Gemini 4 Carbon vs Claude Fable 5.2: Price and Benchmarks
Gemini 4 Carbon (gemini-4-carbon), GA since 10 November 2026, is Google's strongest reasoning model. It costs $2/$10 per million tokens until 30 November and $4/$20 from 1 December 2026, half of Claude Fable 5.2's $8/$40. Consumers need Google AI Ultra. Fable 5.2 leads WebDev Arena (1,823 vs 1,769); Carbon leads Text Arena. Checked 20 November 2026.
Rédaction assistée par IA : ce guide a été rédigé avec Claude Sonnet 5.5, et chaque fait renvoie à ses sources.
Gemini 4 Carbon is Google's strongest reasoning model, generally available since 10 November 2026 on the paid Gemini API and, in Google's apps, only on AI Ultra. It costs $2/$10 per million input/output tokens until 30 November 2026 and $4/$20 from 1 December, half of Claude Fable 5.2's $8/$40. Fable 5.2 leads most front-end boards; Carbon is cheaper per task.
What is Gemini 4 Carbon?
Gemini 4 Carbon (`gemini-4-carbon`) is what Google's developer docs call "our strongest reasoning model", built for long agentic work in software engineering, science and cyber defense. Google announced it on 28 October 2026 for a small group of cyber defenders and made it generally available on 10 November 2026. It outputs text only.
Its model card says Carbon "is based on Gemini 4 Argon". Neon, launched on 18 November, does not replace it: Google's Neon post says "Carbon tackles your hardest, long-horizon projects" while Neon "balances depth and efficiency".
How does Gemini 4 Carbon compare with Claude Fable 5.2?
From 1 December 2026 Carbon costs half as much per token as Claude Fable 5.2: $4/$20 against $8/$40 per million input/output tokens (Google, Anthropic), and a quarter as much until then. Fable 5.2 returns up to 128K tokens per synchronous request against Carbon's 65,536 and leads most front-end boards. Both have 1M-token context windows. Checked 20 November 2026.
| Gemini 4 Carbon | Claude Fable 5.2 | |
|---|---|---|
| Model ID | `gemini-4-carbon` | `claude-fable-5-2` |
| Released | GA 10 Nov 2026 | 9 Nov 2026 (announcement) |
| Price per 1M tokens | $2/$10, then $4/$20 from 1 Dec 2026 | $8/$40 (model page) |
| Cached input | $0.10, then $0.20 | $0.20 (pricing) |
| Context window | 1,048,576 tokens | 1M tokens |
| Max sync output | 65,536 | 128K |
| Thinking | Low, medium (default), high | Always on; high on the API, medium in Claude Code |
| Consumer plans | AI Ultra | Pro (usage credits), Max, Ultra, Team premium, Enterprise (pricing) |
| Agent tool | Antigravity | Claude Code, `/model fable` (docs) |
| Cloud | Agent Platform, verified enterprise | Claude API, Bedrock, Google Cloud, Foundry (overview) |
| AA cost per index task | $7.76 at high | $26.54 at max (AA) |
Antigravity also offers Claude Opus 5.5 and Sonnet 5.5, but not Fable 5.2, to paid Pro, Ultra and Enterprise users (models). For Fable's animation results and effort costs, see Claude Fable 5.2 for animation.
How much does Gemini 4 Carbon cost, and what are its specs?
