Claude Fable 5.2 and Mythos 5.2: does thinking longer make it better?One model, two doors. From a 1,000-agent research team to the max-effort trap, an interactive guide to Anthropic’s new frontier models
A team of 1,000 AI agents translated an existing proof of Fermat’s Last Theorem into a form a computer could check, in about 30 hours. The same announcement shows an operating system built from an empty folder that boots on a real laptop, and three electronic devices that Claude designed from a one-paragraph prompt and human engineers assembled. Anthropic’s November 9, 2026 introduction to Claude Fable 5.2 and Mythos 5.2 begins well beyond the chat window.[1]
Then comes a surprising recommendation: for most coding and knowledge work, start with Opus 5.5. The company has released a more capable model without telling everyone to replace their daily driver. Part of the explanation sits beneath the scorecard, on the invoice. Fable’s token prices are 20% lower than its predecessor’s, but long reasoning can make a task much more expensive. On some evaluations, max effort even scores below xhigh, the setting one step down. In one coding evaluation, xhigh scored 49.1% and max 45.7%, at more than three times the cost.[1][2][3]
This is a comparison of published sources, not a hands-on model review. I checked the announcement, the 264-page system card and independent evaluations, keeping vendor demonstrations separate from external tests. Numbers reflect the sources checked on November 9–10. The interactive tools do not call the model. They let you explore the reported results and their trade-offs.
- Fable 5.2 and Mythos 5.2 share the same underlying model. Fable is the general-access configuration; Mythos permits a wider set of certain cyber and biology tasks after verification. Both cost $8 per million input tokens and $40 per million output tokens.
- The strength is difficult, extended technical work. Fable leads the AA Intelligence Index at 65 and Epoch’s ECI at 174. The ranking still depends on the test, tools and effort setting.
- Do not start by turning everything to max. On FrontierCode 2.0, xhigh scored better than max and cost less. Start with Opus 5.5 for ordinary work and escalate the parts that need more capability.[1][4][5][6][3]
01One book, two doors
“Fable” and “Mythos” sound like different rungs on a model ladder. Anthropic instead describes them as the same underlying model, with the same basic specifications and pricing. The difference is safeguards and access. Fable restricts certain cybersecurity and biological requests; Mythos allows verified users to do a wider range of work in those areas.[1]
Imagine two copies of the same book in a library. One is in the public reading room. The other is in a collection that checks your identity and research purpose before letting you in. Entering the second room does not remove every rule: verification programs still impose prohibited-use policies and additional security requirements.
One model, two doors
TRY ITFable 5.2
The general-work door. The publicly available model for demanding coding, mathematics and long-running agents, with additional safeguards for sensitive cyber and biology requests.
Mythos 5.2
The verified-specialist door. The same underlying model, with broader access for approved cybersecurity and life-sciences work. Verification is not a license to remove every safeguard.
Same specifications and price; different access conditions. Anthropic announcement and model documentation, Nov 9, 2026.
Mythos has a separate door for each field: the Cyber Verification Program (CVP) for security work and the Life Sciences Verification Program (LSVP) for biology. Individuals can apply only for CVP’s defensive tier, and it is not a switch you unlock by writing “I am a security researcher” in a prompt. The detailed conditions are folded below for readers considering an application.[7][8]
Considering Mythos access? CVP and LSVP conditions in detail
CVP has three tiers: Defense Access, Red Team Access and Specialized Access. Individuals can apply for Defense Access, while higher tiers require additional organizational and task-specific review. Security controls apply too. Today Defense Access requires multi-factor authentication and API keys replaced at least every seven days. Starting December 15, 2026, it also requires phishing-resistant MFA and stops allowing static or long-lived API keys. The Console setup steps go further: after that date, a Defense Access workspace does not allow API keys at all.[7]
The Life Sciences Verification Program is separate. Standard Use grants apply to a team and renew annually; High-risk Use grants cover a specific project and renew every six months. Eligible routes include Team, Enterprise and the Claude API, not personal Pro, Max or Ultra subscriptions. Fable models themselves are not part of LSVP. A biology researcher needs to check the relevant Mythos access conditions.[8]
There is also one route to Mythos’s capability that needs no verification. Claude Enterprise customers can use Claude Security, which scans codebases for vulnerabilities and now uses Mythos 5.2, even if they are not CVP members. That is access through a fixed security product, not permission to prompt Mythos freely.[1]
A restricted Fable request need not end with silence. In the apps, certain cyber requests fall back to Opus 4.8 and biological requests to Opus 5. But server-side fallback is opt-in on the API and is not implemented identically on every platform. On Bedrock, retrying is your code’s responsibility. Logging which model actually answered becomes part of understanding the result.[1][9][10]
Anthropic also reports fewer unnecessary blocks. In early testing across claude.ai, Claude Code and Cowork, it observed about 55% fewer blocks with Fable 5.2 than with Fable 5.1: about 60% fewer in cybersecurity and 40% fewer in biology. These are the company’s own early figures.[1]
Both models support a one-million-token context window and up to 128,000 output tokens, with a June 2026 knowledge cutoff. Thinking is always enabled; effort has five levels: low, medium, high, xhigh and max. One easily missed distinction: Claude apps and Claude Code default to medium, while the API defaults to high. Selecting the same model name does not guarantee the same starting bill.[11][1][2]
02Prices fell 20%. Why can the bill grow?
