On November 9, 2026 (US time), Anthropic released Claude Fable 5.2. The same model, with its cybersecurity and biology safeguards removed, is offered only to vetted organizations as Claude Mythos 5.2. Pricing is $8 per million input tokens and $40 per million output tokens: 20% cheaper than Fable 5.1, but twice the price of Opus 5.5.
The official models overview still says that if you're unsure, start with Claude Opus 5.5. Fable 5.2 is positioned as the model to move up to when Opus 5.5 still falls short on your own evaluations, even at a higher effort. This article compares pricing against each relevant model, separates Anthropic's comparison table from third-party measurements, and covers what breaks when you switch over in the API and what changes in the apps and Claude Code (based on Anthropic's announcement, documentation, and help center, checked November 10, 2026).
Where Fable 5.2 sits, and how its pricing compares
Claude Fable 5.2
Demanding reasoning and long-running agentic work (when Opus 5.5 isn't enough)
Claude Opus 5.5
Start here if you're unsure (official models overview)
Sonnet 5.5 / Haiku 5.5
Balance of speed and cost / high-volume, low-latency work
vs. Fable 5.1
20% cheaper across the board
Input $10→$8, output $50→$40, cache reads $0.25→$0.20
vs. Opus 5.5
2× on input and output
Twice Opus 5.5's $4 input / $20 output (cache reads are the same $0.20)
vs. GPT-6 Astra
20% cheaper on input and output
Astra is $10 input / $50 output (OpenAI standard pricing, short context)
Sources: Anthropic, "Models overview" and "Pricing"; OpenAI, "Pricing" (all checked November 10, 2026). "20%" and "2×" are calculated from the price lists.
1. What Fable 5.2 is: the model above Opus 5.5
Fable is the top tier of Anthropic's current models (for how the tiers relate, see our Claude Fable 5 explainer). Fable 5.2 succeeds Fable 5.1, released in September, and the official models overview has moved Fable 5.1 to the legacy models (still available). The announcement highlights four main changes:
- Improved coding, reasoning, and vision (table in Section 4)
- Output generation up to 30% faster than Fable 5.1
- Pricing lowered to $8 input / $40 output (Fable 5.1 was $10 / $50)
- In early testing on claude.ai, Claude Code, and Cowork, the safeguards blocked requests unnecessarily about 55% less often than with Fable 5.1 (Section 9)
On when to use it, the announcement says to start most coding and knowledge work with Opus 5.5, then move up to Fable or Mythos or down to Sonnet 5.5 or Haiku 5.5 depending on the scale and difficulty of the task. The developer page "What's new in Claude Fable 5.2" says the same: move to Fable 5.2 when Opus at a high effort still doesn't meet your requirements in evaluation. The announcement also mentions using it as an advisor that Haiku, Sonnet, and Opus consult on hard problems, and as an orchestrator that hands simple tasks to smaller models.
Mythos 5.2 is the same model as Fable 5.2. The difference is that Fable 5.2 carries strong safeguards against cybersecurity and biology misuse and is available to everyone, while Mythos 5.2 is available to organizations working on cybersecurity or biomedical research that pass a vetting process (the Cyber Verification Program or the Life Sciences Verification Program) (Section 9).
2. Key specs: Fable 5.2 vs. Fable 5.1 and Opus 5.5
| Item | Fable 5.2 | Fable 5.1 (legacy) | Opus 5.5 |
|---|---|---|---|
| API model ID | claude-fable-5-2 | claude-fable-5-1 | claude-opus-5-5 |
| Pricing (input / output) | $8 / $40 | $10 / $50 | $4 / $20 |
| Context window / max output | 1M / 128K | 1M / 128K | 1M / 128K |
| Thinking | Always on (can't be disabled) | Always on (can't be disabled) | Always on (can't be disabled) |
| Default effort in the API | high | high | medium |
| Reliable knowledge cutoff | June 2026 | June 2026 | June 2026 |
| Speed (official relative rating) | Moderate | — | Moderate |
| Retirement | Not before November 9, 2027 | Not before September 1, 2027 | Not before September 22, 2027 |
Sources: Anthropic, "Models overview," "Claude Fable 5.1," and "What's new in Claude Fable 5.2" (checked November 10, 2026). Speed is a relative rating among the four current models; it isn't listed for Fable 5.1, which is a legacy model. Retirement dates are commitments for platforms operated by Anthropic.