On the Gemini API, Carbon costs $2 per million input tokens and $10 per million output tokens, thinking included, through 30 November 2026, then $4 and $20 from 1 December 2026, per Google's pricing page checked 20 November. It has no free tier and one flat price across its 1,048,576-token context window.
| Item | Gemini 4 Carbon | Source |
|---|---|---|
| Model ID | `gemini-4-carbon`, stable only, no preview alias | Model page |
| Status | GA 10 Nov 2026; no shutdown date | Deprecations |
| Price per 1M tokens | $2 input, $10 output, $0.10 cached to 30 Nov; $4, $20, $0.20 from 1 Dec 2026 | Pricing |
| Batch, Flex, Priority | Batch and Flex at half price; Priority $3.60/$18, then $7.20/$36 | Pricing |
| Free tier | None; only Hydrogen has one | Pricing |
| Context window | 1,048,576 tokens, input and output shared | Model page |
| Max output | 65,536 per synchronous request; about 1M with long decode | What's new |
| Thinking levels | Low, medium (default), high; no xhigh | Thinking |
| Modalities | Text, image, audio, video, PDF in; text out | Model page |
| Knowledge cutoff | March 2026 | Model card |
| Caching | Implicit only, from 4,096 tokens | What's new |
On Gemini Enterprise Agent Platform, the new name for Vertex AI, non-global endpoints cost $2.20/$11 now and $4.40/$22 from 1 December, and the November discount is paid as "50% credits back" (pricing). Non-streaming output there stops at 32,768 tokens per segment (model page).
Who can use Gemini 4 Carbon?
Anyone on the paid Gemini API tier can call Carbon. In Google's consumer products it is Google AI Ultra only, at $99.99 or $199.99 a month in the US per the subscriptions page: the Gemini app, AI Mode, AI Studio and Antigravity. On Agent Platform it is GA but limited to verified enterprise customers, as of 20 November 2026.
| Surface | Who gets Carbon | Source |
|---|---|---|
| Gemini API | Paid tier only | Pricing |
| AI Studio, incl. Build | Paid API tier or AI Ultra | Plans, guide |
| Gemini app | AI Ultra, picked by hand (Auto never routes to it); starts at Low thinking | Help |
| AI Mode in Search | AI Ultra, 18 and over | Help |
| Antigravity app, CLI, IDE | AI Ultra and Enterprise; CLI since v1.5.8, 13 Nov 2026 | Models, changelog |
| Agent Platform | Verified enterprise customers; global, us, eu | Model page |
| GitHub Copilot | Pro+, Max, Business, Enterprise | Docs |
| Cursor | Premium, Super, Ultra | Docs |
| OpenRouter | Slug `google/gemini-4-carbon` | Models |
| Gemini CLI | Not documented; README: "Not tuned for the Gemini 4 series" | Repo |
Is Gemini 4 Carbon on Google AI Pro?
No. As of 20 November 2026, Gemini Apps Help lists Carbon only for Ultra, AI Mode help names Ultra only, and Antigravity marks it unavailable on Pro. AI Pro gets Neon and Hydrogen. A plain-text read of Google's subscriptions page can suggest otherwise, because each "Includes Gemini 4" badge sits above the next plan's card.
What does Google say about Carbon's front-end taste and motion design?
Google's 10 November 2026 launch post says Carbon "checks both its code and the rendered UI" and comes "paired with improved front end taste and motion design skills in Google Antigravity". Its demos are Three.js builds, including an Eiffel Tower construction exhibition built by a /swarm team of Carbon agents and a 906K-triangle office reconstruction. These are Google's claims.
The skill is not tied to Gemini and installs in Claude Code too. Its motion rules ask for 140 to 240 ms ease-out transitions, never ease-in or `transition: all`, and `prefers-reduced-motion` support; it does not mention Three.js or GSAP. For Claude Code skills aimed at motion-heavy pages, see animated scroll websites with Claude Code.
Carbon outputs text, so its animations are code (HTML, CSS, SVG or Three.js) that you render afterwards; see how to turn an HTML animation into an MP4. Google says one demo, an Atlantis scene, started "from a video reference generated with Gemini Omni", a separate model.
How does Gemini 4 Carbon score against Claude on benchmarks?
Claude Fable 5.2 leads six of the nine independent boards below, checked 20 November 2026, including WebDev Arena, 1,823 against 1,769, where Carbon is #4 behind three Claude models. Carbon leads Text Arena and Design Arena's SVG board, and the two split ARC-AGI-3. Google's model card compares Carbon with Fable 5.1, not 5.2.