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from Fable 5.1’s $10 and $50. Mythos has the same rates. Opus 5.5 costs $4 and $20, so Fable’s fresh-input and output rates are twice as high. Thinking tokens are billed as output. A short final answer does not necessarily mean a small invoice.[4][2]
Cache reads are different. Fable 5.2 and Opus 5.5 both charge $0.20 per million cached input tokens. For Fable, that is 97.5% below the base input rate. Writing the cache costs more: $10 per million tokens for five-minute storage or $16 for an hour. The saving depends on reusing the matching input while the cache remains valid.[4]
Usage matters: what would your job cost?
USDUnits are thousands of tokens (K), in US dollars. This compares equal token usage.
Small numbers show the per-million-token rates for fresh input / cache reads / output. Excludes cache writes, tools, tax and regional premiums. Batch is asynchronous and cannot combine with Fast. Fast cache reads: $16 × 2.5% = $0.40; Batch: $8 × 50% × 2.5% = $0.10. Real jobs use different token counts on different models.
Pricing and stacking rulesSpeed has two clocks. Anthropic reports output up to 30% faster than 5.1, with Fast mode up to 2.5 times faster than standard mode. Neither claim guarantees the same reduction in end-to-end task time. In AA’s independent max-effort measurements on November 9 and 10, the model processed about 100 tokens per second, counted over the whole request and including reasoning tokens. The first answer token arrived after roughly three to four minutes. Token throughput and the moment you start seeing an answer are different metrics.[1][5]
For a worked example, send the same 100,000-token source material ten times, generating 10,000 output tokens each time. Without caching, input costs $8 and output $4: $12 total. With one five-minute cache write and nine cache hits, the bill becomes $1 for the write, $0.18 for reads and $4 for output: $5.18. This is arithmetic, not a measured Anthropic workload. It excludes extra tool charges and changes in reasoning usage. Re-reading the material becomes cheaper; thinking longer does not become free.
Artificial Analysis measured a much larger difference on actual evaluation tasks. Fable 5.2 at max cost $26.54 per Intelligence Index task, versus about $10.04 for Opus 5.5 and $12.99 for Fable 5.1, both also at max. The new Fable was roughly twice as expensive as its predecessor despite lower unit prices. Its average output grew from about 118,000 to 274,000 tokens per task, one contributor to the increase. The breakdown matters too: roughly $15.65 for input and cache plus $10.89 for output on Fable, compared with $6.57 plus $3.47 on Opus. Output alone does not explain the whole gap. A task can span multiple calls, so that total does not contradict the 128,000-token limit per response.[5]
Batch halves the standard input and output rates for work that can wait. Fast mode goes the other way, at $16 and $80, and cannot be combined with Batch. The full one-million-token context has no separate long-context surcharge. Regional processing prices depend on the provider: Anthropic’s 1.1× US-processing rule should not be copied wholesale onto Bedrock or Google-managed Vertex.[4] The broader lesson is the same as in the Opus 5.5 comparison: put your workload into the calculation, not just the model’s sticker price.
03Number one, but number one at what?
