About Services Consulting Workshop Certificates Events Work Blog Links Contact 🇵🇦 Español Book VIP Consulting →
← All articles
November 11, 2026 · 9 min read · By Chiragx

Claude Fable 5.2: more capable than Opus 5.5 and 20% cheaper

AIClaudeAnthropicAI modelsClaude Fable
Claude Fable 5.2: more capable than Opus 5.5 and 20% cheaper

On November 9, 2026, Anthropic released Claude Fable 5.2, its most capable model open to everyone. Three facts sum up the launch: it beats Opus 5.5 on almost every test Anthropic published, it costs 20% less than Fable 5.1, and it writes up to 30% faster.

According to Anthropic, in prerelease testing it completed software migrations that were not feasible before, designed three working electronic devices from scratch, solved two open problems in black hole physics and formalized Fermat's Last Theorem in 30 hours.

Here is everything: what it is, what changes, the results, what it costs, where to use it, how to set it up and when it is really worth it. The facts come from Anthropic's announcement and official documentation, linked at the end.

What Claude Fable 5.2 is

Fable is Claude's most powerful line, above Opus. It is built for demanding reasoning and long agentic work: large, loosely defined problems that can run for hours or days, such as migrating a whole system, coordinating dozens of subagents or researching a scientific problem.

Alongside Fable 5.2 came Claude Mythos 5.2. According to Anthropic, they are the same underlying model. The difference is the safeguards: Fable has strong cybersecurity and biology protections and is open to everyone; Mythos does not have those blocks and only verified organizations can use it.

Diagram: Claude Fable 5.2 and Claude Mythos 5.2 are the same model. Fable is for everyone, in the Claude apps, the API and the clouds, with cybersecurity and biology safeguards. Mythos is only for organizations verified through Anthropic's verification programs
Fable 5.2 and Mythos 5.2: the same model, with different safeguards and different access.

Key specs

  • API name: claude-fable-5-2.
  • Context window: 1 million tokens, what it can read at once. That is several books or a large code repository.
  • Max output: 128,000 tokens.
  • Knowledge cutoff: June 2026.
  • Thinking: always on. It cannot be turned off; you control it with the effort level.
  • Data: requires 30-day retention for safety monitoring. It does not run under zero data retention unless Anthropic expressly authorizes it.
  • Predecessor: Claude Fable 5.1, released September 1, 2026.

What changes from Fable 5.1

  • Cheaper. $8 per million input tokens and $40 per million output tokens, down from $10 and $50.
  • Faster. It writes up to 30% faster. It also adds a fast mode that speeds up generation by up to 2.5x, at $16 and $80 per million.
  • Fewer false blocks. Anthropic says it kept the same cybersecurity and biology protection with about 55% fewer blocks: about 60% fewer in cybersecurity and 40% fewer in biology, based on early testing.
  • More capable. Better at coding, reasoning and vision. The numbers are in the next section.
  • Better protected against copying. According to Anthropic, its anti-distillation defenses (against others training models on its answers) cut the data those actors could harvest by 90% to 99%.

The test results

Anthropic compared Fable 5.2 with Fable 5.1, with Opus 5.5 and with models from other companies. This chart puts eight of those tests side by side against its own models:

Bar chart of eight tests published by Anthropic: Fable 5.2 scores 49.1% on FrontierCode 2.0, 69.9% on FrontierSWE v2, 74.6% on Terminal-Bench-Science, 74.2% on Humanity's Last Exam, 41.3% on Remote Labor Index, 89.3% on OSWorld 2.1, 64.9% on RoboDojo Sim and 73.4% on Chartography, above Opus 5.5 and Fable 5.1 on all of them
Fable 5.2 against Opus 5.5 and Fable 5.1, according to Anthropic's figures.