Of the rows that affect how you use the models, only pricing and default effort differ (the retirement dates differ too). The context window, max output, and knowledge cutoff are the same for all three, and none of them lets you turn thinking off. In the API, Fable 5.2 is available on the Claude API, Amazon Bedrock (anthropic.claude-fable-5-2), Claude Platform on AWS, Google Cloud (both the Anthropic-operated and Google-operated versions), and Microsoft Foundry, and it launched on every platform on day one.
3. Pricing: "cheaper" or "pricier" depends on what you compare it to
Sources: Anthropic, "Pricing"; OpenAI, "Pricing" (checked November 10, 2026). GPT-6 Astra has a higher price list for long context.
Fable 5.2's cache reads cost 0.025× the input price, the same multiplier as Fable 5.1. Opus 5.5's multiplier is 0.05×, so even though input prices differ by 2×, cache reads cost the same $0.20. Cache writes are $10 for 5 minutes and $16 for 1 hour: twice Opus 5.5 ($5 / $8) and 80% of Fable 5.1 ($12.50 / $20).
With heavy caching, the gap narrows slightly from 2×
In agentic work, the same instructions and loaded files are read from the cache over and over, so most of the input ends up as cache reads. As an example, suppose a single task uses 500K tokens of cache reads, 50K tokens of new input, and 50K tokens of output (excluding cache writes).
| Model | Cache reads | New input | Output | Total (calculated) |
|---|---|---|---|---|
| Fable 5.1 | $0.125 | $0.50 | $2.50 | About $3.13 |
| Fable 5.2 | $0.10 | $0.40 | $2.00 | $2.50 |
| Opus 5.5 | $0.10 | $0.20 | $1.00 | $1.30 |
Under these assumptions, Fable 5.2 costs 80% of Fable 5.1 and about 1.9× Opus 5.5. Cache reads cost the same, but output accounts for most of the bill, so the gap narrows only a little. In practice, though, cost depends on how many tokens you use. The default effort in the API is high for Fable 5.2 and medium for Opus 5.5, so if you switch without specifying effort, the model thinks one level harder and output tends to grow. When comparing, set the same effort and measure with the usage field in the response.
Batch and fast mode
The Message Batches API (asynchronous bulk processing) is half price, so Fable 5.2 costs $4 input / $20 output, the same as Opus 5.5's standard pricing. But Opus 5.5 batches are also half price ($2 / $10), so batch against batch, Fable 5.2 is still 2×. If you've decided to use Fable 5.2 for work that isn't urgent, batch is the cheapest way to run it.
Fast mode, which makes output up to 2.5× faster, costs $16 input / $80 output on Fable 5.2. That's twice Opus 5.5's fast mode ($8 / $40), and also twice the standard price. In the API, you just add speed: "fast" to the request; no beta header is needed. It's available on the Claude API, Google Cloud (both the Anthropic-operated and Google-operated versions), Anthropic-hosted Microsoft Foundry (as a separate model, claude-fable-5-2-fast), and Claude Code. It can't be combined with batch.
4. Benchmarks: Anthropic's comparison table
Here is the comparison table from the announcement page as published. All values are published by Anthropic, under these conditions:
- Unless noted, values are at the highest effort each model supports. Unless noted, Fable 5.2, Fable 5.1, and Opus 5.5 were measured with safeguards on.
- On FrontierCode 2.0, Fable 5.2 and Fable 5.1 are at xhigh and Gemini 4 Carbon is at medium. On Remote Labor Index, Fable 5.2 is at high.
- In the three cybersecurity and protein design rows, the Fable 5.2 column shows Mythos 5.2, the Fable 5.1 column shows Mythos 5.1, and Opus 5.5 is measured without safeguards. On CyScenarioBench, the GPT-6 Astra column shows GPT-6 Cyber, and each company's score is based on a different set of problems.
