Claude Fable 5.2 vs Opus 5.5 vs GPT-6 Astra vs Gemini 4 (2026)

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026 (Anthropic announcement). Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, exactly twice Claude Opus 5.5. So the Fable 5.2 vs Opus 5.5 question is about cost per task more than cost per token.

The other flagships on launch day are OpenAI's GPT-6 Astra and Google's Gemini 4 Argon, plus Gemini 4 Carbon, which Google has opened only to a small group so far. Pick the wrong one and you pay twice for work a cheaper model does as well.

This page puts six models in one table: price, context window, inputs and who can use them. Then come the benchmark claims each vendor published, with the source next to each number, and a pick for coding, agents, writing and cost. We did not run these models for this post. Every score below is the vendor's own claim, read on November 9, 2026.

TL;DR: Keep Claude Opus 5.5 as your default and send a task to Claude Fable 5.2 at High effort when Opus falls short. Anthropic reports that setting matches Opus 5.5's best FrontierCode score for about half the cost per task. If token price matters most, Gemini 4 Argon costs $2 / $10 per million tokens through November 30, 2026.

Claude Fable 5.2 vs Opus 5.5 vs GPT-6 Astra vs Gemini 4: a quick overview

If you can use any model today:

• Claude Fable 5.2: best for agent runs that last hours and for coding tasks Opus 5.5 cannot finish. It leads every coding row in Anthropic's table.
• Claude Opus 5.5: best default for daily coding and agent work. Anthropic's own docs tell most teams to start here.
• GPT-6 Astra: best for teams already built on OpenAI's API and tools, such as the Responses API.
• Gemini 4 Argon: best for the lowest token price and for audio, video or PDF input.

If you work in security or life sciences:

• Claude Mythos 5.2: the same model as Fable 5.2 with fewer safeguards, for organizations Anthropic has verified.
• Gemini 4 Carbon: Google's newest frontier model, open to a subset of cyber defenders.

Prices and specs below come from each vendor's own docs, checked on November 9, 2026. Prices are per million tokens, input / output.

Sources for the table: Claude models overview, Claude pricing, GPT-6 Astra model page, Gemini models, Gemini pricing, Gemini 4 Carbon launch post.

Official benchmark claims, side by side

Anthropic published one table that puts Fable 5.2 next to Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. The rows below are copied from the Fable 5.2 announcement. Anthropic says every score uses each model's highest available effort unless a cell says otherwise, with safeguards on for its own models.

Source for every row: Anthropic, November 9, 2026.

Three things to read before you quote a cell. First, Fable 5.2 does not win every row: Gemini 4 Carbon leads AutomationBench, and Opus 5.5 has the higher AA-Briefcase Elo. Second, OpenAI's developer page for GPT-6 Astra lists no benchmark scores, so every Astra number here comes from Anthropic's runs. Third, Google's own Gemini model page prints the same Carbon scores on the shared rows (61.2% FrontierSWE, 53.9% AutomationBench, 81.5% OSWorld, 72.5% Chartography), which is a good sign the numbers were not cherry-picked.

Two independent leaderboards now agree with Anthropic's table. On the Vals Index, Fable 5.2 ranks first of 36 models at 71.31%. On the Artificial Analysis Intelligence Index v5.0.1, Claude Fable 5.2 (max) scores 65 and ranks #1 of 74 (checked November 9, 2026). Gemini 4 Carbon (high) scores 60, Opus 5.5 (max) 59 and GPT-6 Astra (max) 57.

That top score is expensive. Artificial Analysis puts Fable 5.2 (max) at $26.54 per index task, against $10.04 for Opus 5.5 (max).

Fable 5.2 vs Opus 5.5: cost per task and the verdict

Fable 5.2 beats Opus 5.5 on 9 of the 10 rows above, and it costs twice as much per token. Anthropic's answer to the price gap is effort: at a lower effort setting Fable 5.2 uses fewer tokens and can still match Opus 5.5 at its highest setting.

