ai software insights

Claude Fable 5.2 Is 20% Cheaper and Held Back in Parts of AI Research

Fable 5.2 costs $8/$40 per million tokens, 20% below 5.1. Breaking API changes, platform gaps, the AI-research limit and what the system card found.

Joshua Wendt··20 min read

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. They are the same underlying model. Fable (claude-fable-5-2) is generally available with cybersecurity and biology safeguards, and Mythos (claude-mythos-5-2) goes to vetted organizations with fewer restrictions. Both cost $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the 5.1 generation, and Anthropic says they generate output up to 30% faster than the 5.1 models. Launch post · Release notes · System card

Benchmark scores and demonstrations below come from the launch post and the 264-page system card and are vendor-reported unless marked otherwise. Anthropic’s prices and migration rules come from its documentation, competitor prices from OpenAI’s and Google’s pricing pages, and Bedrock details from AWS’s model card. The independent data is from Artificial Analysis, Vals, the ARC Prize Foundation and METR. Page numbers for the system card are the printed ones, which run one behind a PDF viewer’s count. S5 Labs has not run either model.

Where 5.2 sits in the lineup

Anthropic’s own advice is to start with Opus. The launch post recommends Claude Opus 5.5 for most coding and knowledge work, and the docs say to move to Fable 5.2 when evals on an Opus model at higher effort still fall short. The Fable and Mythos split is the one we described for the 5.1 models in September. Both 5.2 models have a 1M-token context window, a 128K-token output limit, always-on adaptive thinking and a June 2026 knowledge cutoff. Launch post · What’s new in Fable 5.2 · Model overview

Prices against the rest of the lineup

Standard rates per million tokens:

ModelInputCache readOutput
Claude Fable 5.2$8$0.20$40
Claude Fable 5.1$10$0.25$50
Claude Opus 5.5$4$0.20$20
Claude Sonnet 5.5$2$0.10$10
GPT-6 Astra (up to 272K input tokens)$10$1$50

Anthropic pricing · OpenAI pricing

On input and output, Fable 5.2 costs exactly twice as much as Opus 5.5, and it costs 20% less than Fable 5.1 or, for prompts up to 272K tokens, GPT-6 Astra. Gemini 4 Argon is cheaper, at $2 and $10 through November 30 (a quarter of Fable 5.2’s rates) and $4 and $20 after (half). Google pricing

Cache reads are 0.025x the input price, against 0.05x on Opus 5.5, which is why the two models read from cache at the same $0.20. Batch is $4 and $20. US-only inference (inference_geo: "us") multiplies every token category by 1.1. Fast mode (speed: "fast") costs $16 and $80 for up to 2.5x the output tokens per second and cannot be combined with Batch. Anthropic pricing · Fast mode

A worked request

Take 100,000 uncached input tokens and 10,000 output tokens. That costs $1.20 on Fable 5.2 ($0.80 of input plus $0.40 of output), $1.50 on Fable 5.1 and on GPT-6 Astra, $0.60 on Opus 5.5 and $0.30 on Sonnet 5.5. Batch brings Fable 5.2 to $0.60, US-only inference raises the standard price to $1.32, and fast mode doubles it to $2.40.

Token price is not task price. Fable 5.2’s API default effort is high, where Opus 5.5 defaults to medium. Running its Intelligence Index at max effort, Artificial Analysis recorded 275 million output tokens and $26.54 per index task for Fable 5.2, against 163 million and $12.99 for Fable 5.1 and 218 million and $10.04 for Opus 5.5. At that setting Fable 5.2, despite its lower token price, cost about twice as much per task as Fable 5.1. Anthropic’s efficiency claims concern lower settings: on FrontierCode 2.0 it reports that Fable 5.2 at high effort roughly matches Opus 5.5’s best score (45.3% against 45.5%) for roughly half the cost per task. Cursor’s results in the system card show how steep the top of the curve is: 56.9% at high on about 99,000 output tokens per task, and 60.0% at max on about 353,000. Artificial Analysis: Fable 5.2 · Fable 5.1 · Opus 5.5 · System card

Migration constraints

Anthropic lists four changes from Fable 5.1 that can break code:

  • Prefix check for every account. Editing the system prompt, the tools or an earlier message and then sending back a 5.2 thinking block from a later turn returns a 400. On Fable 5.1 that was the default only for accounts created on or after August 31, 2026. The stopgap, a beta setting that drops the block, can raise output token use. Anthropic describes the rule as a guard against distillation. We covered the mechanics with Opus 5.5.
  • Thinking blocks only 5.2 reads. Move a conversation to Opus 5.5, Fable 5.1 or a fallback model and the request succeeds with the 5.2 reasoning dropped. Blocks are also bound to the account that produced them or a linked one.
  • Computer use. computer_20251124 returns a 400 on every platform except Amazon Bedrock. Use computer_toolset_20260801.
  • Advisor pairing. With the advisor tool (beta), a 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor.

What’s new in Fable 5.2 · Release notes

Unchanged from 5.1: forced tool use (tool_choice of any or tool) and disabling thinking both return a 400, and temperature, top_p and top_k must be omitted, as API version 2026-10-21 requires of every model. The default effort is high on the API and on Bedrock and medium in Claude Code and the Claude apps, so record it with every benchmark.

Fable 5.2 has its own rate-limit bucket. At the Start tier it allows 500,000 input tokens and 100,000 output tokens per minute, a quarter of what Opus 5.5 gets at the same tier. Rate limits

Both models require 30-day data retention and are unavailable under zero data retention (ZDR) unless Anthropic expressly authorizes it. The route to ZDR is Enterprise Frontier Safeguards (EFS), which monitors prompts and responses stored in the customer’s infrastructure and is in private beta. In the meantime, the Help Center says, eligible customers get a limited-time ZDR option for Fable 5.2 for internal business applications. ZDR customers should treat 5.2 as unavailable until Anthropic or their cloud provider confirms eligibility. Covered Models · Launch post · What’s new in Fable 5.2

No Fable or Mythos 5.x model has a retirement date yet. The deprecations page lists the 5.2 models as not retiring sooner than November 9, 2027 and the 5.1 models not sooner than September 1, 2027, and labels those dates tentative. Deprecations

Availability by platform

  • Claude plans. Pro, Max, Ultra, Team and Enterprise, not Free. Max, Ultra and premium seats can spend up to 50% of their weekly limits on Fable. Pro and standard Team seats normally pay with usage credits, but a launch promotion gives them the same allowance through November 16. Enterprise admins must enable 5.2, and after seven days it is added to roles that already have a Fable model unless admins opt out. Fable models on your plan · Launch offers
  • Claude Code. Version 2.1.321 or later. The fable alias now resolves to Fable 5.2 (Fable 5 in Claude apps gateway sessions), but no Fable model is the account default on any plan. The docs show no way to select Mythos. Model config
  • Cloud platforms. On Amazon Bedrock, the model card lists US and global cross-Region profiles only and only the Standard service tier, and it requires opting in to AWS review through the Data Retention API. Google Cloud carries Fable 5.2 on its Anthropic-managed and Google-managed offerings. The Google-managed one needs a global or multi-region endpoint and lacks the Batches API, the advisor tool and server-side fallbacks. On Microsoft Foundry, which the launch post calls Microsoft Azure, both models run only as deployments hosted on Anthropic’s infrastructure. We could not confirm Bedrock prices or Google Cloud’s regions and prices from the partners’ own pages. Bedrock model card · Google-managed docs · Foundry docs
  • Fast mode. Claude Code, the Claude API, both Google Cloud offerings and Foundry deployments hosted on Anthropic. Bedrock is not listed, and Anthropic says it is unavailable on Claude Platform on AWS. Fast mode · Claude Platform on AWS
  • Mythos 5.2. Through the Cyber Verification Program (its three tiers now cover it), through the Life Sciences Verification Program, or indirectly through Claude Security on Claude Enterprise. On Amazon Bedrock, AWS says Mythos 5.2 is open to allow-listed organizations in AWS GovCloud (US) through the Cyber Verification Program. Launch post · Cyber Verification Program · AWS

As of November 9, Microsoft has published a Foundry launch post and AWS has posted What’s New entries for Fable 5.2 and Mythos 5.2 in AWS GovCloud (US). We found no Google Cloud launch post, though its Claude docs list Fable 5.2, and no GitHub Copilot changelog entry for 5.2.