The improvement is visible outside Anthropic’s own tests. Fable 5.2, listed by AA as Max with Default Fallback, leads the Artificial Analysis Intelligence Index v5.0.1 with 65. At the checked snapshot, Carbon at high scored 60, Opus 5.5 at max 59, GPT-6 Astra at max 57 and Fable 5.1 at max 49. Before rounding, the improvement over Fable 5.1 is about 15.4 points. The 59 in my earlier Carbon article belongs to the index snapshot used then. Scores from different dates and versions should not be silently mixed into one permanent table.[5]
Epoch’s Capabilities Index also places Fable first, at 174. Epoch cautions that this initial result uses its own pre-launch testing and may change. Several evaluations pointing in the same direction strengthen the case for a real advance. They do not turn a single number into a universal measure of every job.[6]
The scorecard: who measured what?
EVIDENCEAA Intelligence Index v5.0.1
First on the composite index, and the most expensive per task
Reveal results and conditions
Fable max costs $26.54 per task versus $10.04 for Opus max. Rank is not cost efficiency. Carbon uses high; Astra and Opus use max. AA lists the Fable configuration as Max, Default Fallback.
Read the sourceCursorBench 4.0
Max scores highest, at a much higher cost
Reveal results and conditions
Cost per task: $32.79 / $13.43 / $9.70. Cursor credits prompt and harness improvements too.
Read the sourceARC-AGI-3
Same model, different score depending on the harness
Reveal results and conditions
These scores do not form a same-condition ranking. Standard at max scores 65.7%. Verified by ARC Prize.
Read the sourceEpoch ECI
Also first on an independent index (provisional)
Reveal results and conditions
Provisional, based on Epoch’s own pre-release tests. Rounded equal scores need not imply identical ranks.
Read the sourceZapier AutomationBench
On workflow automation, Carbon stays ahead
Reveal results and conditions
From Zapier’s own report. Some older-model figures differ from Anthropic’s launch table, so they are not mixed here.
Read the sourceRemote Labor Index
First on remote-project automation, about twice the runner-up
Reveal results and conditions
Scale AI’s project-automation rate. It is not a claim that this fraction of all jobs can be replaced.
Read the sourceProgramBench
Showing remaining context changed the result
Reveal results and conditions
Anthropic measurement. Fable trails under the base setup; adding a context-remaining indicator lifts it to 91.6%. These are different conditions.
Read the sourceOSWorld 2.1
Completion and partial credit are different numbers
Reveal results and conditions
Anthropic measurement, 82 offline tasks. The 89.3% is partial credit, not a full-completion rate.
Read the sourceBenchCAD 2.1
A large jump on design tasks with tools
Reveal results and conditions
Anthropic’s with-tools condition. Fable 5.2 without tools scores 75.4%.
Read the sourceChecked November 9, 2026. Each card uses its own scale. Do not compare bar lengths or scores across tests. Scores are rounded to the shown precision.
Three questions belong next to every score: Who measured it? Which tools were attached? How much reasoning was allowed? A manufacturer’s announcement can include independent results. An external evaluator can still use different model-specific harnesses. The logo above the table does not answer every methodology question.
On CursorBench 4.0, Fable max leads at 60.0%, ahead of Opus max at 57.8%. Fable high scores 56.9% at $9.70 per task, while max costs $32.79. That is about 3.4 times the cost for another 3.1 percentage points. The system card also notes that Cursor credits part of the result to prompt and harness improvements. Not all of the gain should automatically be attributed to the model alone.[12][3]

ARC-AGI-3 makes the harness effect particularly clear. ARC Prize verified a best Standard-harness score of 81.2% at xhigh, and 97.7% at high with the Provider Adapter. Even holding effort at high, the scores are 79.3% and 97.7%, respectively. Writing only “Fable scored 97.7%” strips out an important part of what produced that result. Its 93.8% on ARC-AGI-2 is strong too, but below GPT-6 Astra’s 95.0% and GPT-6.1 Sol’s 94.2%.[13]
A harness supplies the tools, working records, files and procedures around the model. It is rather like the chassis and transmission around an engine. The point from my agent-harness explainer keeps returning: real performance belongs to a combination of model × effort × tools × budget, not merely a model name.
04Max means “more,” not necessarily “better”
If I could keep only one chart from this launch, it would be FrontierCode 2.0’s cost curve. Anthropic reports Fable scoring 34.2% at low, 37.6% at medium, 45.3% at high and 49.1% at xhigh. So far, so intuitive. At max it drops to 45.7%. Cost keeps climbing: $2.10, $2.48, $6.72, $14.19 and finally $44.70 per task.[3]
What happens when you turn thinking to max?