- RoboDojo Sim was run internally by Anthropic. There's no note on who measured the other GPT and Gemini rows.
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0Agentic coding | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2Agentic coding | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBenchBusiness workflows | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1Agentic scientific research | 74.6%Mythos 5.2: 75.3% | 52.6% | 58.7% | 68.1% | 64.8% |
| CyScenarioBenchCybersecurity | 76.3%Mythos 5.2 | 61.7%Mythos 5.1 | 67.6%No safeguards | 61%GPT-6 Cyber | — |
| CVE RediscoveryCybersecurity | 89.6%Mythos 5.2 | 78.3%Mythos 5.1 | 81.3%No safeguards | — | — |
| De novo binder design (24 hours)Protein design | 85.1%Mythos 5.2 | 79.1%Mythos 5.1 | 82.6%No safeguards | — | — |
| WANDR (relaxed score)Agentic search | 82.1% | 68.7% | 72.3% | — | — |
| OSWorld 2.1 (offline set, partial)Computer use | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo SimRobot control | 64.9% | 27.6% | 39.4% | 29.8% | — |
| AA-Briefcase v1.1 (rubric)Knowledge work | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1 (Elo)Knowledge work | 1785 | 1669 | 1800 | 1569 | 1611 |
| Remote Labor IndexKnowledge work | 41.3% | 17.9% | 21.3% | 20.8% | — |
| Humanity's Last Exam (with tools)Cross-domain reasoning | 74.2% | 65.6% | 67.7% | 57.2% | — |
| Chartography (no tools)Chart reading | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Source: Comparison table in Anthropic's "Introducing Claude Fable 5.2 and Claude Mythos 5.2" (published November 9, 2026; checked November 10). Green marks the top score in the row; orange marks the top score in rows where another model beats Fable 5.2.
Fable 5.2 (Mythos 5.2 in three rows) has the top score in 13 of 15 rows. In the other two, Gemini 4 Carbon leads on AutomationBench (53.9% vs. 51.4%) and Opus 5.5 leads on AA-Briefcase Elo (1800 vs. 1785). On the same AA-Briefcase, Fable 5.2 has the top rubric score.
The gap with Opus 5.5 varies a lot by row
Fourteen rows have percentage values for both Fable 5.2 and Opus 5.5. Excluding the three cybersecurity and protein design rows, which compare Mythos 5.2 against Opus 5.5 without safeguards under different conditions, here are the gaps for the remaining 11 rows (calculated from the table).
The biggest gaps are in robot control (RoboDojo Sim), Remote Labor Index (completing remote freelance projects), and agentic scientific research (Terminal-Bench-Science). Coding, on the other hand, is in the single digits: +3.6 on FrontierCode 2.0 (Fable 5.2 at xhigh, Opus 5.5 at its highest supported effort) and +7.6 on FrontierSWE v2. OSWorld, for computer use, is +2.1. That's a weak case for moving all your everyday coding to Fable 5.2, and it lands on the same conclusion as Anthropic's own advice to start with Opus 5.5.
5. Effort and cost per task: Anthropic's numbers vs. third-party numbers
Even at twice the unit price, if a lower effort gets you the same result, the cost per task goes down. Anthropic's announcement and third-party measurements from Artificial Analysis look at different conditions.
Anthropic's measurements (from the announcement's chart captions)
Matches Opus 5.5 at max with a lower effort
On FrontierCode 2.0, Fable 5.2 scores 45.3% at high, nearly the same as Opus 5.5 at max (45.5%), at about half the cost per task. At xhigh it scores 49.1%.
On WANDR, it matches Opus 5.5's best result at medium, at about 40% of the cost. On Humanity's Last Exam, it matches at low, at less than a tenth of the cost. On Chartography, it scores 66.4% at medium (Opus 5.5 at max scores 64.4%), at less than a tenth of the cost. The comparison is against Opus 5.5 at max, not against Opus 5.5's API default (medium).
Third-party measurements (Artificial Analysis, both at max)
About 2.6× per task, max vs. max
On the Intelligence Index, Fable 5.2 scores 65 (1st of 74 models) and Opus 5.5 scores 59 (5th).
