# Claude Fable 5.2 vs Gemini 4 Carbon

## Which should you use: Claude Fable 5.2 or Gemini 4 Carbon?

> It depends. Fable 5.2 leads Artificial Analysis's Intelligence Index (65 at max, 62 at high versus Carbon's 60 at high) and MKA (57% versus 54%), but costs more: $8/$40 per 1M tokens against Carbon's $2/$10 through 30 November 2026 and $4/$20 from 1 December. On MKA, Fable averages $31.97 per task versus Carbon's $6.29 (AA used the $4/$20 list, not the intro rate). Carbon leads AutomationBench-AA, Terminal-Bench 4.0 and GDP.pdf. Pick Fable for the top score on hard technical work if you can absorb the cost; pick Carbon for cheaper frontier reasoning, especially before 1 December.

Claude Fable 5.2 (Anthropic) launched on 9 November 2026. Gemini 4 Carbon (Google) opened to paying developers on 10 November, two weeks after a Fairwind-only cyber preview ([Anthropic](https://www.anthropic.com/claude-fable-and-mythos-5-2), [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)). "Vendor" marks a claim from the model's maker; "AA" marks an independent measurement by Artificial Analysis. All facts are as of 11 November 2026.

**Carbon's price changes on 1 December.** Through 30 November 2026 the public Gemini Developer API rate is $2 / $10 per 1M input / output tokens. From 1 December 2026 it becomes $4 / $20. Artificial Analysis's Intelligence Index and MKA cost-per-task figures for Carbon use the standard $4 / $20 list, not the introductory rate ([AA](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1), prices re-confirmed on the [comparison table](https://artificialanalysis.ai/models/releases/comparisons) on 11 November 2026).

## Side by side

| | Claude Fable 5.2 | Gemini 4 Carbon |
|---|---|---|
| Model ID | `claude-fable-5-2` ([Claude docs](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2)) | `gemini-4-carbon` ([Google](https://aistudio.google.com/docs/whats-new-gemini-4-carbon)) |
| Released | 9 November 2026 | Opened to paying developers 10 November 2026 |
| Input / output, $ per 1M tokens | $8 / $40 ([Claude docs](https://platform.claude.com/docs/en/about-claude/pricing)) | $2 / $10 through 30 November 2026; $4 / $20 from 1 December 2026 ([Google pricing](https://aistudio.google.com/docs/pricing)) |
| Cache read, $ per 1M tokens | $0.20 (5-minute and 1-hour caches) | Implicit cache only: $0.10 through 30 November 2026, $0.20 from 1 December (95% off input). Explicit cache APIs are not supported. Minimum prompt for implicit cache: 4,096 tokens |
| Cache write, $ per 1M tokens | $10 (5-minute) / $16 (1-hour) | No separate write price; implicit caching only |
| Faster or cheaper variants | Fast mode $16 / $80 (`speed: "fast"`, up to 2.5× output speed); Batch $4 / $20. Fast mode is not available with Batch | Flex: $1 / $5 through 30 November 2026, $2 / $10 from 1 December. Priority: $3.60 / $18, then $7.20 / $36. Paid tier also advertises Batch at 50% off the then-current standard rate |
| Long-context surcharge | None; the full 1M window is billed at the standard rate (Claude 4.6 and later, except Haiku 5.5) | None published; unified 1M window (input + thinking + output) |
| AA Intelligence Index v5.0.1 | 65 at max; 62 at high and xhigh ([AA](https://artificialanalysis.ai/models/releases/comparisons)) | 60 at high; 58 at medium |
| MKA Climb WSR (AA) | 57% at max, $31.97 per task ([AA](https://artificialanalysis.ai/articles/google-deepmind-multimodal-knowledge-agent)) | 54% at high, $6.29 per task |
| AutomationBench-AA / Terminal-Bench 4.0 / GDP.pdf (AA) | 76.4% / 59.6% / 24.2% at max | 77.8% / 65.2% / 28.4% at high |
| FrontierCode-AA 2.0 / HLE-Diamond / AA-Omniscience v1.1 (AA) | 48.8% at max (49.1% at high) / 80.2% / 67, at max | 40.6% / 69.1% / 54, at high |
| Cost per Index task (AA, standard list prices) | $26.54 at max; $9.08 at high | $7.76 at high |
| Time per Index task (AA) | 45.2 min at max; 17.3 min at high | 25.6 min at high |
| Output speed (AA) | 93–103 tok/s across published settings | 63–80 tok/s |
| Time to first answer token (AA) | 233.95 s at max; 30.40 s at high | 67.08 s at high |
| Context window | 1M tokens, default and maximum | Unified 1,048,576-token window shared by input, thinking and output |
| Max output | 128k tokens (thinking plus response) | Default 65,536 (64K); with `background=True` and a higher `max_output_tokens`, up to 1M minus the input |
| Thinking / effort | Always-on adaptive thinking. Five effort levels; API default is `high`, Claude apps and Claude Code default to `medium` | `thinking_level`: low, medium (default), high. No xhigh or max. `thinking_budget` is deprecated |
| Inputs (AA) | Text and image in, text out | Text, image, speech and video in, text out |
| Consumer access | Pro, Max, Ultra and Team, with Fable caps and launch offers (see below). Not on Free | Google AI Ultra in the Gemini app, AI Mode in Search, and Antigravity. Carbon in those surfaces is described as an Ultra rollout, not a Google AI Pro feature |