What stands out:

  • Real work: on Remote Labor Index, which measures real remote jobs end to end, it goes from 17.9% to 41.3%. More than double Fable 5.1, and almost double Opus 5.5.
  • Robotics: from 27.6% to 64.9% on RoboDojo Sim.
  • Science: 74.6% on Terminal-Bench-Science, 16 points above Opus 5.5.
  • Computer use: 89.3% on OSWorld 2.1, where Opus 5.5 was already high (87.2%).

Two caveats, to read the numbers calmly:

  • It does not win everywhere. On AutomationBench, which measures business workflows, Gemini 4 Carbon scores 53.9% against 51.4%. And on the AA-Briefcase Elo score for office work, Opus 5.5 is just ahead: 1800 against 1785.
  • These are Anthropic's figures, and effort matters. The 49.1% on FrontierCode 2.0 is at xhigh effort. The first independent tests point the same way, with different numbers: Vals AI ranks it first of 36 models on its overall index, but second on SWE-bench Multimodal, behind Opus 5.5.

What it costs

Price table per million tokens: Opus 5.5 costs $4 input and $20 output, Fable 5.2 $8 and $40, Fable 5.1 $10 and $50, and Fable 5.2 in fast mode $16 and $80. Example task with 100,000 input tokens and 10,000 output tokens: $0.60 on Opus 5.5, $1.20 on Fable 5.2 and $1.50 on Fable 5.1
API prices per million tokens, and the cost of an example task.

Fable 5.2 costs 20% less than Fable 5.1, but still twice as much as Opus 5.5 per token. Two details bring the bill down:

  • Caching: reading context that is already cached costs $0.20 per million, the same as on Opus 5.5. If you reuse the same context a lot (a manual, a repository, a long contract), the price gap shrinks.
  • Batch: work that is not urgent costs half with the batch API: $4 and $20.

If you need inference to run only in the United States, the price is 1.1x.

Where you can use it

  • Claude apps (web, desktop and mobile): on the Pro, Max, Ultra, Team and Enterprise plans. In the apps it runs at medium effort.
  • Usage limits: on Pro and on standard Team seats, Fable can take up to 50% of the weekly limit. In launch week, paid plans that include it get extra usage and one to three Fable resets, depending on how long you have been subscribed.
  • Claude Code: also at medium effort by default, with fast mode available.
  • API: for all customers, as claude-fable-5-2.
  • Clouds: Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

Mythos 5.2 is requested separately: cybersecurity professionals apply to the Cyber Verification Program, life sciences researchers to the Life Sciences program, and Enterprise customers can use it inside Claude Security.

How to choose the effort level

Effort is the most important setting on Fable 5.2: it decides how much the model thinks before answering, how long it takes and how many tokens it spends. There are five levels:

Guide to the five effort levels of Claude Fable 5.2: low for simple, high-volume tasks, medium for chat and everyday work (the default in the apps and in Claude Code), high for hard code and research (the default in the API), extra high only for the hardest work and max to measure the model's ceiling
Which effort level to use for each case, based on Anthropic's guide.

Three things Anthropic points out in its guide:

  • At high effort or above, it thinks before almost every reply, even a greeting. For easy turns, lower the effort.
  • Asking it in the prompt to think less no longer works at medium effort or above. If you want less thinking, lower the level.
  • Effort barely changes how long the replies are. Ask for length and tone in the prompt.

What it is good for: what early testers reported

Anthropic recommends it for coding across large repositories, agents that work for days, scientific research, data analysis, financial analysis and 3D, hardware and robotics design. These are some of the companies that tested it before launch:

  • Uber ran a 10 to 12 hour security audit with about 170 reviewer agents, and the model was honest about what it could not cover.
  • iGent built a hypervisor with more than 3,500 machine-checked proofs in a single session.
  • Discovered Materials made Quantum ESPRESSO's phonon code run about 36 times faster on average.
  • Box saw it find inconsistencies in documents, such as salary deposits that did not match the bank balance in a loan application.
  • Harvey got clearer legal compliance guidance than with Fable 5.1 in an I-9 review.
  • Cognition noted that it admits when it does not know something instead of making it up.
  • Datadog uses it to coordinate faster models, like an incident commander.
  • Snowflake says it is the strongest model it has measured on its data engineering benchmark.

