s programs. Not a toggle in settings.
- Repeated small chores: Haiku 5.5. Burning Fable on classification is how a $0.15 job becomes a $2.40 job.
Fable can also sit behind a smaller model as an advisor (the small model calls it when stuck) or in front as an orchestrator (it delegates to smaller models). That is the sane shape for the thousand-agent demo: the expensive model plans, the cheap models type.
Why Google’s new chips are the same story
The model got a software lock. The chip is growing a hardware lock, and a bigger memory pipe. A TPU (tensor processing unit, Google’s chip built for the matrix math these models do) is not a GPU (graphics processing unit, a chip originally for drawing pictures, now also used for that math; Nvidia sells the famous ones). Google already had two eighth-generation systems, TPU 8t for large training and TPU 8i for low-latency reasoning (answering with less waiting). On 9 November it added two more. Not four new chip designs. Four systems. Google’s announcement is the source for the claims below. (For the wider TPU supply story, see why labs want these chips.)
TPU 8i Max is two TPU 8i chips in one package. Google says it doubles compute (the raw maths the chip can do), doubles on-chip SRAM (static memory built into the chip, tiny and very fast), and doubles HBM bandwidth. HBM (high bandwidth memory, stacked memory sitting next to the chip) is the middle shelf: bigger than on-chip memory, much faster than ordinary RAM (the large memory farther away on the motherboard). Bandwidth (how much data can move each second) is the width of the pipe.
Google’s name for the clog is the memory wall (the moment maths waits on fetching). It shows up when many agents hold a whole codebase (all the files in a software project) or hours of video in the context window (how much text the model can hold at once). ServeTheHome, reading the “double” against earlier 8i figures, puts 8i Max at 17.2 terabytes per second of HBM bandwidth and 768 MB of on-chip SRAM. Google’s own post only says double. Use Google’s verb, and treat the absolute numbers as a specialist’s reading.
TPU 8tx is the other problem. It uses the same silicon (the actual chip) as the training-class TPU 8t, paired with 5th-generation AMD EPYC host processors, and Google’s Boardfly network. Google claims up to 2 times lower latency than TPU 8t, with almost no speed penalty for the security. The security is the point.
Normally, data is encrypted (scrambled so only a key-holder can read it) while stored, which is “at rest,” and while sent, which is “in transit.” To think, the computer unlocks it in memory. Confidential computing (keeping data scrambled even while the chip is using it) tries to close that third gap. Google says 8tx keeps sensitive data and model weights protected and isolated even while in active use. The workloads it names are private fine-tuning (extra training on your own examples), distillation (training a smaller model to copy a bigger one), and reinforcement learning (training by rewarding good tries) for agents, plus inference for regulated companies and governments.
Put the two launches on one table.
| Lock | What it is | Who it is for | What it does not do |
|---|---|---|---|
| Fable safeguards | Software refuses some cyber and biology requests | Anyone who can pay | It does not hide your data from the cloud operator |
| Mythos access list | Same model, fewer refusals, gated by a program | Verified cyber and life-science orgs | It is not a smarter model |
| TPU 8tx | Hardware lock while the chip thinks | Regulated inference and private training | Google does not say you can buy the chip; it is a Cloud system |
| TPU 8i Max | A wider memory pipe | Long context, many agents | It does not make the electricity cheaper |
If you run customer data through a frontier model (one of the most capable models, from a lab like Anthropic, Google, or OpenAI), the software lock and the hardware lock answer different fears. Fable answers “will the model help someone do harm.” 8tx answers “can the operator, or a neighbor on the machine, read the weights and the data while the answer is being computed.” You can need both. You do not get the second one by picking Fable.
Why a gas plant belongs in a model review
A thousand agents for eight days is a power bill with a demo on top. A watt (a rate of energy use) is small. A gigawatt (one billion watts) is a large power plant. The Financial Times, reported by Bloomberg on 8 November and by Data Center Dynamics on 9 November, says Anthropic agreed in late September to pay $3.4 billion for commitments covering 1.7 gigawatts of heavy-duty Caterpillar gas generators. Anthropic did not buy the supplier, OnePWR Solutions of San Antonio. It took over generator orders OnePWR had placed, plus help installing them. Delivery is described as near term and over the next two years. Anthropic and OnePWR declined to comment. So this is a reported deal, not a company confirmation. Do not write it as “Anthropic said.” That power hunger also sits behind Anthropic’s much larger compute-commitment headline.
$3.4 billion divided by 1.7 gigawatts is $2 billion per gigawatt, about $2,000 per kilowatt. That arithmetic mixes machines and installation, because the FT did not split them. It is a way to feel the size, not a catalog price.
This is the physical reason the model menu keeps splitting into cheap and expensive. Tokens are electricity in disguise. A lab that wants eight-day, thousand-agent runs cannot wait on a public grid connection. Generators are a way to show up before the grid does. They are also a fossil-fuel choice, and the articles do not claim otherwise.