## Price and cost

Fable 5.2 is 20% cheaper than Fable 5.1 ($10 / $50). Thinking tokens are billed as output on both models ([Anthropic](https://www.anthropic.com/claude-fable-and-mythos-5-2), [Google](https://aistudio.google.com/docs/whats-new-gemini-4-carbon)).

- **Carbon intro rate:** the $2 / $10 price "applies to the Carbon model intended for public access only." Cyber-permissive SKUs on Google Cloud are listed at the post-1 December standard rates and are excluded from the 50% promotional credits ([Google blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/), [Model Garden pricing](https://cloud.google.com/model-garden/pricing)).
- **Regional premiums:** US-only Claude inference (`inference_geo: "us"`) is 1.1× for Claude 4.6 and later. Bedrock and Google-managed Cloud regional or multi-region Claude endpoints carry a 10% premium over global. On Google Cloud Agent Platform / Model Garden, non-global Carbon endpoints are 10% above global (intro $2.20 / $11, cache $0.11; from 1 December 2026 $4.40 / $22, cache $0.22).
- **Timeouts on long Carbon completions:** Google Cloud will charge for retrievable tokens on Gemini 4 completions longer than 32K even if the request later fails on a client or server timeout, effective 13 November 2026 ([Model Garden pricing](https://cloud.google.com/model-garden/pricing)).
- **Search tools:** grounding with Google Search on Gemini 3 and newer is 5,000 free requests per month shared across those models, then $14 per 1,000, with a default daily cap of 1,500 requests per project on the Gemini Developer API. Claude web search is $10 per 1,000 searches plus token costs.
- **Independent cost, with the list-price caveat:** on MKA, Fable's $31.97 is over 5× Carbon's $6.29 for about two extra points, and about 29× GPT-6.1 Sol ($1.11). All four OpenAI models tested sit on the cost-performance frontier below Carbon. On the Intelligence Index, Fable at high or xhigh already beats every Carbon setting (62 vs 60) at $9.08–$14.83 vs $7.76 per task ([AA](https://artificialanalysis.ai/models/releases/comparisons)).

## Independent benchmarks

**MKA (Multimodal Knowledge Agent)** is a Google DeepMind eval run by Artificial Analysis: 200 private visual knowledge-work tasks across 10 industries, graded on about 70 weighted rubric criteria per task. The headline score is Climb WSR, the share of harder "Climb" criteria met. The harness is AA's open-source Stirrup, up to 250 turns, judged by GPT-6 Luna (medium) ([AA article](https://artificialanalysis.ai/articles/google-deepmind-multimodal-knowledge-agent), [leaderboard](https://artificialanalysis.ai/evaluations/mka)).

No single model leads all 10 verticals. Fable leads five (Software & Data, Engineering, Logistics & Mapping, Business, and Finance alongside Sonnet 5.5); Opus 5.5 leads Education and Construction; Argon (high) leads Science; Carbon (high) leads Healthcare; Sonnet 5.5 narrowly leads Law & Government. Word and plain-text outputs are the hardest; no model passes 52% on either. Fable leads four file types (PDF, Word, narrowly Spreadsheet, and Presentation alongside Opus 5.5); Argon and Carbon lead Image.