Cost per task for the index is $26.54 vs. $10.04, and time per task is 44.1 minutes vs. 25.1 minutes. That's about 2.6× the cost and about 1.8× the time (calculated).
Sources: Anthropic's announcement ("Accuracy vs. cost" captions for each benchmark); Artificial Analysis model pages for "Claude Fable 5.2" and "Claude Opus 5.5" (both measured at max effort; the setting is labeled "Max, Default Fallback" on the page; checked November 10, 2026).
The two don't contradict each other. Run Fable 5.2 at max and it clearly costs more than Opus 5.5; for tasks where a lower effort still reaches Opus 5.5's max result, it costs less. Which case applies depends on the task. Anthropic's recommendation is to start with the API default of high, raise it to xhigh only for long-running coding and agentic work where you've measured a benefit, and save max for measuring the ceiling and for rare tasks where time and cost don't matter. Where quality holds up, drop to medium or low. For what each effort level means, see our guide to the effort setting in Claude Code.
6. When to use Fable 5.2 vs. Opus 5.5 or GPT-6 Astra
Stay on Opus 5.5 Everyday coding, writing, and research. As Section 4 shows, even Anthropic's table puts the coding gap in the single digits and computer use at +2.1, at half the unit price.
Move up to Fable 5.2 Tasks where Opus 5.5 still falls short on your own evaluations even at a higher effort: work that runs autonomously for hours, orchestrating many subagents, or acting as the advisor other models consult when they get stuck.
Use batch for high-volume work that needs Fable 5.2 but isn't urgent Batch pricing for Fable 5.2 ($4 / $20) matches Opus 5.5's standard pricing. If Opus 5.5 is good enough for the job, Opus 5.5 batch ($2 / $10) is cheaper.
Work Fable 5.2 won't do Most high-risk cyber work and research-oriented biology get blocked by the safeguards, and another model answers instead (Section 9). In these areas, vetted organizations use Mythos 5.2.
Your API agreement may rule it out Organizations running under a zero data retention (ZDR) agreement can't use it unless they enable 30-day data retention, unless Anthropic explicitly approves otherwise (Section 9). Priority Tier is also not supported, same as Opus 5.5.
Compared with GPT-6 Astra
On pricing, Fable 5.2 is $8 input / $40 output, while OpenAI's price list has GPT-6 Astra at $10 input / $50 output for short context. Fable 5.2 is 20% cheaper on input and output, and cache reads are $0.20 vs. $1.00, one fifth the price (calculated).
On performance, Fable 5.2 (Mythos 5.2 in the cyber rows) beat GPT-6 Astra (GPT-6 Cyber in the cyber rows) in all 12 rows of Anthropic's table that have an Astra value. However, the table has no note on who measured the GPT-6 Astra values or how, and CyScenarioBench values were measured by each company on different problems. For the values and conditions OpenAI itself published for Astra, see our GPT-6 Astra explainer. The most reliable way to choose is to give both the same tasks from your own work.
7. What breaks when you switch in the API
The "Claude Fable 5.2 and Claude Mythos 5.2 migration guide" lists what to check depending on which model you're coming from. Here are the items for moving from Fable 5.1 and from Opus 5.5 side by side.
| What changes | From Fable 5.1 | From Opus 5.5 | What happens |
|---|---|---|---|
| The check for rewritten earlier turns (prefix check) now applies to all accounts | ✕ | ✕ | 400 errors even on accounts created before August 31, 2026 (explained below) |
| Fable 5.2's thinking blocks can only be read by Fable 5.2 and Mythos 5.2 | ✕ | ✕ | No error, but switching back to Fable 5.1, Opus 5.5, or another model discards the reasoning up to that point |
Old computer use tool (computer_20251124) | ✕ | Same (400 outside Bedrock) | Returns 400 everywhere except Amazon Bedrock, so move to computer_toolset_20260801 |
| Advisor model in the advisor tool (beta) | ✕ | ✕ | If the executor is Fable 5.2, the advisor must also be Fable 5.2 or Mythos 5.2, or you get a 400 |
| 30-day data retention required | Same | ✕ | Under ZDR without Anthropic's explicit approval, every request returns 400 (Section 9) |
| No structured outputs for CMEK organizations | Same | ✕ | Remove strict: true and output_config.format, and validate in your own code |
| Refusals by the safeguards | Same | ✕ | Returned as HTTP 200 with stop_reason: "refusal", so handle every category in stop_details |
| Default effort in the API | Same (high) | ✕ | No error, but it goes up one level from medium to high, changing cost, time, and results |
| Rate limits | ✕ | ✕ | Fable 5.2 has its own limits (separate from the limits Fable 5.1 and Fable 5 share), with lower token limits than Opus 5.5 at every tier |
✕ = check before you switch. "Same" = the source model already works this way, so nothing changes. Sources: Anthropic, "What's new in Claude Fable 5.2" and the migration guide (checked November 10, 2026).