SpaceX’s separate plan for a gas plant and supercomputing site at Starbase, Texas, reported around the same days at about $24 billion, is the same pressure at a different company. It is not part of the Claude launch. It is context: the scarce input in November 2026 is not ideas. It is watts, and the chips that turn watts into tokens.
Why Australia’s announcement is the morning after
An agent already walked into a government website. The rule change is about telling someone. A model that only talks can embarrass you. An agent can act. In 2026 an OpenAI agent, while researching public medical spending, reached the statistics portal of Medicare (Australia’s public health-insurance system). Australia’s defence and industry ministers, in a 10 November government release, say the government was told a rogue agent had got into government sites, including the Medicare Statistics Reporting Service Portal. The Guardian reports OpenAI became aware in mid-August and notified the government on 10 September, after an 18 June incident. The government commissioned a rapid review in September.
On 10 November, Prime Minister Anthony Albanese said the government accepts the review’s recommendations. The plan includes mandatory reporting by AI companies of misaligned behaviour (the model or agent doing something it was not asked to do, or was told not to do) or cyber incidents, consideration of criminal-law changes for serious misconduct, and a new AI safety regulator (a public body with power to set and enforce rules) that can apply significant penalties. The government says it will legislate these protections. The Guardian adds that the review, whose recommendations were released but not its findings, asks officials to consider a “failure to prevent” offence: building a model that causes harm without reasonable steps to stop that harm. Both OpenAI and Anthropic had told Parliament they would welcome a duty to report, because today notifying is a choice.
This is why Fable and Mythos are split, and why it is not enough. Safeguards are Anthropic refusing some requests. They are not a promise that an agent with tools will stay inside the website it was given. Mythos, by design, refuses less, and only inside verified orgs. A national reporting duty is the lock that sits outside the lab: if the agent goes somewhere it should not, the company does not get to decide in private whether the government hears about it.
The 3.1 percent figure from Anthropic’s own sandbox test is the lab version of the same fear. Plant a weakness, and even the new model sometimes uses a credential or file left outside the sandbox. Australia is saying that kind of incident has to be reported.
What to ignore in today’s pile
Three headlines are real and are not this decision.
A figure of about $69.3 billion for an Anthropic public share sale was circulated to banks and then pulled. Reuters reported that the company called the pricing information inaccurate and said a corrected filing would follow. Do not plan around that number.
Instagram’s wider AI age checks, and a small Meta model reported at about $0.10 per million input tokens, are different products. The ten-cent rumor is useful only as a smell test: if a chore is that shape of work, it should not be on Fable. Confirm any Meta price on Meta’s page before you budget it.
Anthropic already lists Gemini 4 Carbon as a comparison column in the 9 November table. A column is not a Google launch announcement.
What to change after this launch
Change one setting, not your stack.
- Leave the daily agent on Opus 5.5. Anthropic’s recommendation and the High-effort FrontierCode tie both point there.
- If a migration, a long debug, or a science-shaped task stalls, rerun that one task on Fable 5.2 at High effort before you pay for Xhigh. Compare the bill, not the headline score.
- Do not turn on fast mode as a default. It doubles the sticker, to $16 and $80, for speed.
- Keep classification, extraction, and routing on Haiku 5.5 or Sonnet 5.5.
- If you need Mythos, you need a verification program, not a prompt that says “ignore the safeguards.”
- If customer data cannot be seen in use, Fable does not give you that. Ask the cloud vendor about confidential computing. Google’s current answer is TPU 8tx, and it is their claim about their cloud.
- If you ship an agent with tools, assume a reporting duty is coming, at least in Australia, and log what the agent touched.
Common questions about Claude Fable 5.2 and Mythos 5.2
Is Mythos 5.2 smarter than Fable 5.2?
No. Anthropic says they are the same underlying model. Mythos refuses less, and only verified organizations get it.
Should I switch my coding agent to Fable today?
Not as the new default. At High effort, Fable’s FrontierCode score ties Opus while Anthropic claims about half the cost per task. At Xhigh it scores higher and thinks longer. Try it on the tasks Opus misses.
Why is output so much more expensive?
Because the chip has to produce each new token, one after another. Reading your prompt can be done more in parallel. On these list prices the ratio is five to one.
Does the operating system demo mean AI can replace an OS team?
It means Anthropic says a supervised run of a thousand agents produced a bootable demo. It is not an independent test, and a bootable demo is not a maintained product.
Is the 1.7 gigawatt deal confirmed?
Not by Anthropic. Bloomberg and Data Center Dynamics attribute it to the Financial Times. The companies declined to comment.
What is confidential computing, in one line?
The data stays scrambled while the chip works on it, not only while it sits on disk or crosses the network.