Here is every cost-per-task claim on the announcement page, all from Anthropic's own early testing:

Verdict. Keep Opus 5.5 as the default, as Anthropic's models overview advises. Move a task to Fable 5.2 when it runs for hours, spans many files, or still fails on Opus 5.5 at high effort. Start Fable at High effort, not Xhigh. High is the API default for Fable 5.2 and the setting behind the half-cost claim. Save Xhigh for the problems where the extra 3.8 points on FrontierCode (49.1% vs 45.3%) are worth a bigger bill.

Is Fable 5.2 worth 2x the price of Opus 5.5?

Fable 5.2 is worth 2x per token when the task is long or hard enough that Opus 5.5 needs its Max effort setting to finish. On Anthropic's runs, that is where Fable at a lower setting costs less per task than Opus at its top one.

The price facts, from the Claude pricing page on November 9, 2026:

• Base price: Fable 5.2 $8 / $40, Opus 5.5 $4 / $20.
• Cached input costs the same on both: $0.20 per million tokens (2.5% of base on Fable 5.2, 5% on Opus 5.5).
• Batch API: 50% off both, so Fable 5.2 in batch ($4 / $20) costs the same as Opus 5.5 at standard rates.
• Fast mode: Fable 5.2 is $16 / $80 for up to 2.5x faster output, Opus 5.5 is $8 / $40.

Opus 5.5 is still the safer pick in three cases:

• You need zero data retention, which Fable 5.2 does not support unless Anthropic authorizes it (Fable 5.2 model page).
• Your code forces tool use, which returns a 400 error on Fable 5.2.
• The task is short, where a cheaper token beats a smarter one.

1. Claude Fable 5.2, best for long agent runs and the hardest coding

Claude Fable 5.2 is Anthropic's most capable generally available model, built for demanding reasoning and agent sessions that run for hours (Fable 5.2 model page). It shares one underlying model with Mythos 5.2, with stronger cyber and biology safeguards so anyone can use it.

Anthropic's examples are big: a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours, and agents migrated a production service from Python to Rust (announcement). Those are Anthropic's own reports.

Key features

• Adaptive thinking that is always on, steered by effort from Low to Max
• 1M token context window and 128K output tokens
• Fast mode at up to 2.5x the output speed
• Up to 30% faster output than Fable 5.1, per Anthropic
• Can act as an advisor or orchestrator for Haiku, Sonnet and Opus

Best for

• Engineering teams running agents for hours on migrations, refactors or research
• Tasks where Opus 5.5 at high effort still fails your evals

Pricing $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1's $10 / $50. Batch is 50% off and cached input is $0.20. In the Claude apps, Fable is included for up to 50% of weekly usage limits on Pro plans and standard Team seats (as of Nov 9, 2026, Claude pricing).

Pros

• Highest score on 8 of 10 rows in Anthropic's comparison table, and ahead of Opus 5.5 on 9
• Lower effort settings match Opus 5.5's best on several tests at a fraction of the cost
• Cheaper per token than GPT-6 Astra

Cons

• Blocks high-risk cyber work and most research biology tasks; Opus 4.8 or Opus 5 answers those instead
• On Amazon Bedrock only through US Geo or Global cross-Region profiles, with no single-Region option
• Needs 30-day data retention unless Anthropic authorizes otherwise

2. Claude Opus 5.5, best default for daily coding and agents

Claude Opus 5.5 is Anthropic's model for long-running agentic coding and knowledge work, and the one its models overview tells most teams to start with. The Fable 5.2 announcement calls it Anthropic's "strongest daily driver."

It has the same 1M context window and 128K output as Fable 5.2 at half the token price. It runs on the same clouds, including Amazon Bedrock as anthropic.claude-opus-5-5.

Key features

• 1M token context window, 128K output tokens
• Adaptive thinking, default effort Medium on the API
• Fast mode at $8 / $40
• Available on the Claude API, Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS

Best for

• Developers in Claude Code who want one model for most tickets
• Teams that need zero data retention

Pricing $4 per million input tokens and $20 per million output tokens. Batch is 50% off and cached input is $0.20 (as of Nov 9, 2026, Claude pricing).