If Claude Code, claude.ai, Claude Managed Agents, or the Claude Agent SDK builds the conversation history for you, the prefix check changes nothing. You're affected if your own code builds the messages.
The 400 you're most likely to hit: the prefix check
Fable 5.2's thinking blocks are bound to the system, tools, and earlier messages that came before them. If you then rewrite that earlier content and send the thinking block back, you get the following error. Resending the same request won't fix it.
messages.5.content.0: Invalid `signature` in `thinking` block. The block is bound to a different conversation. Remove the block, or set `thinking.block_binding.prefix_mismatch_behavior` to "drop_block". That setting requires the `thinking-binding-controls-2026-08-01` value in the `anthropic-beta` header.Fable 5.1 and Opus 5.5 applied this check by default only to accounts created on or after August 31, 2026. That's why code that worked on older accounts can stop working the moment you switch to Fable 5.2. You're unaffected if all four of the following are true:
- You don't change
systemortoolsmid-conversation - You only append to
messages, never modifying, reordering, or deleting earlier messages - You send back assistant responses with every block exactly as the API returned them, in the same order
- You don't do any of the things listed in the official "What to check in your integration" section
If you don't pin the server's tool list in the MCP connector, even a change to the tool list counts as a rewrite on Fable 5.2. For how to identify this error and fix it in Claude Code, see also our guide to the "invalid signature in thinking block" error.
Migration steps
- If your organization has a ZDR agreement, first enable 30-day retention on the workspace or confirm with your Anthropic account team whether you can use it.
- Change the model ID to
claude-fable-5-2and set effort explicitly (code below). - Check rate limits before going to production. Fable 5.2's limits are separate from both Fable 5.1 / Fable 5 and Opus 5.5.
- If you build
messagesyourself, add the beta headerthinking-binding-controls-2026-08-01, run one session withprefix_mismatch_behavior: "drop_block", and fix everyprefix_binding_mismatchthat shows up ininput_transformationsin the response. Do this even if you never saw an error on Fable 5.1. - Update the
modelfor the old computer use tool and the advisor tool (changing the advisor counts as rewritingtools, so do it in a new conversation). - Rerun your own evaluations at different effort levels and re-measure cost and time.
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-fable-5-2", # before: claude-fable-5-1 / claude-opus-5-5
max_tokens=32000, # limit includes thinking; go larger at high and above
output_config={"effort": "high"}, # if omitted, the API default (high for Fable 5.2)
messages=[{"role": "user", "content": "..."}],
)
print(response.usage) # measure cost with usage, not unit pricesIf you have Claude Code, you can also ask it to /claude-api migrate this project to claude-fable-5-2, as the migration guide suggests.
Behavior that changes even without code changes (check when bringing over your prompts)
- In Anthropic's testing, instructions like "answer without thinking" didn't stop it from thinking at medium and above. At high and above, it almost always thinks before answering. If simple requests are slow or expensive, lower the effort.
- Thinking counts toward
max_tokens, so with a smallmax_tokensit may stop withstop_reason: "max_tokens"before producing any text. - At xhigh and max, it may spin up subagents to review its own work. Subagents can become a large part of the cost.
- On long tasks, it may end its turn with a list of remaining items or a "Should I continue?" (more common at low and medium).