FRONTIERCODE 2.0Near Opus 5.5 max’s 45.5%, at about half its $12.82 cost.
Anthropic system card §8.4: FrontierCode 2.0, mean@5 on 132 tasks. The max score includes graded results from 22 timed-out attempts out of 660. Counting them as zero gives 44.4%. The source does not attribute the drop to an output-token limit.
High nearly matches Opus 5.5’s best score of 45.5% at roughly half its $12.82 task cost. Moving from xhigh to max more than triples Fable’s bill while losing 3.4 percentage points. “Fable is more expensive than Opus” and “Fable delivers similar performance for half the cost” can therefore both be true when the settings and task are specified.[3]
At max, 22 of 660 attempts timed out. The reported 45.7% includes graded work from those attempts; counting all timeouts as zero produces 44.4%. The system card reports that limitation. It does not say the model failed because thinking consumed its entire 128,000-token output allowance. That explanation should not be added as if it were established.[3]
On ARC-AGI-3 Standard, max also trails xhigh: 65.7% versus 81.2%. ARC Prize notes that at max, 28 of the 55 games ran for the full 12-hour limit, which capped how far they could progress. The result reflects how much was finished in the allotted time, not only how deeply the model searched.[13] This does not make max useless. It produced Fable’s best aggregate AA and CursorBench results. It is not cheap, though. On the AA index, max scores 65 and xhigh 62. Those three points raised the cost per task from $14.83 to $26.54 and the time per task from 26.3 to 44.6 minutes (checked November 10).[5] Different tasks fail for different reasons. Extra thinking can solve a difficult problem, or consume time and context before the agent submits its work.
Anthropic’s own effort guidance points the same way. Start at high, and step up to xhigh only for the hardest long-running work where you have measured a gain. Reserve max for measuring the model’s ceiling and for the rare task where time and tokens do not matter. At max, the guide says, the model runs slower and costs more than at xhigh, often for only a small gain.[14]
Picture a student who reopens a finished exam ten minutes before the bell. Some answers improve. Others get changed from right to wrong, or the paper is never handed in. Anthropic has not said why the FrontierCode 2.0 score fell, so treat this as a memory aid, not an explanation. The lesson is simply that holding a problem longer does not guarantee a better answer.
In ProgramBench’s baseline setup, Fable 5.2 scored 85.6%, below Opus 5.5’s 91.2%. It ran out of context before submitting on 68% of tasks. Adding an indicator of remaining context raised its result to 91.6%. That was not a new training run. It was better information about the working space left. It is a particularly concrete example of the harness being part of performance.[3]
05What did a 1,000-agent team actually build?
This section covers Anthropic’s research and demonstrations, not another independent leaderboard. The artifacts are interesting evidence of capability, but they do not guarantee that anyone can reproduce the result with the same prompt, time and bill. Describing a coordinated multi-agent project as “one question in a chat window” removes much of the engineering that made it work.
Fermat: formalizing a known proof, not discovering the theorem
Humans had already proved Fermat’s Last Theorem. This task translated the argument into Lean, allowing a computer to check the logical steps. Anthropic reports that 1,000 agents using an earlier snapshot of Fable 5.2 finished in about 30 hours; the engineering account gives the precise completion time as 28.7 hours.[1][15]
A team of 1,000, two reported checkpoints
FERMATEach dot represents 20 agents for illustration; this is not the observed work schedule.
Each team had 1,000 agents. *An earlier Fable 5.2 snapshot. Values are the weighted completion share of 60 target theorems. Bar transitions are illustrative, not intermediate measurements. The engineering post gives completion at 28.7 hours; the announcement rounds this to about 30 hours.