Pros

• Half the token price of Fable 5.2 with the same context and output limits
• Highest AA-Briefcase Elo in Anthropic's table (1800)
• Widest cloud availability of the Claude flagships

Cons

• Behind Fable 5.2 on 9 of 10 rows in Anthropic's table, by 7.6 points on FrontierSWE v2
• Needs Max effort to reach scores Fable 5.2 hits at High, which raises cost per task

3. GPT-6 Astra, best for teams built on OpenAI's API

GPT-6 Astra is OpenAI's flagship, described on its model page as "our most capable model for the most demanding work." It takes text and images, returns text, and has a 1.05M token context window with 128K output.

Its edge over the Claude models is the OpenAI tool set: the Responses API gives it web search, hosted shell, computer use, MCP, code interpreter and skills. It is also listed on Amazon Bedrock as openai.gpt-6-astra.

Key features

• 1.05M token context window, up to 922K input tokens
• Reasoning effort from low to max
• Responses API tools: hosted shell, apply patch, computer use, MCP
• Knowledge cutoff April 30, 2026

Best for

• Teams whose agents already run on OpenAI's Responses API
• Bedrock users who want an OpenAI model next to Claude

Pricing $10 per million input tokens and $50 per million output tokens for prompts up to 272K tokens. Longer prompts cost 2x on input and 1.5x on output for the whole request. Batch and Flex are 50% off; fast mode is 2x (as of Nov 9, 2026, OpenAI pricing).

Pros

• Largest context window of the open flagships (1.05M tokens)
• Deep built-in tool support in one API
• Second on FrontierSWE v2 in Anthropic's table (65.5%)

Cons

• Highest token price here, and long prompts cost more
• Behind Fable 5.2 on every row where Anthropic scored both
• OpenAI's developer page publishes no benchmark scores to check Anthropic's numbers against

4. Gemini 4 Argon, best for low token cost and audio or video input

Gemini 4 Argon is the model Google's Gemini API docs call "our flagship reasoning model," with a shared 1M token context window and long outputs. It is the only flagship here that lists audio, video and PDF as inputs.

Google launched it to developers on October 13, 2026, through the paid Gemini API tier in Google AI Studio and Android Studio. Google AI Ultra subscribers can also use it in Antigravity.

Key features

• 1M token shared context window
• Output listed as "65K / 1M tokens"
• Accepts text, image, audio, video and PDF
• Batch, Flex and Priority tiers

Best for

• High-volume agents where token price decides the budget
• Apps that read video, audio or long PDFs

Pricing $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 / $20 from December 1. Not on the Gemini API's no-cost tier. Batch and Flex are half price (as of Nov 9, 2026, Gemini pricing).

Pros

• A quarter of Fable 5.2's token price until November 30
• 51.3% on AutomationBench in Google's own table, close to Fable 5.2's 51.4% in Anthropic's
• Widest range of input types

Cons

• Price doubles on December 1, 2026, to match Opus 5.5
• Behind Carbon on every row of Google's own table (55.0% vs 61.2% on FrontierSWE 2)
• Knowledge cutoff not stated on the model page

5. Claude Mythos 5.2, best for verified cyber and life sciences teams

Claude Mythos 5.2 is the same underlying model as Fable 5.2, offered with fewer safeguards to organizations verified through Anthropic's Cyber Verification Program or Life Sciences Verification Program (announcement). Anthropic calls it its most capable model for cybersecurity engineering.

Claude Enterprise customers can also reach it through Claude Security without joining a verification program. Everyone else cannot use it.

Key features

• Same specs and pricing as Fable 5.2, per the Fable 5.2 model page
• 89.6% on Anthropic's CVE Rediscovery test, vs 78.3% for Mythos 5.1
• Fast mode for verified organizations
• Powers Claude Security for Claude Enterprise

Best for

• Security teams doing penetration testing and vulnerability research
• Life sciences labs working on protein design

Pricing $8 per million input tokens and $40 per million output tokens, $4 / $20 in batch, listed as limited availability (as of Nov 9, 2026, Claude pricing).