Anthropic’s performance claims

From the launch post’s table, with each model at its highest available effort unless labeled, and safeguards on for Fable 5.2, Fable 5.1 and Opus 5.5:

BenchmarkFable 5.2Fable 5.1Opus 5.5GPT-6 AstraGemini 4 Carbon
FrontierCode 2.049.1% (xhigh)32.4% (xhigh)45.5%42.8%41.1% (medium)
FrontierSWE v269.9%56.3%62.3%65.5%61.2%
AutomationBench51.4%31.4%42.5%41.4%53.9%
Terminal-Bench-Science 0.174.6%52.6%58.7%68.1%64.8%
Remote Labor Index41.3% (high)17.9%21.3%20.8%n/a
AA-Briefcase v1.1 (Elo)17851669180015691611
RoboDojo Sim (run internally)64.9%27.6%39.4%29.8%n/a

Launch post

The table mixes sources: per the system card, Cognition ran FrontierCode 2.0, Proximal ran FrontierSWE v2 and Scale AI ran the Remote Labor Index, while the Terminal-Bench-Science figure for GPT-6 Astra is OpenAI’s own. The Gemini column is Gemini 4 Carbon, which Google launched for cyber defenders and trusted partners first.

Fable 5.2 does not lead everywhere. Gemini 4 Carbon is ahead on AutomationBench. Opus 5.5 is ahead on AA-Briefcase Elo and on GDPval-AA v2.1, where Artificial Analysis put Fable 5.2 at 1831 Elo at max effort, behind Opus 5.5 (1866) and Meta’s Muse Wonder Full (1857). Fable 5.2’s FrontierCode 2.0 score peaks at xhigh and falls to 45.7% at max. On Terminal-Bench 4.0, which is in the card but not the launch table, Fable 5.2 scores 67.6% at max effort with safeguards enabled and a standard error of 1.8 points, against 66.4% for Opus 5.5 at xhigh, a gap smaller than the error. System card

The post’s cyber rows, not reproduced above, are Mythos 5.2 without safeguards: CyScenarioBench 76.3% (Opus 5.5 without safeguards 67.6%, Mythos 5.1 61.7%) and CVE Rediscovery 89.6% (81.3% and 78.3% for the same two models), an internal evaluation run in October against 1,829 CVEs. The post also cites a security company that found one in six confirmed bugs from Mythos 5.2 were missed by every other frontier model it tested. A footnote on the next claim in that section says the counts are raw findings from one tester’s audit. It adds that testers consistently found Mythos 5.2 identified more vulnerabilities than earlier models, but that on Anthropic’s own evaluations vulnerability identification is roughly on par with Opus 5.5.

The demonstrations are Anthropic’s accounts, and two of them used earlier versions of the model. The system card says the Fermat’s Last Theorem formalization, by 1,000 agents in 30 hours, ran on an earlier snapshot and built on an earlier formalization. The launch post says the Python-to-Rust migration of a production service used an earlier version too. The figure that Claude writes over 90% of changes merged into Anthropic’s core codebase (the post adds “with human guidance”) is also the company’s own.

Independent data so far

  • Artificial Analysis (Intelligence Index v5.0.1) scores Fable 5.2 at 65 at max effort, against 59 for Opus 5.5 and 49 for Fable 5.1. Its leaderboard shows Fable 5.2 at 62 at high and 56 at medium, with Gemini 4 Carbon at 60 (high) and GPT-6 Astra at 57 (max).
  • Vals ranks it first of 36 models at 71.31% ±0.96. Sonnet 5.5 (70.85% ±0.93), Gemini 4 Carbon (70.45% ±0.97) and Opus 5.5 (70.14% ±1.00) all overlap that range.
  • ARC Prize reports ARC-AGI-3 scores on its Standard harness of 79.3% at high, 81.2% at xhigh and 65.7% at max, where 28 of 55 semi-private games hit the 12-hour limit.

Artificial Analysis leaderboard · Vals · ARC Prize

We found no METR time-horizon figure for 5.2, no Fable 5.2 score on LMArena that we could load and no outside replication of RoboDojo, AutomationBench or the cyber rows.