What do Google and Anthropic report?
Google runs Gemini at high thinking, while the API defaults to medium (evaluation notes). On its card Carbon scores 65.2% on Terminal-Bench 4.0 (an Artificial Analysis run) against 59.6% for Opus 5.5, but trails Opus 5.5 on FrontierSWE 2 (61.2% against 62.3%) and OSWorld-2.1 offline (81.5% against 87.2%).
Anthropic's Fable 5.2 announcement has a Carbon column but does not say where the numbers come from. Fable 5.2 leads FrontierSWE v2 (69.9% against 61.2%) and FrontierCode 2.0 (49.1% against 41.1%, Fable at xhigh, Carbon at medium); Carbon leads AutomationBench, 53.9% against 51.4%.
What do independent boards measure?
| Board | Gemini 4 Carbon | Claude Fable 5.2 | Ahead |
|---|---|---|---|
| AA Intelligence Index v5.0.1 | 60 at high, 58 at medium (API default); #4 by model | 65 at max; #1 | Fable 5.2 |
| Arena Text | 1,527, #1 of 426 | 1,503 at max, #7 | Carbon |
| Arena WebDev | 1,769 ±15, #4 of 162 | 1,823 ±16, #1 | Fable 5.2 |
| Agent Arena Index, 18 Nov | 1,303, #4 of 68 | 1,448, #1 | Fable 5.2 |
| Vals Index v2.2 | 70.45%, #4 | 71.31%, #1 | Fable 5.2 |
| Vals Vibe Code Bench v1.1 | 89.06%, #11 | 93.77%, #1 | Fable 5.2 |
| Terminal-Bench 4.1 official board | 64.5% | 65.8% (max, Claude Code) | Fable 5.2 |
| ARC-AGI-3 | 65.1% standard; 99.1% provider adapter | 81.2%; 97.7% | Fable 5.2 standard, Carbon adapter |
| Design Arena SVG | 1,394, #3 of 123 | Outside the visible top 15 | Carbon |
On WebDev Arena, Carbon's confidence range overlaps Sonnet 5.5's 1,771, and Arena lists only Carbon's high-thinking variant.
What did creators find when they tested Carbon against Claude?
In two scored same-prompt tests published by 20 November 2026, Carbon finished below Claude Fable 5.2 and cost far less: Pat Simmons reports $27.68 of API-equivalent spend against $883.30. These are individual creators' runs, not controlled benchmarks, and one review was written with early access from Google.
- Simon Willison gave Fable 5.2, Carbon and Neon one Three.js prompt, Yonkers rising as a giant mech and stomping across New York. On Hacker News he called Carbon's result "pretty great" after one follow-up, and he did not rank the three (his post).
- Pat Simmons (AI for Mortals, 11 November) ran Carbon in Antigravity against Fable 5.2 and Opus 5.5 in Claude Code and GPT-6 Astra in Codex on a website clone with "every animation", a launch film rendered to MP4 and a flight sim. Carbon placed fourth, fourth and third of four models.
- AICodeKing (video, 11 November) scored Carbon 53.3% against 70% for Fable 5.2 on six app builds, widest on a 3D room designer (6.3 against 9.2), at about $23 against $222. His advice: Fable "when the result has to work", Carbon "for cheap and fast first versions that you check yourself".
- Every (Katie Parrott, 10 November) discloses that "Google provided Every with early access" and says Fable 5.2 "has raised expectations for what a flagship should deliver".
What does motif's archive show about Gemini 4 Carbon?
Nothing yet: none of the 28 Gemini 4 posts in motif's archive on 20 November 2026 names Carbon, Hydrogen or Neon, and its 22 same-prompt Argon runs against Claude are covered in Gemini 4 Argon vs Claude Opus.
