Gemini 4 Carbon vs Claude Fable 5.2
The Short Version
These are the strongest models most developers can get in November 2026, and they reached developers a day apart. Claude Fable 5.2 has the better record on published coding and agent benchmarks, and Artificial Analysis’s independent index puts it a few points ahead. Gemini 4 Carbon is much cheaper per token, especially until its introductory price ends on November 30, accepts audio and video as well as text and images, and leads on Zapier’s business-automation benchmark.
Neither wins outright. Fable 5.2 is the safer starting point for hard software engineering. Carbon is the stronger candidate for automation across business apps, multimodal research, and teams already on Google’s stack, as long as you account for how many tokens it spends.
Since this was written: Google released Gemini 4 Neon on November 18, 2026, a mid-tier model at $3 / $15 per million tokens that scores 61 on Artificial Analysis’s Intelligence Index at xhigh, one point above Carbon at high. Google says Neon costs 32% less per task than Carbon. Carbon is still Google’s strongest model for the hardest work, and OpenAI’s GPT-6.2 Astra (62 on the same index) is now the other top contender.
At a Glance
| Gemini 4 Carbon | Claude Fable 5.2 | |
|---|---|---|
| Developer | Google DeepMind | Anthropic |
| Released | Announced Oct 28 (cyber defenders); paid API and Ultra from Nov 10, 2026 | Nov 9, 2026 |
| API model ID | gemini-4-carbon | claude-fable-5-2 |
| Context / max output | 1M / 65,536 by default, up to 1M with long decode | 1M / 128K |
| Input | Text, images, audio, video, PDF | Text and images |
| Reasoning control | Thinking levels: low, medium, high | Effort: low, medium, high, xhigh, max |
| API price (per 1M tokens) | $2 / $10 through Nov 30, 2026; $4 / $20 from Dec 1 | $8 / $40; fast mode $16 / $80 |
| Cached input | $0.10, then $0.20 (implicit caching only) | $0.20 cache reads; cache writes $10 (5 min) or $16 (1 hour) |
| Batch | Half price | Half price |
| AA Intelligence Index v5.0.1 | 60 (high) | 65 (max), 62 (high), 56 (medium) |
| AA cost per index task | $7.76 (high) | $26.54 (max), $9.08 (high), $3.80 (medium) |
| AA output speed | ~80 tokens/s | ~96 to 98 tokens/s |
| Consumer access | Google AI Ultra: Gemini app, AI Mode, Antigravity | Claude paid plans and Claude Code |
| Restricted sibling | Carbon for cyber, via the Fairwind Program | Claude Mythos 5.2, via Anthropic verification programs |
Artificial Analysis (AA) figures were read on November 12, 2026, from its model pages. Its Carbon page lists the $4 / $20 rate. Prices and specs are from Google’s Gemini API pricing and model pages and Anthropic’s model overview.
Two Different Philosophies of Effort
Gemini 4 Carbon is built to explore. Google says it spends more tokens than Gemini 4 Argon to try approaches and verify its work, and it bills those thinking tokens as output. On Artificial Analysis’s index it produced about 205 million output tokens at high thinking, one of the highest totals on the board.
Claude Fable 5.2 gives you a wider dial. Anthropic documents five effort levels and says most of the gain arrives by high, the API default; xhigh helps only on the hardest long-running coding and agent work, and max is mainly for measuring the model’s ceiling. The spread shows up in the numbers: at medium, Fable used about 41 million output tokens for an index score of 56; at max, about 275 million for 65. The Claude apps and Claude Code default to medium.
This is the most important thing to understand about cost. A per-token price tells you little until you know how much each model writes on your kind of task.