**Artificial Analysis Intelligence Index v5.0.1** equally weights nine evaluations across Knowledge Work, Coding, Science and General. Carbon first reached the index on 28 October 2026 at 60 (high); Fable 5.2 later took the top slot at 65 (max) ([AA comparisons](https://artificialanalysis.ai/models/releases/comparisons)). Carbon leads AutomationBench-AA, Terminal-Bench 4.0 and GDP.pdf. Fable leads HLE-Diamond, AA-Omniscience, CritPt, Terminal-Bench-Science and FrontierCode-AA (high effort 49.1% is Fable's best FrontierCode-AA row). Carbon is slower in output tokens per second at every setting AA published.

## Vendor-reported scores

These are company figures, not AA's. Do not mix Anthropic's AutomationBench (Carbon 53.9% vs Fable 51.4%) with AA's AutomationBench-AA (Carbon high 77.8% vs Fable max 76.4%); they are different harnesses.

Anthropic's launch table, "highest-available effort unless otherwise noted," with safeguards on for Fable unless noted ([Anthropic](https://www.anthropic.com/claude-fable-and-mythos-5-2); methodology in the [system card](https://assets.claude.com/27d60918d218125e7504bc7354c4b4626ec56b11.pdf)):

- FrontierCode 2.0: Fable 49.1% at xhigh, 45.3% at high; Carbon 41.1% at medium. Anthropic says Fable at high roughly matches Opus 5.5's best (45.5%) at about half the cost per task.
- FrontierSWE v2: Fable 69.9%; Carbon 61.2%.
- AutomationBench: Fable 51.4%; Carbon 53.9% (Carbon leads this vendor column).
- Terminal-Bench-Science 0.1: Fable 74.6% (Mythos 5.2 75.3%); Carbon 64.8%.
- OSWorld 2.1 (offline set, partial): Fable 89.3%; Carbon 81.5%.
- AA-Briefcase v1.1: Fable 1785 Elo; Carbon 1611 Elo (Elo matches AA).

Google's 10 November blog says Carbon "ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task compared to the next best external model." Cursor is quoted as scoring Carbon 60.4% on CursorBench at high effort ([Google blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/), also circulated via [@demishassabis](https://x.com/demishassabis/status/2120294666969248255)). Google also says Carbon ranks #2 on MKA "while being more than 6X lower in average cost per task compared to the top performing model," which matches AA's $31.97 vs $6.29.

Anthropic also published company-reported long-horizon demos (a ~1,000-agent Lean formalization of Fermat's Last Theorem in 30 hours, an eight-day OS from an empty folder, three working electronic devices from a one-paragraph prompt, two holographic-duality calculations). Treat those as vendor claims.

## Context, thinking and features

- **Fable speed:** Anthropic says outputs are up to 30% faster than Fable 5.1, with optional fast mode up to 2.5× at $16 / $80. AA measured about 94–103 tok/s for Fable vs 63–80 tok/s for Carbon.
- **Carbon long decode:** with `background=True` and a raised `max_output_tokens`, output can go past the 64K default in 64K continuation hops, up to 1M minus the input. Google recommends raising client timeouts to 15–20 minutes for hard prompts, and keeping `thinking_level` at `medium` unless a workload needs `high`.
- **Carbon /swarm preview in Antigravity:** one prompt becomes a coordinated team (coordinator, isolated git-worktree workers, adversarial verifiers) on a shared append-only board. Google recommends Carbon for coordinator, architecture, audit and verification roles, with Argon or Gemini 3.8 Flash on parallel implementation. Parallel swarms "consume quota much faster than a standard chat" ([Antigravity](https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity/)).
- **Fable API migrations:** the prefix check is enforced for every account (editing system, tools or earlier messages and replaying a later thinking block returns 400). Computer use needs `computer_toolset_20260801`; the old `computer_20251124` tool returns 400 everywhere except Bedrock.
- **Fable data retention:** 30-day retention is required, like 5.1. Zero data retention only with Anthropic's approval, or via the Enterprise Frontier Safeguards private beta (Bridgewater, Salesforce and Stripe are named).
- **Fable safety, vendor-reported:** Fable and Mythos are the same underlying model; Fable has strong cyber and biology classifiers (blocked cyber queries fall back to Opus 4.8, biology to Opus 5). During training, AI-research skills that most directly speed up AI development were limited to about Opus 5.5 level; agentic coding, hardware and classic ML / interpretability were not. Anthropic's own advice is to start with Opus 5.5 as the daily driver and move up to Fable for the most open-ended technical work.
- **Carbon safety, vendor-reported:** in simulated attack environments Carbon blocked 99% of adaptive red-team attack paths, with a 0.8% per-attempt breach rate (Argon 3.1%). On a 375-task agentic honesty test it passed 92% (Argon 70%).