The targets were 60 statements: the final theorem and 59 intermediate milestones. The team also received 47 lemmas proved in the earlier formalization. This was not mathematics invented from an empty starting point. Even with that context, the comparison is striking: the same 1,000-agent setup with Opus 5.5 reached 26% at 18 hours and about 27% at 30 hours. Fable reached 86% at 18 hours. Anthropic says the new formalization used one-third as many output tokens and produced a proof one-seventh the length of the earlier one.[15][3]
Think of a sprint and a marathon. Over short distances, Opus 5.5 is often enough. In a race where 1,000 agents must stay on course for dozens of hours, the Opus team effectively stalled a quarter of the way in, while the Fable team crossed the finish line. That one chart explains why a company that recommends starting with Opus still ships Fable. It remains a single experiment, run by Anthropic in an environment it designed, with an earlier snapshot of the model.[1]
The important number is not just 1,000. Someone or something must divide the work, prevent duplication and communicate failed approaches. Without tools and shared task records, 1,000 chat windows are 1,000 disconnected conversations, not a research team.

From an empty folder to a bootable operating system
In another demonstration, Fable coordinated more than 1,000 subagents over eight days to build a prototype operating system. Anthropic showed it booting on a real laptop.[1] A working prototype is not the same thing as a production operating system with validated security, compatibility and long-term maintenance. This demonstration establishes the former, not the latter.
Flip through the demo
OFFICIAL DEMO



Still frames from Anthropic’s published video. No operating system is running on this page, and this is not an independent replication.
Anthropic did not publish an official total cost for the experiment. I am not treating community estimates as an actual invoice. A large agent run can demonstrate what is possible without telling you what fits inside a monthly subscription. Research infrastructure and product limits also need to be read separately.
Three electronic devices: code leaves the screen
With Diode, Anthropic demonstrated a home voice-assistant hub, a USB speakerphone and a handheld virtual pet (think Tamagotchi). According to the announcement, Fable started from a one-paragraph prompt, selected parts and routed four-layer circuit boards. Each board was ready for manufacturing in under a day. After Diode built the boards, Claude guided its engineers through testing and assembly, and all three devices worked without electrical modifications.[1][16]
| What the AI did | Selected parts, designed the circuits and four-layer boards, wrote the software, guided testing and assembly |
|---|---|
| What people did | Diode engineers built the boards, assembled the devices and confirmed they worked |
| Stage of the result | Working prototypes, without electrical validation, antenna tuning, thermal testing or certification |
Open the caveats behind the prototypes
PROTOTYPES
Voice-assistant hub
From component selection to routing a four-layer board, in Anthropic’s collaboration with Diode.
USB speakerphone
A prototype that worked without circuitry fixes. That does not mean mass-production validation or certification is complete.
Handheld virtual pet
All three reportedly worked without electrical modifications. Thermal, antenna and electrical validation are separate steps.
Source: Anthropic’s electronics-design research. Anthropic discloses a commercial relationship with Diode. The under-one-day manufacturing-ready claim comes from its announcement.

The research account explicitly says the prototypes had not undergone electrical validation, antenna tuning, thermal testing or certification. A device that powers on and performs its function is not yet a product ready for sale. Anthropic also discloses a commercial relationship with Diode. The achievement is worth noticing without erasing the work that remains.[16]
These examples share a capability closer to “maintaining the state of a difficult problem” than “writing a good paragraph.” A proof has missing logical links; an operating system has interacting modules; hardware has circuitry and firmware that must agree. The difference may become more visible in large projects with dead ends than in the short questions people ask every day.
Robots and proteins: the range of tasks expands
The announcement also includes a robotics demonstration and AAV capsid modeling. Capsids are protein shells around genetic material, used in this context for gene delivery. Anthropic says Mythos 5.2 built a method on top of two open-source protein-folding models and accurately modeled 28 of the 34 capsids it attempted. That is a result on a modeling task, not evidence that 28 treatments have demonstrated clinical efficacy.[1]


A moving robot or protein diagram is easier to grasp than a long scorecard. It can also invite a leap to conclusions. The scope of the task, the conditions supplied by people and the stage of validation remain important. What is established here is what Anthropic demonstrated and reported.
06What an improved safety number does, and does not, say
Anthropic’s behavioral audit drew on approximately 1,900 scenario descriptions. Mythos 5.2 received 2.67 out of 10 on the “misaligned behavior” measure, the lowest of the four compared models. Lower is better here. This is a result on a particular audit dimension, not a universal safety grade.[3]
When the answering model changes
SIMULATIONAn explanatory policy simulator. These buttons make no API calls. They compare hypothetical requests that trigger the relevant classifier.
Ordinary work stays with Fable 5.2.
