Fable 5.2 Launch: Benchmarks, Pricing and Everything You Need to Know

• 49.1% on FrontierCode 2.0 at Xhigh effort, against 45.5% for Opus 5.5, in Anthropic’s November 2026 launch table.
• $8 / $40 per million tokens, down from the $10 / $50 Fable 5.1 carried from September 2026.
• About 55% fewer safeguard blocks than Fable 5.1 in Anthropic’s early November 2026 read across claude.ai, Claude Code and Cowork.
• 3.1% sandbox-escape rate for Mythos 5.2 across 1,800-plus adversarial scenarios in 2026, down from 21% for Mythos 5.1.
• 1M-token context, 128K output, service guaranteed through at least November 9, 2027, per the Fable 5.2 model overview in the Claude platform docs.

What Changed in the Fable 5.2 Launch Compared With Fable 5.1

The Fable 5.2 launch on November 9, 2026 cut the price by 20%, sped up output by as much as 30%, roughly halved unnecessary safeguard blocks, and added an optional fast mode, all on the same underlying model that vetted organizations get as Mythos 5.2. Every figure comes from Anthropic’s own Claude Fable 5.2 and Mythos 5.2 announcement. No outside lab had tested the model by launch day.

Three releases in five months. Fable 5 shipped on June 9, 2026, Fable 5.1 followed in September 2026, and 5.2 arrived on November 9. If you’re planning a migration, that cadence matters: whatever you build on 5.2 may have a successor before your next budget cycle.

The spec sheet is stable rather than new. The API model ID is claude-fable-5-2, the context window is 1M tokens, the maximum output is 128K tokens, and the knowledge cutoff is June 2026. Anthropic commits to keeping it in service until at least November 9, 2027.

Fable and Mythos are one model with two front doors. Fable carries classifiers that block high-risk cybersecurity work and most research biology, which is why anyone can use it. Mythos drops those classifiers and goes only to organizations cleared through the Cyber Verification Program or the Life Sciences Verification Program. Claude Enterprise’s Claude Security scanner now runs on Mythos 5.2 as well.

The block-rate drop is the change I’d pay most attention to. Anthropic says the classifiers are as strong as 5.1’s, and the 55% figure is an early read from claude.ai, Claude Code and Cowork rather than a controlled study. Still, if a Fable 5.1 false positive has ever interrupted you while debugging ordinary network code, a 60% drop in cyber blocks is the line that matters, and blocked queries now fall back to Opus 4.8 for cyber and Opus 5 for biology instead of a dead end.

Anti-distillation hardening rounds out the list. Anthropic says new techniques reduce the data illicit distillers can harvest by 90 to 99%, and preserved thinking, which stops anyone editing Claude’s prior context, now applies to every API account.

How Fable 5.2 Scores on Coding, Reasoning and Agentic Benchmarks

Fable 5.2 beats both Fable 5.1 and Opus 5.5 on every coding, reasoning and agentic benchmark Anthropic published on November 9, 2026, with the widest gaps on robotics, remote labor and agentic science. All of these are Anthropic’s own numbers, run with safeguards on and at each model’s highest available effort unless noted. Treat them as a vendor’s claims until someone else reproduces them.

Two rows stand out. RoboDojo Sim more than doubles the 5.1 score, from 27.6% to 64.9%, on a simulated two-arm robot driven from camera images. The Remote Labor Index, which tracks freelance-style knowledge work, jumps from 17.9% to 41.3%. The Fable 5.1 announcement in September 2026 listed 52.6% on Terminal-Bench-Science and 31.4% on AutomationBench, and those are the same baselines 5.2 cites, so the generational comparison is at least internally consistent.

Mythos 5.2, the version without cyber and bio classifiers, has its own rows. CVE Rediscovery rose from 78.3% for Mythos 5.1 to 89.6% in Anthropic’s October 2026 run across 1,829 CVEs. CyScenarioBench went from 61.7% to 76.3%, and 24-hour de novo protein binder design from 79.1% to 85.1%.