Pros

• Does the cyber and biology work that Fable 5.2 blocks
• Found 10.2% more vulnerabilities than Opus 5.5 on Anthropic's internal test

Cons

• Closed to the public; access goes through a verification program
• Anthropic's own evals put its vulnerability identification "roughly on par with Opus 5.5"

6. Gemini 4 Carbon, best for cyber defenders in Google's early access group

Gemini 4 Carbon is Google's newest frontier model, announced on October 28, 2026, and described as Google's strongest model for cybersecurity and long tasks (Google launch post). Access started with a subset of cyber defenders and is expanding to trusted partners through Google's Fairwind Program.

Google says it will open Carbon to developers, businesses and consumers "as soon as possible," with no date. Context window and input types are not stated yet.

Key features

• 59 on the Artificial Analysis Intelligence Index v5.0, ranked first in Google's post. On v5.0.1, Carbon (high) scores 60 and sits behind three Fable 5.2 effort settings (Artificial Analysis, Nov 9, 2026)
• 88.2% on SEC-bench Pro (finding vulnerabilities) and 71.0% on CWE-bench (patching)
• 0.0% attack success on the Gray Swan indirect prompt injection test
• Cached input at 95% off

Best for

• Security teams already in Google's early access group
• Teams planning a Gemini migration once Carbon opens

Pricing $2 per million input tokens and $10 per million output tokens at launch, rising to $4 / $20 after the introductory period (as of Nov 9, 2026, Google launch post).

Pros

• Beats Fable 5.2 on AutomationBench (53.9% vs 51.4%) in Anthropic's own table
• Half Fable 5.2's token price even after the introductory period

Cons

• Not open to developers yet
• Behind Fable 5.2 on every coding row Anthropic published

Claude Fable 5.2 vs GPT-6 Astra

Claude Fable 5.2 scores higher than GPT-6 Astra on every benchmark where Anthropic tested both, and it costs less per token. The gaps are widest on knowledge work and reasoning:

• Remote Labor Index: 41.3% vs 20.8%
• Humanity's Last Exam with tools: 74.2% vs 57.2%
• OSWorld 2.1: 89.3% vs 79.5%
• FrontierSWE v2: 69.9% vs 65.5%, the closest coding row

On price, Fable 5.2 is $8 / $40 and GPT-6 Astra is $10 / $50, and Astra's rate rises past 272K input tokens. All of these scores come from Anthropic's runs (announcement). Astra still makes sense if your agents depend on the Responses API tools.

Fable 5.2 vs Gemini 4 Carbon: where Gemini wins

Gemini 4 Carbon beats Claude Fable 5.2 on business workflow automation and on price, and Fable 5.2 wins on coding and computer use. Gemini wins in three places:

• AutomationBench: Carbon 53.9%, Fable 5.2 51.4% in Anthropic's own table.
• Price: Carbon's introductory $2 / $10 is a quarter of Fable 5.2's rate; Argon matches it until November 30.
• Inputs: Argon takes audio, video and PDF. Fable 5.2 takes text and images.

Fable 5.2 leads Carbon on FrontierCode (49.1% vs 41.1%, with Carbon at Medium effort), FrontierSWE v2 (69.9% vs 61.2%) and OSWorld 2.1 (89.3% vs 81.5%). Chartography is close: 73.4% vs 72.5%. The cyber scores do not compare: Google reports SEC-bench Pro and CWE-bench, and Anthropic reports CVE Rediscovery and CyScenarioBench.

How to choose: Fable 5.2 vs Opus 5.5 vs Sonnet 5.5 by task

Pick by task, not by brand. Claude Sonnet 5.5 ($2 / $10) and Claude Haiku 5.5 (from $0.10 / $0.50) belong in the same routing plan (Claude models overview, Nov 9, 2026).

1) Coding: Opus 5.5 first, Fable 5.2 for what it cannot finish

The best Claude model for coding depends on the ticket. If your tickets are small and many, Opus 5.5 at Medium is the cheaper path. If a migration or refactor keeps failing, move it to Fable 5.2 at High before you try Opus at Max. Run both on 10 of your own tickets and compare cost per merged change, not cost per token.

