KOContact
IT

Claude Fable 5.2 Is Here: What Changed, Benchmarks, and Pricing

Claude Fable 5.2 launched November 9, 2026: what changed from Fable 5.1, benchmarks vs Opus 5.5, lower API pricing, and developer notes from official docs.

12 min read
Claude Fable 5.2 Is Here: What Changed, Benchmarks, and Pricing

Anthropic released Claude Fable 5.2 on November 9, 2026. Compared with Fable 5.1, it generates output up to 30% faster, its API price is 20% lower, and it beat Claude Opus 5.5 on most of the benchmarks the company published. This post draws only on Anthropic's launch announcement, developer docs, and Help Center articles, not on other blogs or news coverage. Everything here reflects those pages as of November 12, 2026.

A processor chip at the center of a dark green circuit board, surrounded by densely packed small components

Key Takeaways

  • Release: November 9, 2026. Claude Mythos 5.2, which runs on the same model, launched alongside it, but Mythos is only available to organizations that pass one of Anthropic's verification programs.
  • Speed and price: Output is up to 30% faster than Fable 5.1, and API pricing dropped from $10 to $8 per million input tokens and from $50 to $40 per million output tokens.
  • Performance: It scored higher than Opus 5.5 on 11 of the 12 benchmark metrics included in this post. Keep in mind that these are Anthropic's own measurements, and most come from the highest effort setting.
  • Fewer unnecessary blocks: Anthropic says blocks dropped by about 55% compared with Fable 5.1 while keeping the same level of cyber and biology safeguards.
  • Watch out, developers: The API has four changes that can break existing code, including a conversation history modification check (prefix check) that now applies to every account.

What Is Claude Fable 5.2?

Claude Fable 5.2 is the model Anthropic positions one tier above Opus 5.5, built for demanding reasoning and long-running agentic work. The developer docs point to long agent sessions lasting hours, research that takes many steps, and analysis that ends in a deliverable such as a report, spreadsheet, or slide deck.

Claude Mythos 5.2, released at the same time, is the same model as Fable 5.2. What separates them is the safeguards. Fable comes with classifiers that block misuse in cybersecurity and biology, so anyone can use it without extra verification. Mythos has looser restrictions, but only organizations that have passed the Cyber Verification Program or the Life Sciences Verification Program can use it.

Anthropic recommends starting most coding and knowledge work with Claude Opus 5.5. According to the developer docs, you should move up to Fable when a task is so large and complex that raising Opus's effort still doesn't get you the result you need.

What Changed from Fable 5.1

ItemFable 5.1Fable 5.2
API price (input/output per million tokens)$10 / $50$8 / $40
Cache reads (per million tokens)$0.25$0.20
Output generation speedBaselineUp to 30% faster
Fast modeNo comparison in launch materialsUp to 2.5x generation speed, $16 / $80
Safety classifier blocksBaselineAbout 55% fewer (cyber about 60%, biology about 40%)

These figures come from Anthropic's launch announcement and developer docs. Anthropic says the block reduction figures come from early testing in claude.ai, Claude Code, and Cowork. Cache read prices are calculated from the docs' statement that, for both models, they are 0.025x the base input price. Fable 5.2 has a 1 million token context window and a maximum output of 128,000 tokens.

Beyond pricing, three things changed.

  • Communication: According to Anthropic, early testers found its explanations clearer than those of previous Fable and Mythos models.
  • Token efficiency: Anthropic says that at medium effort (what the Claude apps and Claude Code use by default) it needs fewer tokens than Fable 5.1 to finish difficult tasks.
  • Distillation defenses: Anthropic says it has the strongest defenses of any of its models so far against distillation, where someone harvests Claude's outputs at scale to train another model.

Fable 5.2 Benchmarks Compared with Opus 5.5

The table below picks the main rows from the table Anthropic published in its announcement. According to the company, the Fable 5.2, Fable 5.1, and Opus 5.5 scores were measured with safeguards turned on. Scores without a label come from the highest effort each model supports; a label in parentheses means the score was measured at a different effort.