Coding and Agent Work
The published head-to-head numbers come mostly from the vendors. In Anthropic’s launch table, run at each model’s highest available setting:
| Benchmark (vendor-reported) | What it measures | Gemini 4 Carbon | Claude Fable 5.2 |
|---|---|---|---|
| FrontierCode 2.0 | Hard agentic coding tasks | 41.1% | 49.1% (xhigh); 45.3% at high |
| FrontierSWE v2 | Real-repository software engineering | 61.2% | 69.9% |
| AutomationBench (Zapier) | Business workflows across apps | 53.9% | 51.4% |
| Terminal-Bench-Science 0.1 | Scientific tasks in a terminal | 64.8% | 74.6% |
| OSWorld 2.1 (partial credit) | Operating a desktop computer | 81.5% | 89.3% |
Anthropic’s table marks Carbon’s FrontierCode result as a medium-thinking run, so that row is not a like-for-like comparison at the top setting.
Google’s own launch claims point the other way on some tests: it says Carbon ranks first on CursorBench 4.0 (coding tasks from Cursor), Perplexity’s WANDR (web research), and AutomationBench, each at less than half the cost per task of the next-best outside model, but it did not publish the scores. Anthropic reports 82.1% for Fable 5.2 on WANDR and lists no Carbon result. Treat every figure in this section as a vendor’s selection of tests, not as settled rankings.
The independent signal is Artificial Analysis’s index, which combines nine evaluations: Fable 5.2 scores 62 at high effort and 65 at max, Carbon 60 at high.
What It Costs to Finish a Task
On Artificial Analysis’s run, Carbon at high cost about $7.76 per task and Fable 5.2 at high about $9.08, for scores of 60 and 62. Fable at medium cost $3.80 for a score of 56, and Fable at max $26.54 for 65. If AA’s Carbon figure is computed at the $4 / $20 rate its page lists, the introductory price would roughly halve it until November 30.
To show how token prices turn into a bill, here is a hypothetical agent session: 3 million input tokens over many turns, of which 2.4 million are repeated context served from cache and 600,000 are new, plus 150,000 output tokens. The token counts are invented for illustration, and in practice the two models would not use the same number of tokens on the same task.
| Hypothetical session | Gemini 4 Carbon (to Nov 30) | Gemini 4 Carbon (from Dec 1) | Claude Fable 5.2 |
|---|---|---|---|
| 600K new input tokens | $1.20 | $2.40 | $4.80 |
| 2.4M cached input tokens | $0.24 | $0.48 | $0.48 |
| 150K output tokens | $1.50 | $3.00 | $6.00 |
| Total | $2.94 | $5.88 | $11.28 |
The sum leaves out Anthropic’s cache-write charges and any batch or fast-mode pricing. If Carbon writes three times as many output tokens as Fable on your task, its output line triples and much of the gap closes. The only reliable number is cost per completed task measured on your own work, including retries and the time someone spends fixing results.
How They Handle Five Common Jobs
Refactoring an application. Fable 5.2’s lead on FrontierSWE v2 and FrontierCode 2.0 makes it the default to try first, at high effort. Google positions Carbon for multi-package migrations, and in Antigravity its preview /swarm mode can split a large refactor across parallel agents in separate git worktrees with Carbon coordinating, which is a different way to attack the same job.
Debugging a difficult issue. Both are built for this. Google specifically names cross-subsystem root-cause investigations as a Carbon strength. Fable’s high effort tends to think before every reply, which helps on a hard bug and wastes tokens on a simple one. Give each the same failing test and logs and compare.
Reviewing a pull request. This rarely needs the top setting. Fable at medium or Carbon at a lower thinking level is usually enough, and a cheaper model such as Claude Opus 5.5, Sonnet 5.5, or Gemini 4 Argon may be the better value. Keep a human reviewer on anything that ships.
A multi-stage research task. Carbon accepts audio, video, and PDFs directly, so a brief built from recorded interviews and slide decks is simpler to feed it. Fable 5.2 needs text or images, but Anthropic’s WANDR result is strong. Both companies claim the lead on WANDR; neither claim is independently confirmed.
Building an automated workflow. AutomationBench, which tests workflows across business apps, is the one published benchmark where Carbon leads Fable in Anthropic’s own table. Combined with its lower token price, that makes Carbon worth testing first for business automation.