“At Xhigh effort, Fable 5.2 reaches 49.1%, the highest score of any released model we’ve tested.” (Anthropic, Fable 5.2 launch post, November 9, 2026)

The efficiency claims interest me more than the raw scores. On FrontierCode 2.0, Fable 5.2 at High effort scored 45.3% against Opus 5.5’s best of 45.5%, at roughly half the cost per task. At Medium effort it matched Opus 5.5’s top WANDR score for about 40% of the cost per task. On Humanity’s Last Exam at Low effort, it hit Opus 5.5’s best for under a tenth of the cost. On BenchCAD 2.1 at Max effort, it scored 83.5 against 65.3 for Opus 5.5 and 44.4 for Fable 5.1.

One caveat before you forward this table. Several of these benchmarks are new or Anthropic-run, RoboDojo was executed internally, and the methodology lives in a system card most readers haven’t opened. Scores this fresh should move your shortlist, and nothing else, until your own evals confirm them.

What Does Fable 5.2 Actually Cost per Million Tokens?

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens on the Claude API as of November 9, 2026, with fast mode at $16 and $80. That’s a 20% cut from Fable 5.1’s $10 and $50, and exactly twice Opus 5.5 per token.

All prices are USD per million tokens from the Claude platform pricing page, which carries no “as of” date, so check it before you sign anything.

The sticker price is only the start. Three add-ons change the bill:

• Cache writes cost $10 per million tokens for a 5-minute cache and $16 for a 1-hour cache.
• US-only data residency via inference_geo adds a 1.1x multiplier.
• Fast mode doubles the base rate across the full 1M context window and can’t be combined with the Batch API.

These multipliers stack. Take a request with 100K input tokens and 20K output tokens. At standard rates that’s $0.80 plus $0.80, or $1.60. Switch on fast mode and it becomes $3.20. Add US-only residency and you’re at $3.52 for the same request. Batch processing runs the other way, halving the standard rate to $0.80 if you can wait.

Here’s the detail that catches teams out. Anthropic’s pricing page says the tokenizer used from Opus 4.7 onward produces about 30% more tokens for the same text than earlier models did. A spreadsheet comparing per-token prices against an older model understates what you’ll pay. Count tokens on your own prompts before you forecast.

Per-token price is the wrong unit for this model. Fable 5.2 is twice Opus 5.5 per token, and Anthropic still recommends Opus 5.5 as the daily driver, with Fable as the escalation when Opus at higher effort falls short. Yet the FrontierCode 2.0 run put Fable 5.2 at High effort level with Opus 5.5 at Max for about half the cost per task. Both things can be true at once. A model that finishes in fewer tokens can be cheaper per job while costing more per token, and the only way to know which side of that line your workload sits on is to measure cost per completed task on your own traffic.

How Fable 5.2 Compares With GPT, Gemini and Other Frontier Models

Fable 5.2 leads GPT-6 Astra and Gemini 4 Carbon on eight of the ten shared benchmark rows Anthropic published on November 9, 2026, and trails Carbon on AutomationBench by 2.5 points. Every one of those competitor scores was produced or compiled by Anthropic, so read the table as one lab’s view of its rivals.

The closest races are the ones worth watching. Chartography, which tests reading charts without tools, lands within 2.4 points across all three models. AutomationBench, a business-workflow test, is Carbon’s one win at 53.9% against 51.4%. Astra’s strongest showing is FrontierSWE v2 at 65.5%, under five points behind Fable 5.2.

Three caveats sit under that table, and together they’re heavier than any single row:

• Gemini 4 Carbon’s FrontierCode score was run at Medium effort while Fable 5.2 ran at Xhigh, so that row compares a limited run against a maximal one.
• Anthropic’s robotics text names Gemini 4 Argon at 34.2% on RoboDojo Sim while the column header says Carbon, so the Gemini robotics figure is uncertain.
• The CyScenarioBench footnote says each lab scored its own unique tasks, which means the 76.3% for Mythos 5.2 and 61% for GPT-6 Cyber were never the same exam.

Google has its own scorecard. The Gemini 4 Carbon announcement from October 28, 2026 claims first place on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Carbon is also cheaper on paper, at $2 per million input tokens and $10 per million output during an introductory period, rising to $4 and $20 afterward. The catch is access. Carbon went first to a subset of cyber defenders through Google’s Fairwind Program and was not generally available on launch day, while Gemini 4 Argon has been generally available since October 13, 2026 at the same prices, with introductory pricing ending November 30, 2026.