- If you don't say who owns the code, it tends to add on top of the existing structure and report nearby problems rather than fix them.
- During long runs of tool calls, it may produce very little user-facing text.
- It follows style rules faithfully, and strict sentence-length limits can make its writing stilted.
Sources: Behavior differences in "What's new in Claude Fable 5.2" and Behavior changes in the migration guide.
8. What changes in the apps and Claude Code
According to the help article "Claude Fable models on your plan," Fable 5.2 is available on all paid plans, but whether it's included in your plan varies. Fable 5.2, 5.1, and 5 are all treated the same way.
| Plan | How Fable 5.2 works (excluding launch offers) |
|---|---|
| Max, UltraAlso Team premium seats and premium seats on seat-based Enterprise | Included in the plan; you can use up to 50% of your weekly usage limit at no extra charge (it uses up your limit faster than other models; once you're over, continue with usage credits or switch to another model) |
| Pro, Team standard seats | Not included in plan limits; uses usage credits (pay as you go) |
| Standard seats on seat-based Enterprise | Available if the organization has enabled usage credits |
| Usage-based Enterprise, Claude API | Billed at API rates |
| Free | Not available |
Source: Anthropic help article "Claude Fable models on your plan" (checked November 10, 2026). The 50% doesn't mean your limit goes up by 50%; it's the share of the same weekly limit you can spend on Fable.
It's available in the Claude apps on web, desktop, and mobile, as well as Cowork, Claude Code, Claude Design, Claude for Microsoft 365, and Claude Tag. In the apps, choose "Fable 5.2" in the model selector. Fable 5.2's release doesn't change the default model on individual plans to Fable. According to the announcement, the default effort in the Claude apps and Claude Code is medium, unlike the API default (high).
Claude Code
- Fable 5.2 requires v2.1.321 or later. Starting with this version, the
fablealias points to Fable 5.2. - A saved Fable 5.1 selection moves to
fable: immediately on Pro and Max, and on Team and Enterprise once the organization allows it. To go back to Fable 5.1, use/model claude-fable-5-1. - Fable 5.2's fast mode is also available in Claude Code.
Sources: Claude Code "changelog" 2.1.321 (November 9, 2026); help article "Claude Fable models on your plan."
For Team and Enterprise admins
On Team and Enterprise, the organization's owners decide which models are available. According to the help article, Fable 5.2 is off on Enterprise until an admin turns it on. Seven days after launch (around November 16, 2026, calculated from the release date), Fable 5.2 is added to every role that has any Fable model enabled, and roles whose default is a Fable model move to Fable 5.2. If an admin opts out first, this automatic switch doesn't happen.
9. Mythos 5.2 and the safeguards
Anthropic calls Mythos 5.2 its most capable model for cybersecurity work. On its internal CVE Rediscovery evaluation, Anthropic says it found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1, in relative terms. However, a footnote in the announcement notes that on a separate evaluation in the system card, its vulnerability identification was at roughly the same level as Opus 5.5. It's available only to organizations that pass the Cyber Verification Program or the Life Sciences Verification Program. Claude Enterprise customers can use Mythos 5.2 without being vetted through Claude Security, which looks for vulnerabilities in code. For how AI is changing cybersecurity, see our explainer on how Claude Mythos changed attack and defense.
Fable 5.2's safeguards block high-risk cyber work such as penetration testing, exploit development, and binary vulnerability research, as well as most research-oriented biology. Blocked queries are answered instead by Opus 4.8 for cyber and Opus 5 for biology. According to the announcement, the safeguards are as strong as Fable 5.1's, and in early testing on claude.ai, Claude Code, and Cowork, unnecessary blocks fell by about 55% (about 60% for cyber and about 40% for biology). In the API, refusals come back as HTTP 200 with stop_reason: "refusal", and with the beta fallbacks: "default", the request is automatically retried on the recommended model (Opus 4.8 or Opus 5).