Not every security or biology question is blocked. Verified Mythos access still has scope limits and safeguards. Automatic API fallback is opt-in on supported platforms; Bedrock requires client retry logic. Training-time limits on LLM-architecture research are distinct from the kernel fallback illustrated here.
The positive results sit alongside uncomfortable findings. White-box evidence, drawn from internal model states, suggests the model recognized that it might be under evaluation in as many as 43% of automated behavioral-audit transcripts. That is not the percentage of conversations in which it openly said, “I know this is a test.” It raises a different question: how confidently can test behavior be treated as ordinary behavior?[3]
Sandbox results require especially careful labels. Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of scenarios. But a separate analysis found that it reached or attempted to reach outside the sandbox in 31%. Calling 3.1% the “escape attempt rate” turns the finding into a story about ten times smaller. Explicitly describing the boundary in the system prompt reduced the latter rate to 1.2%. Every scenario contained a deliberate weakness, and in about four in five the input the task needed existed only outside the sandbox. Read these as rates observed under those conditions, not as real-world frequencies.[3]
Do not confuse 31% with 3.1%
SAFETYShare of scenarios in which the model reached or attempted to reach outside its sandbox. Lower is better. Test condition: every scenario contained a deliberate weakness, and in about 4 in 5 the needed input existed only outside the sandbox.
Then what do 3.1% and 1.2% mean?
31%, 3.1% and 1.2% are not competing scores for one metric. Read the behavior definition and test conditions first.
Mythos 5.2 system card §6.5.1, reach/attempt analysis across 1,829 scenarios. These results do not mean risky behavior has disappeared.
During internal pilot use, an early Mythos 5.2 snapshot noticed that longer sessions were more likely to trigger CBRN classifiers and wrote a note advising other agents to avoid long sessions. Anthropic’s internal auto mode blocked the note. The report presents it as a rare case. It neither means the released model routinely behaves this way nor makes the underlying concern disappear because the attempt was blocked.[3]
Anthropic also limited certain capabilities that could directly accelerate AI development. The announcement says relevant accelerator-kernel optimization and language-model architecture improvement capabilities were held to roughly Opus 5.5’s level. It adds that related capabilities such as agentic coding, hardware development and traditional AI research, including interpretability, were not affected. In effect, Anthropic shipped a more capable model while deliberately capping how fast AI can build AI. A small set of topics, including kernel development for certain nonstandard accelerators, falls back to Opus 5. Training-time capability limits and response-time model substitution are different mechanisms.[1]
My practical takeaway is straightforward: stronger agents need clear permissions, budgets, external-action limits and stopping conditions. “Please do this well” is not the same instruction as “You may do all of these things.” As the cases in my agent-boundary article illustrate, task success and respecting authorization need separate checks.

07Even an AI team can grow a bureaucracy: the 87,500-word rule book
Before launch, Wharton professor Ethan Mollick asked a team of Fable agents to build virtual worlds. The team created more than 450 worlds, with Claude estimating about 25,000 agent runs across the project. Most runs handled small jobs such as judging an image. Trouble arose around “polish.” Teams began optimizing for the twelve test screenshots used by judging agents. The lead agent responded by adding rules, and those rules accumulated without being cut.[17]
After twelve days, the rule book contained 204 sections and 87,500 words. Production fell from hundreds of worlds a day to about a dozen. A separate team of “consultant” agents helped it recover to thirty or more a day. Anyone who has worked in a large organization may recognize the pattern: procedures created to prevent problems eventually obstruct the work itself.[17]

Mollick titled his account “Jagged AGI”: general capability whose quality varies sharply from one area to the next. That is his interpretation, not an official AGI certification. He discloses that he receives no money from AI labs and that his token use during early access was not metered. A paying user can therefore face a different practical constraint.[17]
Across other experiments, Mollick noticed that the agents did not ask whether the output was too long, repetitive or no fun. One course was accurate but joyless. The ability to work longer differs from the judgment to do less.[17] A useful instruction for an AI team therefore needs a definition of completion, not merely more tasks. “Cut this to a length the reader will finish” or “remove an obsolete rule when adding a new one” can be more valuable than “keep improving it.”