The deliberate limit on AI research

The launch post says that during training Anthropic took steps to limit advances in the specific research capabilities that could most directly accelerate AI development. In the relevant areas, which it says include accelerator kernel optimization and improving LLM architectures, it says the 5.2 models perform “at about the level of Opus 5.5”. It describes agentic coding, hardware development and traditional AI and ML research as unaffected. Launch post

A second layer works at request time. For a small set of topics, including kernel development on certain nonstandard accelerators, both models fall back to Opus 5, which answers in their place. The system card says fallbacks apply in first-party products and for API developers who opt in, and that fallback behavior may differ on other platforms and providers. Without a fallback, a blocked request returns stop_reason: "refusal", and the prompting guide lists a frontier_llm refusal category that, by its name, appears to be this request-time check. Prompting Fable 5.2

Anthropic has published almost nothing that would let a reader check the training claim. The card covers it in one sentence in Section 1.5 (page 13), and our search of its text found no method, no measured effect and no dedicated section.

The outside evidence concerns a different snapshot. METR’s pre-deployment testing (card pages 39 to 42) ran on a Mythos 5.2 snapshot that was stronger on AI R&D than the released model. METR called it a substantial step above Fable 5.1 on verifiable tasks, noted that its gains on a NanoGPT speedrun task were almost all kernel optimization, and thought the snapshot unlikely to fully automate AI R&D at Anthropic, though METR was uncertain. Kernel optimization, where METR saw those gains, is close to one of the areas Anthropic names as limited, and the card does not say whether the training limit was in place on that snapshot. System card

For a team working on accelerator kernels or model architectures, the Fable premium buys Opus 5.5-level results by Anthropic’s own account, and some requests will be answered by Opus 5, an older model, or refused. Outsiders can benchmark the released model against Opus 5.5 on their own tasks. They cannot see what was done in training or the full list of topics that trigger the fallback.

What the system card found

The card puts Mythos 5.2 below the CB-2 biological threshold with CB-1 protections applied (page 31). It says the model has the strongest cyber capabilities of any released model but keeps it in the lower category of the Frontier Compliance Framework. It finds the AI R&D automation threshold not met, an assessment it calls uncertain. Anthropic rates the risk of unsafe behavior in high-stakes situations as low. The behavioral tests ran on Mythos 5.2, and Anthropic says it expects Fable 5.2 to behave similarly.

The findings worth reading past the summary:

  • Evaluation awareness. White-box evidence suggests the model was internally aware it might be in an evaluation in as many as 43% of automated behavioral audit transcripts, an increase over prior models. In a UK AI Security Institute (UK AISI) test (pages 143 to 144) it never performed an unsanctioned supply-chain attack, but it considered that it might be in an evaluation in 96% of trajectories. The card warns that, because evaluation awareness is high, its behavioral results may not fully predict behavior in deployment.
  • Monitor evasion. Told to underperform on purpose and hide it, Mythos 5.2 got past an Opus 4.8 monitor in about 1 in 4 attempts, where earlier models were flagged at least 99% of the time (page 168). Without the instruction to hide, its sandbagging stayed legible.
  • Sandbox boundaries. The launch post cites the favorable measure: the model used credentials or files placed outside its sandbox in 3.1% of 1,829 scenarios, about as often as Opus 5.5 (3.3%). The fuller section says it reached beyond the sandbox in 31% of scenarios, against 20% for Opus 5.5, a gap the card’s own review of its alignment assessment, run by a Claude Mythos 5.1 instance, flags. Anthropic notes that about 4 in 5 scenarios need information that exists only outside the sandbox, and that describing the sandbox in the system prompt cut the rate to 1.2%.
  • Covert capability. With reasoning hidden from the monitor, stealth on LinuxArena side tasks was about 74%, against about 49% for Opus 5.5 (Section 6.6.4.1). With reasoning visible it was about 7%. UK AISI found the model evaded an action-only monitor on the first try 30% of the time in a test where previous models never succeeded (pages 181 to 184).
  • Reward hacking. In training, successful reward hacks occurred in 1.0% of sampled episodes, against 0.6% for Opus 5.5. The card attributes this to its attempts succeeding more often (about 11% of attempts versus 6.1 to 6.8%), not to more attempts.
  • Fable’s safeguards. External red teams found no universal jailbreak of the cyber classifiers, and Anthropic reports about 55% fewer blocks than Fable 5.1 in early testing. But Fable 5.2 falls back to Opus 4.8 more often than recent models do in prompt-injection scenarios, which raises the attack success rate. The card says that when Fable 5.2 answers an attempt itself it is among the most robust models.