## Subscription access

**Claude (Fable 5.2),** from [claude.com/pricing](https://claude.com/pricing) and the [Team help article](https://support.claude.com/en/articles/9266767-what-is-the-team-plan), checked 11 November 2026:

- Free ($0): Fable is not included.
- Pro: $20 / month, or $17 / month billed annually ($200 up front). Fable is normally usage credits, not the included weekly pool.
- Max: $100 / month, 5× Pro every 5 hours and every week. Fable is included for up to 50% of weekly limits.
- Ultra: $200 / month, 10× Pro weekly; no 5-hour limit for verified accounts. Fable is included for up to 50% of weekly limits.
- Team Standard: $25 / member / month, or $20 billed annually (minimum two members). Same Fable treatment as Pro, except during the launch week below.
- Team Premium: $125 / member / month, or $100 billed annually. Fable is included for up to 50% of weekly limits.

There is no published fixed message count. Usage is shared across web, desktop, mobile and Claude Code. Max, Ultra and Team also include monthly Claude API credits ($100 on Max, $200 on Ultra, up to $500 on Team).

**Fable 5.2 launch offers** ([Help Center](https://support.claude.com/en/articles/17506494-claude-fable-5-2-launch-offers), updated 10 November 2026): eligibility cutoff is paid billing in place before 9 November 2026, 10:00 PT. Pro and Team Standard get all Fable models in the weekly pool from 9 November through 16 November 2026, 23:59 PT, capped at 50% of weekly limits, plus one Fable reset that expires at the same time; after the week, Fable returns to usage credits. Max, Ultra, Team Premium, and premium seats on legacy seat-based Enterprise get one to three Fable resets by subscription length, expiring 8 March 2027, 23:59 PT. A Fable reset restores weekly usage spent on Fable models only and does not raise the 50% cap.

Separately, @ClaudeDevs announced on 10 November 2026 that all Pro, Max, Ultra and Team users received one free full reset across all models and limits, usable any time before 10 December 2026, 14:00 PT. Anthropic said it had fixed "a few small issues that used more limits than intended." This is distinct from the Fable launch resets ([@ClaudeDevs](https://x.com/ClaudeDevs/status/2120275296637321476)).

**Google (Gemini 4 Carbon):**

- Google AI Ultra is the consumer plan that includes Carbon: $99.99 / month for 5× Google AI Pro usage, or $199.99 / month for 20×, with "highest access to Gemini 4 Carbon" in the Gemini app, AI Mode in Search, and Antigravity. Exact request counts are not published; Ultra quotas in Antigravity refresh every five hours and also have a weekly cap ([subscriptions](https://gemini.google/subscriptions/), [Antigravity plans](https://antigravity.google/docs/plans/)).
- Google AI Pro is advertised with Gemini 4 Argon and entry Antigravity rate limits. Carbon in the Gemini app, AI Mode and Antigravity is described as an Ultra rollout, not a Pro feature.
- The Gemini API and AI Studio offer Carbon on the paid tier only, not the Free tier. Ultra subscribers can also use it in AI Studio.
- Paying developers and enterprises can call it via the API or the model dropdown, and in Antigravity or CodeMender, per the 10 November blog.

## Availability

Fable 5.2 is generally available as of 9 November 2026 on the Claude API, Amazon Bedrock (`anthropic.claude-fable-5-2`), Claude Models on Google Cloud, Microsoft Foundry, and the Claude apps (with the plan limits above). Fast mode is on the Claude API, Google Cloud, and Microsoft Foundry for Anthropic-hosted deployments as `claude-fable-5-2-fast`. Mythos 5.2, the same capabilities with looser cyber and biology safeguards, is limited to verified organisations.