2) Agents: Fable 5.2 for long runs, Argon for cheap volume

For agents that click through software or run for hours, Fable 5.2 has the best published scores. For high-volume workflow agents, Gemini 4 Argon is close on AutomationBench at a quarter of the price until November 30. Recheck the math on December 1, when Argon's price doubles.

3) Writing: test on your own drafts

None of the official pages we read report a writing benchmark. Anthropic says testers found Fable 5.2 communicates more clearly than past Fable models, which is a claim, not a score. Start with Sonnet 5.5 or Opus 5.5 for drafts and move up only if your editors see a difference.

4) Cost: count cost per task

Sonnet 5.5 and Gemini 4 Argon are the cheapest capable options per token this month. Batch pricing halves any Claude model, and cached input costs $0.20 on both Fable 5.2 and Opus 5.5. Effort setting moves cost per task more than the model choice does.

Picking a model is half of a launch-week buying decision. The other half is being named when buyers ask ChatGPT and Perplexity which tool to use, and those answers lean on comparison pages like this one. If you sell a SaaS product, request a spot on AY Listicle to get named in the "best of" lists your buyers read. More comparisons live on the blog.

FAQ

What is Claude Fable 5.2?

Claude Fable 5.2 is Anthropic's model for demanding reasoning and long agent runs, released on November 9, 2026. It costs $8 / $40 per million tokens, has a 1M token context window and returns up to 128K tokens. Its API ID is claude-fable-5-2.

Is Claude Fable 5.2 better than Opus 5.5?

Claude Fable 5.2 scores higher than Opus 5.5 on 9 of the 10 benchmark rows in Anthropic's launch table. Opus 5.5 keeps a higher AA-Briefcase Elo (1800 vs 1785) and costs half as much per token. Anthropic still recommends Opus 5.5 as the starting point for most workloads.

How much does Claude Fable 5.2 cost?

Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens on the Claude API as of November 9, 2026. Fast mode costs $16 / $80, batch requests are 50% off, and cached input costs $0.20 per million tokens.

Can I use Claude Mythos 5.2?

Claude Mythos 5.2 is only for organizations verified through Anthropic's Cyber Verification Program or Life Sciences Verification Program. Claude Enterprise customers can use it through Claude Security. Everyone else gets Fable 5.2, the same model with stronger safeguards.

Is Claude Fable 5.2 free to use?

No, Claude Fable 5.2 has no no-cost API tier: the API price is $8 / $40 per million tokens. In the Claude apps, Fable is included for up to 50% of weekly usage limits on Pro plans and standard Team seats, and launch week adds one to three Fable resets on eligible paid plans.

Is Claude Fable 5.2 on Amazon Bedrock?

Yes. AWS lists Claude Fable 5.2 on Amazon Bedrock as anthropic.claude-fable-5-2, launched November 9, 2026 (Bedrock model card). You call it through us.anthropic.claude-fable-5-2 or global.anthropic.claude-fable-5-2, because there is no single-Region option. Your account must opt in to the aws_review data retention mode first.

Which is cheaper, GPT-6 Astra or Claude Fable 5.2?

Claude Fable 5.2 is cheaper per token: $8 / $40 against GPT-6 Astra's $10 / $50 per million tokens. Astra's price also rises for prompts over 272K input tokens. Both offer batch pricing at 50% off.

Has anyone tested Claude Fable 5.2 independently?

Vals AI ranks Claude Fable 5.2 first of 36 models on the Vals Index, at 71.31%. Artificial Analysis ranks Claude Fable 5.2 (max) #1 of 74 on its Intelligence Index v5.0.1, at 65 (Artificial Analysis, checked November 9, 2026). Every other number on this page is the vendor's own claim.

Sources: Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, Claude models overview, Claude Fable 5.2 model page, Claude pricing, OpenAI GPT-6 Astra model page, OpenAI API pricing, Gemini API models, Gemini API pricing, Google DeepMind Gemini page, Google, Gemini 4 Carbon launch, Google, Gemini 4 Argon for developers, Vals AI Fable 5.2 page, Artificial Analysis, Claude Fable 5.2, Artificial Analysis leaderboard, Amazon Bedrock, Claude Fable 5.2 model card. All checked November 9, 2026.