When is Gemini 4 Carbon worth it over Neon, Argon or Fable 5.2?
Carbon earns its place on long reasoning, debugging and agentic runs where cost per task matters more than first-pass visual polish. At 1 December 2026 list prices and each model's default thinking level, one Artificial Analysis index task costs $6.59 on Carbon, $3.74 on Neon and $8.63 on Argon; Fable 5.2 costs $26.54 at max.
| Model and thinking level | AA index score | Cost per task | Source |
|---|---|---|---|
| Gemini 4 Hydrogen, high (default) | 51 | $1.87 | AA Hydrogen |
| Gemini 4 Neon, high (default) | 60 | $3.74 | AA Neon |
| Gemini 4 Carbon, medium (default) | 58 | $6.59 | AA Carbon |
| Gemini 4 Carbon, high | 60 | $7.76 | AA Carbon |
| Gemini 4 Argon, medium (default) | 48 | $8.63 | AA Argon |
| Claude Fable 5.2, max | 65 | $26.54 | AA Fable 5.2 |
These are AA's figures at list price, except Argon's, which motif doubled because AA still shows its introductory price. Other boards agree on direction: Carbon costs $8.25 per Vibe Code Bench test against $22.13 for Fable 5.2, and $1.61 per median WebDev Arena task against $11.62.
- Carbon over Fable 5.2: long agentic or debugging runs you will review anyway. It is cheaper on every cost figure above, but creators found its claims about working UI need checking.
- Fable 5.2 over Carbon: front-end, 3D and motion work that must look right on the first pass, where Fable leads WebDev Arena and Vibe Code Bench.
- Neon over Carbon: most coding. Neon ties Carbon on WebDev Arena (1,763 against 1,769), scores 93.70% to 89.06% on Vibe Code Bench and costs about 43% less per AA task at defaults; see Gemini 4 Neon vs Claude Opus 5.5.
- Argon: same price as Carbon, a lower score and a higher default cost per task; GitHub Copilot retires it on 14 December 2026 with Carbon as the suggested replacement (changelog).
For the whole family, see which Gemini 4 model to use, compared with Claude; for the cheapest tier, Gemini 4 Hydrogen vs Claude Sonnet 5.5; for Claude alone, which Claude model for animation.
Questions
Is Gemini 4 Carbon free?
No. Gemini 4 Carbon has no free tier on the Gemini API; Hydrogen is the only Gemini 4 model with one. In Google's apps, Carbon needs Google AI Ultra, which costs $99.99 or $199.99 a month in the US. Checked 20 November 2026.
How much will Gemini 4 Carbon cost from 1 December 2026?
$4 per million input tokens, $20 per million output tokens and $0.20 per million cached tokens on the Gemini API, double the introductory $2, $10 and $0.10 that run through 30 November 2026. Batch and Flex stay at half the standard price.
Is Gemini 4 Carbon better than Claude Fable 5.2?
Not on most front-end boards. As of 20 November 2026, Claude Fable 5.2 leads WebDev Arena (1,823 against 1,769) and Vibe Code Bench (93.77% against 89.06%), while Carbon leads Text Arena. Carbon is cheaper: $4/$20 per million tokens from 1 December 2026 against $8/$40, and $6.59 per Artificial Analysis index task against $26.54 for Fable at max.
What is the Gemini 4 Carbon model ID?
`gemini-4-carbon`, on both the Gemini API and Agent Platform, listed as a stable version with no preview alias as of 20 November 2026. A separate `gemini-4-carbon-cyber-permissive` variant is allowlist-only through the Fairwind Program. OpenRouter uses the community slug `google/gemini-4-carbon`.
What are Gemini 4 Carbon's context window and maximum output?
Its context window is 1,048,576 tokens, shared by input and output. A normal synchronous request returns at most 65,536 tokens, and output near 1M needs long decode. Claude Fable 5.2 returns up to 128K tokens per request. Checked 20 November 2026.