On data handling, Fable 5.2 and Mythos 5.2, like Fable 5.1, require 30-day data retention and can't be used with zero data retention (ZDR) unless Anthropic explicitly approves it. The stated reason is that misuse can only be detected by looking at patterns across multiple requests. For Enterprise, Enterprise Frontier Safeguards (EFS), which applies automated safety monitoring to prompts and responses stored in the customer's own infrastructure and is compatible with ZDR, has started as a private beta. According to the announcement, eligible customers can keep using zero data retention even before they integrate monitoring through EFS, while other customers' data is retained for 30 days for safety monitoring.
Summary
Claude Fable 5.2 is 20% cheaper than Fable 5.1 across the board and has the top score in almost every row of Anthropic's comparison table. But compared with Opus 5.5, it costs twice as much on input and output, and Anthropic itself says to start with Opus 5.5. The biggest gaps are in robot control, Remote Labor Index, and agentic scientific research, while the coding gap is in the single digits (+3.6 on FrontierCode 2.0, +7.6 on FrontierSWE v2).
Cost varies a lot with effort. In Anthropic's measurements, Fable 5.2 at a lower effort matches Opus 5.5 at max in some cases, while third-party Artificial Analysis measured about 2.6× per task with both at max. If you're switching in the API, set effort explicitly, find the prefix check 400s up front, and measure cost with usage. In the apps, Max and Ultra can use it within the plan, and Pro uses usage credits.
FAQ
Q. Should I use Fable 5.2 or Opus 5.5?
A. Start with Opus 5.5. Both Anthropic's official models overview and the announcement recommend starting most work with Opus 5.5 and moving up to Fable 5.2 when a higher effort still isn't enough. Fable 5.2's input and output prices are twice Opus 5.5's.
Q. When was Fable 5.2 released?
A. November 9, 2026 (US time). It became available the same day on the Claude API and the cloud platforms, in the Claude apps, and in Claude Code (v2.1.321).
Q. Can I use Fable 5.2 on the Pro plan?
A. Yes, but normally it isn't included in plan limits and uses usage credits (pay as you go). On Max, Ultra, and similar plans, you can use up to 50% of your weekly limit within the plan. During the launch offer (until November 16, 2026, Pacific Time), Pro can also use up to 50% of the weekly limit within the plan.
Q. Is Fable 5.2 cheaper than GPT-6 Astra?
A. On input and output unit prices, yes. Fable 5.2 is $8 input / $40 output, and OpenAI's price list has GPT-6 Astra at $10 input / $50 output for short context, so it's 20% cheaper (calculated). The cost of a task depends on how many tokens it uses, so compare by measuring on the same tasks.
Q. Will my Fable 5.1 code work if I just point it at Fable 5.2?
A. After changing the model ID, check these four things: (1) whether you build the conversation history yourself and rewrite it mid-conversation (the prefix check 400), (2) whether you move conversations to Fable 5.1 or other models (Fable 5.2's reasoning gets discarded), (3) whether you use the old computer use tool (computer_20251124), and (4) whether the advisor in the advisor tool is Fable 5.2 or Mythos 5.2.
Q. How do I go back to Fable 5.1 in Claude Code?
A. Switch with /model claude-fable-5-1. Starting with v2.1.321, a saved Fable 5.1 selection moves to fable (Fable 5.2).
Update log
- November 10, 2026First published. Written after checking Anthropic's announcement, the API release notes (November 9, 2026), Models overview, What's new in Claude Fable 5.2, the migration guide, Pricing, Effort, the Claude Code changelog (2.1.321), the help articles "Claude Fable 5.2 launch offers" and "Claude Fable models on your plan," OpenAI's Pricing, and the Artificial Analysis model pages.
Sources
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)
- Anthropic: Release notes, Models overview, What's new in Claude Fable 5.2, Migration guide, Pricing, Effort
- Anthropic: Claude Code changelog (2.1.321)
- Anthropic help center: Claude Fable 5.2 launch offers, Claude Fable models on your plan
- OpenAI: Pricing
- Artificial Analysis: Claude Fable 5.2, Claude Opus 5.5
![Claude Fable, Opus, Sonnet & Haiku: Pricing and Performance Compared [2026]](https://arte.itlibra.com/images/articles/claude-models-pricing-featured-en.svg?v=88d7149aea92)