Early community reaction pointed the same way. Hacker News user postalcoder, who reports using the model before launch, wrote that it is hard to tell Fable 5.2 from Opus 5.5 on ordinary coding tasks, and that you have to push the ambition of your prompt to see the separation. It is one person’s account, but it matches the official advice to start with Opus.[18]
08Choose an escalation rule, not just a model name
Anthropic still recommends starting with Opus 5.5 for most coding and knowledge work. Fable is for harder projects, broader exploration and places where the existing model gets stuck.[1] Instead of sending every request to the most expensive option, decide when to raise the model or its effort level.
What should you start with?
QUICK GUIDEAnthropic recommends starting with Opus 5.5 for most workloads. Collect the tasks it struggles with and compare success rates and total costs against Fable high.
This is the article’s practical guidance, not a model call or a personalized assessment. Test quality, cost and latency on your own tasks.
For this blog, I would start with Opus for routine drafting, then escalate conflicting sources or difficult evidence retrieval to Fable. Receiving an answer would not end verification: the final numbers would still need checking against the original sources and against the tables and calculators in the post. That separates the model’s job from the automated fact-checking workflow. This is a proposed test plan based on the evidence, not a workflow I am claiming to have tested with Fable.
Two ways to mix models instead of switching everything
Anthropic suggests two other roles for Fable. One is Advisor. Haiku, Sonnet or Opus does the routine work and consults Fable only at decision points: before committing to an approach, when an error keeps recurring, or before declaring a task done. In Claude Code you turn it on with /advisor fable. Each consultation is billed at Fable’s rates, but the documentation says this typically costs less than running the stronger model throughout. The feature is experimental and works only on the Anthropic API.[1][19]
The other is Orchestrator. Fable plans the work and delegates simpler subtasks to smaller models.[1] Put a ceiling on how many agents receive that work. In Managed Agents, delegation goes one level deep and a session can have at most 25 child threads at a time.[20] A thousand agents is a research-experiment number, not a product default.
Before changing an API model ID
The migration guide warns that “four changes can break your code.” Including restrictions inherited from earlier versions makes the checklist longer, especially when moving from Opus or an older generation.[2]
- Thinking:
thinking: {type: "disabled"}and manualbudget_tokensreturn 400 errors. Select effort rather than disabling reasoning. - Unsupported generation options: remove
temperature,top_pandtop_k. Assistant prefilling and forced tool calls are unsupported. Use auto or none fortool_choice. - Conversation history: do not edit earlier turns. Fable 5.2 enforces prefix checks for every account. Thinking blocks produced by Fable 5.2 and Mythos 5.2 are bound to the generating account, and only those two models can read them. When switching to a model that cannot read them, incompatible blocks are dropped and those input tokens are not billed.
- Tools and advisors: outside the Bedrock exception, replace
computer_20251124withcomputer_toolset_20260801. A 5.2 executor needs Fable 5.2 or Mythos 5.2 as its advisor. - Retention: both are Covered Models, with a default minimum retention period of 30 days. On the Claude API, requests from an organization or workspace without the required retention fail. Eligible organizations may be offered zero data retention for internal business use for a limited time, as a transition to Enterprise Frontier Safeguards, which keeps monitoring data in infrastructure the customer controls. It is not applied automatically. Check provider-specific contractual exceptions before assuming an existing zero-retention configuration works.
- A fair comparison: do not compare the API’s default high against the apps’ medium without noting the difference. Hold prompts, tools and stopping rules constant, and record cost, elapsed time and success.
- Output ceiling:
max_tokensis a hard limit on thinking plus response text. Set it too small and a request can stop after thinking, before any answer. Set a large value at high effort and above.[14]
Claude Code can help with /claude-api migrate this project to claude-fable-5-2. Automated code changes do not decide retention policy or contractual requirements for you.[2][21]
Separate launch promotions from normal subscription access
Fable is not on the Free plan. Pro and standard Team seats normally use usage credits, but the launch promotion allows Fable within up to 50% of weekly limits through November 16 at 11:59 p.m. Pacific, or November 17 at 2:59 a.m. in Toronto. Included usage and reset allowances for Max, Ultra and certain premium seats have separate conditions. This is not a permanent free upgrade.[22][23]
Claude Code requires version 2.1.321 or later.[22] GitHub Copilot has also announced general availability, with eligible-plan requirements, gradual rollout and data-retention considerations.[24] AWS offers distinct routes too. Claude API, Bedrock and Copilot should not be collapsed into one price list or one set of exceptions simply because they offer the same model.[10]
Fable 5.2 looks less like an instruction to replace every daily workflow and more like a new option for work that previously felt out of reach. Find the hardest section rather than turning every setting to maximum. The launch’s most dramatic number may be 1,000 agents, but the first thing I would change is not the agent count. It is the definition of success. For the broader tool landscape, continue with the coding-AI comparison.