Limits and cautions

  • It is not for everything. For most everyday coding and office work, Anthropic recommends starting with Opus 5.5: it costs half and answers faster.
  • Cybersecurity and biology. Fable blocks most cybersecurity tasks. In the apps, blocked requests are answered by Opus 4.8 (cybersecurity) or Opus 5 (biology); on the API you need to set up a fallback model.
  • Long turns. On hard tasks, a single reply can take several minutes.
  • It can stop before it is done on long tasks, mostly at low or medium effort. Anthropic suggests telling it in the prompt not to stop while work remains.
  • Model safety. Anthropic says high-stakes risks from misalignment remain low, but admits that building evaluations that catch every failure before a model ships is still unsolved. It also noticed these models suspect more often that they are being evaluated.

If you build on the API: what changes

Moving from Fable 5.1 to Fable 5.2 is mostly a model name change. Check four things:

  1. Do not edit the history. On Fable 5.2 the preserved thinking check applies to every account: if your code changes the system prompt, the tools or an earlier message mid-conversation, the API returns a 400 error.
  2. Its reasoning does not travel back. No earlier model reads its thinking blocks. If you switch models mid-conversation, the other model continues without that reasoning.
  3. Computer use: use the computer_toolset_20260801 tool. The earlier one, computer_20251124, returns an error except on Amazon Bedrock.
  4. Advisor: if you use the advisor tool, the advisor for a Fable 5.2 executor must be Fable 5.2 or Mythos 5.2.

And the usual rules for the Fable line: thinking cannot be turned off (you control it with effort), forcing a tool with tool_choice is not accepted, and you need to handle declined requests, which come back with stop_reason "refusal". A Python example with explicit effort and Anthropic's automatic fallback for refusals:

from anthropic import Anthropic

client = Anthropic()
with client.beta.messages.stream(
    model="claude-fable-5-2",
    max_tokens=64000,
    output_config={"effort": "high"},
    betas=["server-side-fallback-2026-07-01"],
    fallbacks="default",
    messages=[{"role": "user", "content": "Your task here"}],
) as stream:
    msg = stream.get_final_message()

if msg.stop_reason == "refusal":
    print("Declined:", msg.stop_details)
else:
    print("".join(b.text for b in msg.content if b.type == "text"))

Is it worth it for you?

  • Use Opus 5.5 for everyday work: chat, email, content and regular coding. It costs half and answers faster.
  • Use Fable 5.2 when Opus 5.5 at high effort is not enough: large, loosely defined projects, migrations, research, multi-agent analysis or hours of unattended work.
  • If you already use Fable 5.1, switch: Fable 5.2 is better and cheaper.

My take: the most important part of the launch is not the record scores, but that the most capable model got cheaper and shipped on the usual plans. For a business in Latin America, the question is not which model is best, but which work you leave undone today because it is too long, too expensive or too complex for your team. That is where Fable 5.2 starts to make sense.

Bring it to your company

My team helps companies pick the right model for each task and put it to work: agents that serve customers, CRMs that follow up on their own, document analysis and automations that save hours every week. If you want to see what makes sense for your case, book a VIP Consulting session.

Book your VIP Consulting session → (the booking page is in Spanish)

Sources

Chiragx
Written by
Chiragx · Chiragx Bhakta
AI & technology advisor · Founder of CryptoManji, BBR Tek, Bell N Desk and Nova Ignis
About me · LinkedIn

This is where I can help

AI consulting and automation for businesses in Panama
See the service →

Want this for your business?

Book your VIP Consulting session →
Newsletter

Get every article in your inbox

AI and business analysis, no hype. Only when I publish something worth your time.