Which thinking level does Gemini 4 Carbon use?
Medium by default on the Gemini API, with low and high also available and no xhigh. Google's published scores use high, so default calls score a little lower: Artificial Analysis rates Carbon 58 at medium and 60 at high, as of 20 November 2026.
Is Gemini 4 Carbon on Google AI Pro?
No. As of 20 November 2026, Google's Gemini Apps Help, AI Mode help and Antigravity model table list Carbon for AI Ultra only, at $99.99 or $199.99 a month in the US, and AI Pro gets Neon and Hydrogen. Any developer on the paid Gemini API tier can still call Carbon directly.
Has Gemini 4 Neon replaced Gemini 4 Carbon?
No. Neon, launched on 18 November 2026, sits beside Carbon: Google says Carbon "tackles your hardest, long-horizon projects" while Neon "balances depth and efficiency". Carbon has no shutdown date on the Gemini API as of 20 November 2026. GitHub Copilot retires Argon, not Carbon, on 14 December 2026 and suggests Carbon as the replacement.
Can Gemini 4 Carbon make animations or video?
It writes animation code, not video. Carbon outputs text only, so it builds HTML, CSS, SVG or Three.js animations that you render to MP4 afterwards. Google's 10 November 2026 demos are Three.js builds, and its demolab skill sets 140 to 240 ms transitions for interface motion.
Sources
- Google: Gemini 4 Carbon is now available
- Google: Gemini 4 Carbon, our next frontier model
- Google AI Studio docs: Gemini 4 Carbon
- Google AI Studio docs: What's new in Gemini 4 Carbon
- Google AI Studio docs: Pricing
- Google AI Studio docs: Thinking
- Google AI Studio docs: Deprecations
- Google AI Studio docs: Google AI plans
- Google AI Studio: Gemini 4 Carbon developer guide
- Google DeepMind: Gemini 4 Carbon model card
- Google DeepMind: Gemini 4 Carbon evaluation methodology (PDF)
- Google Cloud: Gemini 4 Carbon on Gemini Enterprise Agent Platform
- Google Cloud: Agent Platform generative AI pricing
- Google: Gemini subscriptions (US)
- Gemini Apps Help: Gemini models
- Google Search Help: AI Mode
- Google Antigravity docs: Models
- Google Antigravity docs: Changelog
- Google: Introducing Gemini 4 Neon
- google-gemini/demolab (gemini-frontend-design skill)
- google-gemini/gemini-cli
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2
- Claude Platform Docs: Claude Fable 5.2
- Claude Platform Docs: Pricing
- Claude Platform Docs: Models overview
- Claude Code Docs: Model configuration
- Claude: Plans and pricing
- Artificial Analysis: Models leaderboard
- Artificial Analysis: Gemini 4 Carbon
- Artificial Analysis: Gemini 4 Neon
- Artificial Analysis: Gemini 4 Hydrogen
- Artificial Analysis: Gemini 4 Argon
- Artificial Analysis: Claude Fable 5.2
- Arena: Text leaderboard
- Arena: WebDev leaderboard
- Arena: Agent Arena Index
- Vals AI: Vals Index
- Vals AI: Vibe Code Bench
- Terminal-Bench: Leaderboard
- ARC Prize: Leaderboard
- Design Arena: SVG leaderboard
- GitHub Docs: Supported AI models in Copilot
- GitHub Changelog: Upcoming deprecation of selected Copilot models in December
- Cursor Docs: Models and pricing
- OpenRouter: Models API
- Simon Willison: Gemini 4 Neon, Carbon and Hydrogen tests
- Simon Willison on Hacker News: Carbon Yonkers mech
- AI for Mortals (Pat Simmons): Gemini 4 Carbon review
- AICodeKing: Gemini 4 Carbon vs Claude Fable 5.2
- Every (Katie Parrott): Vibe check on Gemini 4 Carbon