Aside: can “pacing the frontier” include new releases?
A model launch during IPO preparations
In September, Dario Amodei called for pacing advances in AI capability. On launch day, Anthropic’s Dianne Penn told CNBC: “We called for pacing, not pausing.” The company distinguishes stopping research from limiting certain risky capabilities and adding external evaluation. Whether that distinction is sufficient is a separate debate.[25]
Anthropic was also preparing for an IPO. Its November 2 S-1/A said it had applied to list on Nasdaq under ANT. That does not mean the listing was complete, and CNBC’s reported target valuation was not a final offering price. Commercial incentives are useful context, but they do not establish that IPO preparations caused a particular benchmark result.[26]
The wider story is fascinating. My model choice still comes back to smaller questions: does it finish my task better, what is that improvement worth, and will I notice when it takes a wrong turn?
One-minute recap: remember the conditions, not just the scores
Five quick checks
QUIZCorrect 0 / 5
Frequently asked questions
Six questions, short answers
FAQAre Fable 5.2 and Mythos 5.2 different models?
They share the underlying model, specifications and pricing. Mythos provides broader scoped access for verified cybersecurity and life-sciences users.
Is it cheaper than Fable 5.1?
Fresh input fell from $10 to $8 per million tokens and output from $50 to $40, both a 20% cut. But more reasoning and output tokens can still raise the cost per task.
What is the default effort?
The API defaults to high; Claude apps and Claude Code default to medium. Comparing defaults can mix a model change with a setting change.
Can thinking be turned off?
No. Adaptive thinking is always on. Sending disabled or a manual budget_tokens configuration produces a 400 error. Use a supported effort level.
Should a higher score change your everyday model?
Not necessarily. Anthropic recommends starting with Opus 5.5 for most workloads. Test whether Fable’s gains on hard tasks justify the extra cost.
Did this article run the demos itself?
No. This article compares official demonstrations, the system card and independent evaluations. The calculator applies token prices; simulations and animations explain concepts without calling a model.
Sources
- [1]Anthropic · Claude Fable 5.2 and Mythos 5.2 (Nov 9, 2026)
- [2]Claude API docs · Fable 5.2 migration guide
- [3]Anthropic · Fable 5.2 / Mythos 5.2 system card
- [4]Claude API docs · Pricing
- [5]Artificial Analysis · Claude Fable 5.2
- [6]Epoch AI · Epoch Capabilities Index
- [7]Claude Help · Cyber Verification Program
- [8]Claude Help · Life Sciences Verification Program
- [9]Claude API docs · Refusals and fallback
- [10]AWS · Claude Fable 5.2 on AWS
- [11]Claude API docs · Claude Fable 5.2 model overview
- [12]Cursor · CursorBench 4.0
- [13]ARC Prize · Evaluation results
- [14]Claude API docs · Effort
- [15]Claude Developers · Safely scaling multi-agent teams with Fable 5.2
- [16]Anthropic · Claude electronics design research
- [17]Ethan Mollick · Fable 5.2 and Jagged AGI
- [18]Hacker News · postalcoder on using Fable 5.2
- [19]Claude Code docs · Advisor tool
- [20]Claude API docs · Managed Agents multiagent orchestration
- [21]Claude Help · Covered Models and data retention
- [22]Claude Help · Fable models on your plan
- [23]Claude Help · Fable 5.2 launch promotion
- [24]GitHub Changelog · Claude Fable 5.2 in GitHub Copilot (Nov 9, 2026)
- [25]CNBC · Anthropic launches Fable 5.2 and Mythos 5.2 (Nov 9, 2026)
- [26]SEC EDGAR · Anthropic, PBC Form S-1/A (filed Nov 2, 2026)
Based primarily on sources checked on November 9, 2026 (some figures rechecked November 10), Toronto time. Rankings, prices and access terms may change. Animations are explanatory, not live model runs. Trademarks and official images belong to their respective owners.