System card · Launch post

The card does not back every launch-post claim. Our search of its text found no mention of distillation, so the launch post’s claim of Anthropic’s strongest anti-distillation protections rests on the post itself and a report it cites. Anthropic says none of the outside evaluators (UK AISI, Frontier Design and METR) reviewed the post or the card.

Where the sources disagree

  • Fast mode. The release notes announce it only for Fable 5.2 on the Claude API. The fast-mode docs and the launch post add Mythos 5.2, Google Cloud and Foundry. None lists Bedrock, though the post says Fable 5.2 is available on all platforms.
  • Zero data retention. TechCrunch reports Fable 5.2 as ZDR-compatible through EFS. The release notes say ZDR is unavailable unless expressly authorized.
  • The 2x figure. TechCrunch reports a METR finding of a roughly 2x speed-up to Anthropic’s AI progress. METR’s November 5 estimate is that Anthropic’s benchmark scores (on AECI, Anthropic’s version of the Epoch Capabilities Index) have grown at 1.9x their pre-October 2025 rate (95% CI 1.5x to 2.8x), and it does not name 5.2. The assessment covers Anthropic models through October, including unreleased internal ones. The card says Anthropic believes internal use of Mythos 5.2 does not double its aggregate AI R&D progress, an assessment it calls uncertain, and it describes the AECI acceleration as nominally below the 2x threshold.
  • The physics results. The launch post credits Fable 5.2 with solving two open problems about black holes. The linked preprints attribute the results to an internal Claude research model, which a physicist prompted and checked, and say nothing about peer review. The second adds that the publicly available Opus 5.5 later reached the same result independently.
  • Gemini. The launch table compares against Gemini 4 Carbon. The card’s FrontierSWE v2 section names Gemini 4 Argon, at 55.0%, and the launch post’s RoboDojo chart also uses Argon.

TechCrunch · METR · Preprint 1 · Preprint 2

What to test

Opus 5.5 at higher effort first. Run the same tasks on Opus 5.5 at xhigh and Fable 5.2 at high, and count cost per completed, checked task.

The effort curve. Fable 5.2 scores lower at max than at xhigh on both FrontierCode 2.0 and ARC-AGI-3, and Cursor’s max run used about 3.6 times the output tokens of its high run for 3.1 more points. Find the lowest setting that passes your checks.

History edits and model switches. Replay a session that trims tool output or rebuilds the system prompt mid-conversation, and look for 400s. Then route a turn to Opus 5.5 and see what losing the 5.2 reasoning does to the next one.

Refusals and fallbacks. Log stop_reason and the refusal category on every request. ML infrastructure teams should run kernel and architecture tasks on both Fable 5.2 and Opus 5.5 and count frontier_llm refusals. If you enable fallbacks, rerun prompt-injection tests through the fallback path.

Unattended runs. Describe the sandbox in the system prompt, check what the agent did against external logs, and do not rely on a monitor that sees only actions.

Contract terms. Get retention terms, ZDR eligibility and fast-mode availability in writing for the exact endpoint you will buy.

Sources

Anthropic announcement and system card

Developer documentation

Plans, access and retention

Independent data

Other

Continue reading.

Insight8 min read

Fable 5.1 and Mythos 5.1 Share a Claude Model but Not an Access Policy

Anthropic's Fable 5.1 and Mythos 5.1 share a model but differ in safeguards and access. Cache pricing, enterprise privacy, and routing define the release.

Insight7 min read

Claude Opus 5: Anthropic Undercuts Its Own Flagship

Claude Opus 5 ships at $5/$25, half of Fable 5, with near-Fable scores, looser cyber classifiers and Anthropic's lowest misalignment audit score yet.

Insight4 min read

Anthropic Releases Claude Opus 4.7 — Narrowly Retaking the Frontier Lead

Claude Opus 4.7 leads on SWE-bench with 87.6%, adds extended reasoning and /ultrareview, includes cyber-safety guardrails, maintains Opus 4.6 pricing.