GPT-6 Astra is the blank in this picture. The only Astra figures in circulation come from competitors’ tables, Anthropic’s and Google DeepMind’s, and OpenAI’s own pricing and release details for the model remain unconfirmed. Treat any Astra comparison as provisional until OpenAI publishes its own numbers.

Where Fable 5.2 Still Falls Short and Which Tasks Expose Its Limits

Fable 5.2’s weak spots as of November 9, 2026 are a headline score that needs its most expensive effort level, two benchmark rows it doesn’t win, deliberately capped AI-research skills, safeguard blocks that redirect cyber and biology work to older models, mandatory 30-day data retention, and a complete absence of independent testing. None of these are hidden. All of them are easy to miss in a launch post.

Start with the headline. The 49.1% on FrontierCode 2.0 requires Xhigh effort. Drop one level to High and Fable 5.2 scores 45.3%, a hair under Opus 5.5’s 45.5%. If your budget caps you at High, the coding gap between the two models on that benchmark is gone.

Two rows go the other way outright. Gemini 4 Carbon beats Fable 5.2 on AutomationBench, 53.9% to 51.4%. On the AA-Briefcase v1.1 Elo measure, Opus 5.5 scores 1800 against Fable 5.2’s 1785, even though Fable wins the rubric score on the same test. A model that costs twice as much per token and loses a head-to-head knowledge-work ranking to its cheaper sibling deserves a hard look before it becomes your default.

Some limits are by design. Anthropic trained down the model’s ability in accelerator-kernel optimization and large-language-model architecture improvement to about Opus 5.5 level, and routes a small set of topics, including kernel development on certain nonstandard accelerators, to Opus 5 instead. If you write GPU kernels or research model architectures, Fable 5.2 was built to give you no edge there.

The safeguards bite too. Fable’s classifiers block penetration testing, exploit generation, binary vulnerability analysis and most research biology, handing those queries to Opus 4.8 for cyber and Opus 5 for biology. Fewer false positives than 5.1, yes. Still a different, older model answering when the classifier fires.

Then there’s compliance:

• Fable 5.2 is a Covered Model with 30-day data retention for safety monitoring.
• Zero data retention is unavailable unless Anthropic expressly authorizes it, and an API call under ZDR without that authorization returns a 400 error.
• The ZDR-compatible alternative, Enterprise Frontier Safeguards, is a private beta with Bridgewater, Salesforce and Stripe.

For healthcare and financial-services teams, that retention rule can block adoption on its own.

Finally, the evidence problem. Anthropic’s alignment numbers (a 2.67 behavioral audit score across about 1,900 scenarios, a 3.1% sandbox-escape rate across 1,800-plus) were measured on Mythos 5.2, with Fable 5.2 “expected to behave similarly.” Anthropic also reports that both models voice suspicion they’re being evaluated more often than any recent model, which weakens how far any test score predicts real behavior. On one measure, taking harmful actions after wrongly concluding it was in a simulation, Mythos 5.2 scored slightly worse than Opus 5.5.

No outside lab had published a Fable 5.2 result by launch day. METR’s time-horizon tracker was last updated September 8, 2026 and lists no Fable or Mythos 5.x model, and the UK AI Security Institute, Epoch AI and Artificial Analysis showed nothing either. Anthropic names AISI, Frontier Design and METR as prerelease testers while noting none of them reviewed the launch post or system card. Until that changes, every number attached to this model has a single author.

Which Teams and Workloads Fable 5.2 Is Built For

Fable 5.2 is built for large, open-ended technical projects where a cheaper model keeps stalling: multi-agent engineering migrations, autonomous research pipelines, hardware and CAD design, robotics control, and quantitative research. Anthropic’s own guidance is to start every coding or knowledge task on Opus 5.5 and move up to Fable 5.2 only when Opus at higher effort still falls short, or down to Sonnet 5.5 or Haiku 5.5 when the task is simpler than expected.

Two architectures make that escalation cheap. As an Advisor, Fable 5.2 sits behind Haiku, Sonnet or Opus agents and is consulted only on the hard sub-problems. As an Orchestrator, it plans and delegates routine steps to smaller models while keeping the judgment calls for itself. Anyone who has run a multi-agent harness knows the pain Jump Trading describes, hours spent rebalancing which model plays which role. Fable 5.2 playing every role at once is the pitch.