Area (benchmark)Fable 5.2Fable 5.1Opus 5.5GPT-6 AstraGemini 4 Carbon
Agentic coding (FrontierCode 2.0)49.1% (xhigh)32.4% (xhigh)45.5%42.8%41.1% (medium)
Agentic coding (FrontierSWE v2)69.9%56.3%62.3%65.5%61.2%
Business workflows (AutomationBench)51.4%31.4%42.5%41.4%53.9%
Agentic scientific research (Terminal-Bench-Science 0.1)74.6%52.6%58.7%68.1%64.8%
Agentic search (WANDR, relaxed score)82.1%68.7%72.3%--
Computer use (OSWorld 2.1, offline set, partial)89.3%83.5%87.2%79.5%81.5%
Robot control (RoboDojo Sim, internal measurement)64.9%27.6%39.4%29.8%-
Knowledge work (AA-Briefcase v1.1, rubric)66.1%61.5%61.2%52.0%61.8%
Knowledge work (AA-Briefcase v1.1, Elo)17851669180015691611
Knowledge work (Remote Labor Index)41.3% (high)17.9%21.3%20.8%-
Multidisciplinary reasoning (Humanity's Last Exam, with tools)74.2%65.6%67.7%57.2%-
Chart reading (Chartography, no tools)73.4%44.8%64.4%71.0%72.5%

A few things stand out in the table.

  • Coding: FrontierCode 2.0 measures whether a model produces code changes that open-source repository maintainers would accept. Anthropic says Fable 5.2's 49.1% at xhigh was the best result among the publicly available models it tested.
  • Robotics and the Remote Labor Index: Measured as a multiple of Fable 5.1's score, these areas improved the most. RoboDojo Sim jumped from 27.6% to 64.9%, and Remote Labor Index from 17.9% to 41.3%, both more than doubling.
  • It doesn't win everything: On business workflows (AutomationBench), Gemini 4 Carbon scored highest at 53.9%, and on the AA-Briefcase Elo score, Opus 5.5 (1800) edges out Fable 5.2 (1785).

Cost Efficiency by Effort Level

Anthropic stressed cost efficiency as much as raw scores. The table below summarizes what the announcement says about its accuracy-versus-cost charts.

BenchmarkFable 5.2 settingCompared withAnthropic's description
FrontierCode 2.0highOpus 5.5 best score (45.5%)Nearly the same score at 45.3%, at about half the cost per task
WANDRmediumOpus 5.5 best scoreSame score, at about 40% of the cost per task
Humanity's Last ExamlowOpus 5.5 best scoreSame score, at less than one-tenth the cost per task
ChartographymediumOpus 5.5 at max (64.4%)Higher at 66.4%, at less than one-tenth the cost per task

On BenchCAD 2.1 (with tools), a hardware design evaluation, Fable 5.2 reportedly scored 83.5 at max effort, well ahead of Opus 5.5 (65.3) and Fable 5.1 (44.4).

Anthropic's argument is that even though Fable is the pricier model, lowering its effort can get you the same results as Opus 5.5's top setting for less money. All the figures in these tables were measured and published by Anthropic itself, and the company says its evaluation methods are described in the system card. It's worth checking for yourself whether the same gap shows up on the tasks you actually run.

Examples from the Launch Announcement

Anthropic highlighted several things Fable 5.2 accomplished during early testing. Everything below is as reported by the company.

  • Operating system prototype: Starting from an empty folder, it directed more than 1,000 subagents over eight days to build an operating system that boots on a laptop. Anthropic says it included disk, keyboard, and network drivers, file storage, a windowed display, a login screen, and an installer.
  • Formalizing Fermat's Last Theorem: About 1,000 Fable 5.2 agents reportedly translated the proof into Lean, a language computers can use to verify proofs, in 30 hours. For comparison, Opus 5.5 in the same setup had completed 26% by the 18-hour mark and made almost no further progress by the 30-hour mark. When Anthropic updated the announcement on November 11, it added that the agents were given, up front, 47 lemmas already proven in Anthropic's own earlier formalization, a one-sentence sketch of the proof route, and other materials.
  • Electronics design: Working from a one-paragraph prompt with minimal human input, it designed a home voice assistant hub, a USB speakerphone, and a handheld virtual pet, including circuit boards, enclosures, and software. Anthropic says all three devices worked once built, with no changes to the circuits.
  • Internal service migration: Anthropic says agents running a pre-release build of Fable 5.2 rewrote one of its key production services, originally in Python, in Rust; the new version starts roughly 26 times faster.
  • Physics: Anthropic says it solved two open problems in theoretical physics related to black holes.