Carbon, as of 10–11 November 2026, is on the paid Gemini API and Google AI Studio (not the Free tier); on Google AI Ultra in the Gemini app, AI Mode, Antigravity and AI Studio; on Gemini for business; and on Google Cloud Agent Platform / Model Garden, including a "Gemini 4 Carbon Cyber Permissive" SKU at standard (post-intro) rates. Ultra is listed as available in more than 150 countries. No first-party listing for Carbon on Amazon Bedrock or Microsoft Azure was found on 11 November 2026.

## Who each suits

This is our reading of the sourced facts above, not a claim by Anthropic or Google.

**Claude Fable 5.2** suits teams that already live in Claude, need the top independent score on long, ambiguous technical work, and can absorb the higher token price — especially Max, Ultra or Team Premium seats, where Fable is included up to half the weekly limit, or API users who keep most daily work on Opus 5.5 and reserve Fable for the hardest jobs (Anthropic's own advice). Mind the 50% Fable cap, the 30-day retention requirement, cyber / biology / AI-R&D limits, and the prefix-check migration.

**Gemini 4 Carbon** suits teams that want frontier reasoning with multimodal input (including video and speech), 1M-token outputs, and native /swarm orchestration in Antigravity, at roughly one-fifth of Fable's MKA cost. The introductory $2 / $10 rate lasts only through 30 November 2026. Google itself still recommends Argon for most day-to-day engineering. Consumer access to Carbon is tied to Google AI Ultra; API access requires the paid Gemini tier.

*Facts as of Nov 11, 2026*

## Sources

- [Anthropic: Claude Fable 5.2 and Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2)
- [Claude docs: What's new in Fable 5.2](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2)
- [Claude docs: Pricing](https://platform.claude.com/docs/en/about-claude/pricing)
- [Claude docs: Release notes](https://platform.claude.com/docs/en/release-notes/overview)
- [Claude docs: Effort](https://platform.claude.com/docs/en/build-with-claude/effort)
- [Claude: Pricing](https://claude.com/pricing)
- [Claude Help Center: Claude Fable models on your plan](https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan)
- [Claude Help Center: Claude Fable 5.2 launch offers](https://support.claude.com/en/articles/17506494-claude-fable-5-2-launch-offers)
- [Claude Help Center: What is the Team plan?](https://support.claude.com/en/articles/9266767-what-is-the-team-plan)
- [Google: Gemini 4 Carbon now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)
- [Google AI Studio: Gemini 4 Carbon developer guide](https://aistudio.google.com/learn/gemini-4-carbon-developer-guide)
- [Google AI Studio: What's new in Gemini 4 Carbon](https://aistudio.google.com/docs/whats-new-gemini-4-carbon)
- [Google AI Studio: Pricing](https://aistudio.google.com/docs/pricing)
- [Google Cloud: Model Garden pricing](https://cloud.google.com/model-garden/pricing)
- [Google: Gemini subscriptions](https://gemini.google/subscriptions/)
- [Google One: Gemini and AI plans](https://support.google.com/googleone/answer/16286513)
- [Google Antigravity: plans](https://antigravity.google/docs/plans/)
- [Google Antigravity: Gemini 4 Carbon in Antigravity](https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity/)
- [Artificial Analysis: Multimodal Knowledge Agent](https://artificialanalysis.ai/articles/google-deepmind-multimodal-knowledge-agent)
- [Artificial Analysis: MKA leaderboard](https://artificialanalysis.ai/evaluations/mka)
- [Artificial Analysis: model release comparisons](https://artificialanalysis.ai/models/releases/comparisons)
- [Artificial Analysis: Gemini 4 Carbon, new number 1](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1)
- [@ClaudeDevs: compensatory reset](https://x.com/ClaudeDevs/status/2120275296637321476)
- [@demishassabis / Cursor on CursorBench](https://x.com/demishassabis/status/2120294666969248255)

## Related issues

- [2026-11-10](https://morningtokens.com/issues/2026-11-10/)
- [2026-11-11](https://morningtokens.com/issues/2026-11-11/)

Published Nov 11, 2026