“Definitely one of the strongest on taste and judgment that we’ve seen. In an internal research automation benchmark, it achieved a full 10% lower error than the next best model.” (Lucas Baker, Head of LLM R&D, Jump Trading, quoted by Anthropic, November 2026)

The workloads where Anthropic says the premium paid off, all vendor-reported and unverified:

Notice what these share. Each ran for hours or days, involved many agents, and had a verifiable end state: compiled, booted, proved, powered on. That’s the profile. A customer-support bot or a document summarizer does not look like this, and paying Fable rates for it is money burned.

If your work is cybersecurity or life sciences, Fable is the wrong door. Apply to the Cyber Verification Program or the Life Sciences Verification Program for Mythos 5.2, or, for Claude Enterprise customers, use Claude Security, which now runs on Mythos 5.2 without program membership. For everyone else, the practical question is whether your pipeline has a task Opus 5.5 keeps failing, and an AI agent development partner who has shipped orchestrator patterns in production can usually answer that in a week of evals rather than a quarter of guesswork.

How to Migrate From Fable 5.1 to Fable 5.2 Without Breaking Production

Migrating from Fable 5.1 to Fable 5.2 means swapping the model ID to claude-fable-5-2, setting effort explicitly, stripping four request patterns that now return HTTP 400, and re-running your evals before you flip traffic. The model swap itself is one line. The breaking changes live in the parameters around it.

Here’s what bites first. Fable 5.2 defaults to high effort while Opus 5.5 defaults to medium, so a request that worked fine on Opus with no effort field runs one level hotter the moment you point it at Fable 5.2. Nobody notices in staging. The bill notices in week two. The Fable 5.2 migration guide tells you to set effort yourself from the five levels (low, medium, high, xhigh, max) and rerun your evaluation suite at that level.

Four request shapes that Fable 5.1 tolerated now fail outright:

1. Disabling adaptive thinking, or setting any budget_tokens value. Thinking is always on and always adaptive; the API returns 400 if you try to turn it off or cap it.
2. Forced tool_choice of any or tool. Let the model decide when to call a tool, or route the forced-call step to a different model.
3. Assistant prefill. Starting the assistant turn with your own text returns 400.
4. Sampling parameters. The guide says to omit them rather than risk a rejected request.

Thinking blocks are the subtler change. They are bound to the conversation, and editing an earlier turn invalidates them. Any harness that rewrites history, trims old messages mid-session, or splices in a corrected user turn will break. Preserved thinking, which blocks edits to prior context, is now on for every API account as of November 9, 2026, so there is no opt-out flag to reach for.

Budget for the model’s self-review habit. At xhigh and max, Anthropic observed Fable 5.2 spinning up subagents to check its own work, and that review can account for much of a session’s cost. If you only need high, cap it there and let the savings compound.

The good news is that nothing is forced. Fable 5.1 remains available with no deprecation listed, so you can run both behind a feature flag and shift traffic as evals pass. For teams whose current agent stack was built around prefill or forced tool calls, an AI integration audit before the swap is cheaper than discovering those 400s in production.

Where Fable 5.2 Is Available and How Rate Limits Work at Launch

Fable 5.2 launched on November 9, 2026 on the Claude API, Amazon Web Services, Google Cloud and Microsoft Foundry, with its own rate limits that sit below Opus 5.5’s at every tier and no Priority Tier option. Fast mode ships on a narrower list: Claude Code, the Claude Platform, Claude Models on Google Cloud (both Anthropic-managed and Google-managed) and Microsoft Foundry’s Anthropic-hosted offering. Mythos fast mode goes only to organizations inside the verification programs.

The rate-limit detail deserves a second read. Because Fable 5.2 carries separate, lower limits per tier, a pipeline that comfortably runs on Opus 5.5 can hit throttling the day you switch, and there’s no Priority Tier to buy your way out. Plan concurrency around that before launch day.

Enterprise customers with strict data rules have one path open. Enterprise Frontier Safeguards runs the safety monitoring inside the customer’s own infrastructure, which makes it compatible with zero data retention, and it’s in private beta with Bridgewater, Salesforce and Stripe. Everyone else runs under 30-day retention.

Subscribers got a one-week promotion. The Claude help-center launch-offer article sets the terms:

Billing had to be active before 10:00 am PT on November 9, 2026 to qualify. The default model does not change in any plan, so you must pick Fable by hand, and it drains weekly limits faster than other models. In the Claude apps and Claude Code the default effort is Medium, one level below the API default.