Changes Developers Need to Know

On the API, the model name is claude-fable-5-2, and it's available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Its knowledge cutoff is June 2026.

Rules That Apply to Every Request

  • Adaptive thinking is always on. Turning thinking off or setting budget_tokens returns a 400 error. You control how deeply it thinks with effort (low through max), and the API default is high.
  • Forced tool calls are not supported. Setting tool_choice to any or to a specific tool returns a 400 error. If you need JSON in a fixed format, use auto with strict: true, or use structured outputs.
  • 30-day data retention is required. Unless Anthropic grants a separate exception, it can't be used under zero data retention (ZDR) terms.

Four Things That Can Break When Migrating from Fable 5.1

  1. The conversation history modification check applies to every account. If you change system, tools, or earlier messages that come before a thinking block and then send that block again, you get a 400 error. On Fable 5.1, this check only applied to accounts created on or after August 31, 2026. If you use the model through Claude Code, claude.ai, Managed Agents, or the Agent SDK, you don't need to worry about it. If you wrote your own integration, make sure you keep system and tools unchanged for the whole conversation, only append new messages at the end, and send model responses back exactly as you received them. If you can't fix this right away, you can send the thinking-binding-controls-2026-08-01 beta header and set thinking.block_binding.prefix_mismatch_behavior to "drop_block" to drop just the offending block and continue, but the docs treat this as a stopgap.
  2. Fable 5.2's thinking blocks can only be read by Fable 5.2 and Mythos 5.2. If you switch to another model such as Opus 5.5 or Fable 5.1 mid-conversation, the earlier reasoning is left out for that part of the conversation. The request still succeeds, and you aren't billed for the dropped blocks.
  3. The older computer use tool (computer_20251124) returns an error outside Amazon Bedrock. You need to switch to computer_toolset_20260801.
  4. If Fable 5.2 is your executor model, the advisor tool's advisor model must also be Fable 5.2 or Mythos 5.2. Using Fable 5.1 as the advisor returns a 400 error.

Behavior Changes You May See in Outputs

The docs also cover behavior differences that show up even if you don't change any code.

  • In Anthropic's testing, at medium effort and above, the model kept thinking even when the prompt told it to answer directly without deliberating, and at high and above it thought before almost every response.
  • During long tasks, it sometimes ends its turn after leaving only a list of remaining work, and this happens more often at low and medium effort.
  • If the prompt doesn't say who owns the code, it tended to report nearby problems rather than fix them itself.

If your prompts were written for earlier models, it's a good idea to revise them using the Fable 5.2 prompting guide.

Who Can Use It in the Claude Apps

Based on the Claude Help Center articles (retrieved November 12, 2026), here is how access works. Fable 5.2 is available on all paid plans (Pro, Max, Ultra, Team, and Enterprise). On web, desktop, and mobile, choose "Fable 5.2" from the model picker; Claude Code requires version 2.1.321 or later.

  • Max and Ultra plans, and Team and Enterprise premium seats: You can use Fable models for up to 50% of your weekly usage limit at no extra cost. Fable uses up your limit faster than other models.
  • Pro plans and Team standard seats: Normally, you use Fable with usage credits (pay as you go). As a launch offer, until 11:59 pm PT on November 16, up to 50% of your weekly limit is included in your plan.
  • Enterprise: Fable 5.2 stays off until an admin turns it on. Seven days after launch (around November 16), it will be added automatically to roles that have Fable models enabled, and if the default model is Fable, it will switch to Fable 5.2. This won't happen if an admin opts out in advance.

The launch offers also include Fable limit resets, which restore your Fable usage. Max, Ultra, and premium seats get one to three depending on how long they've been subscribed, and Pro and Team standard seats get one. Eligibility, how to use them, and expiry dates are covered in How to Use Your Claude Fable Limit Reset.

The Bottom Line

Claude Fable 5.2 is faster and cheaper than Fable 5.1, and it beat Opus 5.5 on most of the benchmarks Anthropic published. That said, since Anthropic itself recommends starting everyday work with Opus 5.5, it makes more sense to think of Fable 5.2 as a model you reach for on long-running agentic tasks or hard problems Opus can't crack. If you plan to move to it on the API, start by checking whether any of your code modifies conversation history.

Sources

Back to list