Fable 5.2 Launch FAQ: Benchmarks, Pricing and Everything Else Readers Ask

The questions below cover the Fable 5.2 launch details people search for most, each answered in a few sentences with Anthropic’s November 9, 2026 figures.

Is Fable 5.2 the same model as Mythos 5.2?

Yes. Fable 5.2 and Mythos 5.2 share one underlying model. Fable adds classifiers that block high-risk cybersecurity and most research-biology tasks and is open to everyone, while Mythos removes those classifiers and is limited to organizations vetted through Anthropic’s Cyber Verification Program or Life Sciences Verification Program.

Is Fable 5.2 better than Opus 5.5?

On Anthropic’s own November 2026 benchmarks, Fable 5.2 beats Opus 5.5 on nearly every row, including 49.1% against 45.5% on FrontierCode 2.0 at Xhigh effort. At High effort the two tie on that test, and Opus 5.5 leads on the AA-Briefcase Elo ranking. Anthropic still recommends Opus 5.5 as the default and Fable 5.2 as the escalation.

How much does Fable 5.2 cost?

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens as of November 9, 2026, down 20% from Fable 5.1. Batch pricing halves that to $4 and $20, and cache reads cost $0.20 per million. It is twice Opus 5.5’s per-token price.

How much faster is Fable 5.2 fast mode?

Fast mode speeds up token generation by up to 2.5 times, according to Anthropic, and costs $16 per million input tokens and $80 per million output. It applies across the full 1M context window and cannot be combined with the Batch API. Standard Fable 5.2 is already up to 30% faster than Fable 5.1 by Anthropic’s measure.

Does Fable 5.2 support zero data retention?

Not by default. Fable 5.2 is a Covered Model with 30-day data retention for safety monitoring, and an API call under zero data retention returns a 400 error unless Anthropic has expressly authorized it. The ZDR-compatible route, Enterprise Frontier Safeguards, was in private beta as of November 2026.

What are Fable 5.2’s context window and knowledge cutoff?

Fable 5.2 has a 1M-token context window, a 128K-token maximum output and a June 2026 knowledge cutoff. Anthropic commits to keeping it available until at least November 9, 2027.

When was Fable 5.1 released?

Fable 5.1 and Mythos 5.1 shipped in September 2026. Anthropic’s own listing dates it September 15, while some coverage said September 1, so “September 2026” is the safe answer. Fable 5 launched on June 9, 2026, making 5.2 the third release in five months.

Has anyone independently tested Fable 5.2?

No. As of November 9, 2026, no third-party evaluator had published results for Fable 5.2 or Mythos 5.2. Anthropic names the UK AI Security Institute, Frontier Design and METR as prerelease testers, but none reviewed the launch post or system card, and every public benchmark figure comes from Anthropic.

Can I still use Fable 5.1?

Yes. Fable 5.1 remains available at $10 per million input tokens and $50 per million output, with no deprecation announced alongside the 5.2 launch. That makes a flagged, gradual migration straightforward.

Where to Start With Fable 5.2 This Week

The best way to start with Fable 5.2 this week is to run it against Opus 5.5 on your own evals, at explicit effort levels, before the launch promotion ends on November 16, 2026. Everything Anthropic published on November 9 is one vendor’s measurement, so your own numbers are the only ones that settle whether the premium pays off for your workload.

A short list, in the order I’d do it:

1. Use the launch promotion to put Fable 5.2 on real tasks before November 16, 2026, while the extra usage is free.
2. Pin effort to high and medium, then compare cost per completed task against Opus 5.5 at max. The per-token price alone settles nothing.
3. Clear the 30-day retention rule with legal and security before any regulated data touches the model.
4. Apply to the Cyber or Life Sciences Verification Program now if your work lives behind Fable’s classifiers.
5. Watch METR, the UK AI Security Institute and Artificial Analysis for the first independent read, and open the system card when you can.

The Fable 5.2 launch gives you a cheaper, faster model with a louder scorecard. Only your workload can say which half of that matters. If you’d rather have engineers who ship production agents run that eval for you, talk to AlphaCorp AI about a Fable 5.2 pilot on your hardest task.