Availability, Regions, and Data
Carbon is on the paid tier of the Gemini API in “supported countries and territories,” through Google AI Studio, Android Studio, and Gemini for business, and is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode, and Antigravity. It is not on the free API tier, and Google has not published a country list for the rollout.
Fable 5.2 is on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and it is rolling out in Microsoft 365 Copilot’s Cowork and Copilot Studio and on GitHub Copilot’s Pro+, Max, Business, and Enterprise plans. One constraint matters for regulated teams: Anthropic requires 30-day data retention for Fable 5.2 and does not offer it under zero data retention unless expressly authorized.
Safety Boundaries
Both public models refuse some security and science work that their restricted siblings allow. Fable 5.2 blocks high-risk cyber activity such as exploit development and most research biology tasks, and passes those requests to an older Claude model; Anthropic says it blocks about 55% less often than Fable 5.1. Carbon carries Google’s guardrails, with a less-restricted cyber version gated behind the Fairwind Program.
Each company also published lab results: Google reports Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent and passed 92% of an agentic honesty test; Anthropic reports its Mythos 5.2 used planted credentials in 3.1% of deliberately weakened sandbox scenarios, against 21% for Mythos 5.1. These are controlled evaluations of specific behaviors, not guarantees about what either model will do in your environment. Keep an agent sandbox and approval steps in place regardless.
Which to Choose
Start with Claude Fable 5.2 if your work is hard software engineering in large codebases, you already use Claude Code, or you want fine control over effort and cost. Test it at medium and high before reaching for xhigh or max.
Start with Gemini 4 Carbon if cost per token matters, your inputs include audio, video, or long PDFs, you are automating work across business apps, or you build on Google AI Studio, Vertex, or Antigravity. Budget for the December 1 price increase now.
For most everyday work, neither is the right default. Claude Opus 5.5 and Gemini 4 Argon cost less and are what both companies recommend for the bulk of coding and knowledge work; see Claude Fable 5.2 vs Claude Opus 5.5 and model routing.
Related Comparisons
- Gemini 4 Carbon vs GPT-6.1 Sol
- Claude Fable 5.2 vs Claude Opus 5.5
- Claude Opus 5.5 vs Gemini 4 Argon
- Claude vs Gemini
Continue learning
Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.
See more AI tool comparisons
Browse all side-by-side AI tool comparisons on Ainanza.
Frequently Asked Questions
Is Claude Fable 5.2 better than Gemini 4 Carbon?
On most published coding benchmarks, yes, but not on all of them. Artificial Analysis's independent Intelligence Index v5.0.1 scores Fable 5.2 at 65 (max effort) and 62 (high) against 60 for Carbon (high). Anthropic's own launch table shows Fable ahead on FrontierCode 2.0 and FrontierSWE v2 and Carbon ahead on Zapier's AutomationBench. Google says Carbon ranks first on CursorBench 4.0, WANDR, and AutomationBench at lower cost per task, without publishing the scores.
Which is cheaper, Gemini 4 Carbon or Claude Fable 5.2?
Per token, Carbon by a wide margin: $2 per million input tokens and $10 per million output through November 30, 2026, then $4 and $20, against $8 and $40 for Fable 5.2. Per task the gap is smaller, because Carbon writes many more tokens. Artificial Analysis measured about $7.76 per index task for Carbon at high effort and $9.08 for Fable 5.2 at high effort, at the prices it lists.
Can I use Gemini 4 Carbon and Claude Fable 5.2 without the API?
Yes, on certain plans. Carbon is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search, and Antigravity; Google AI Pro gets Gemini 4 Argon instead. Fable 5.2 is on Claude's paid plans and in Claude Code, where it can use up to half of the weekly limit on Pro and standard Team seats.
Which model handles audio and video?
Gemini 4 Carbon. Its API accepts text, images, audio, video, and PDFs. Claude Fable 5.2 accepts text and images, so audio and video need to be transcribed or described first.
What are Mythos 5.2 and the Fairwind version of Carbon?
Both are less-restricted versions for vetted organizations. Claude Mythos 5.2 is the same model as Fable 5.2 with fewer cybersecurity and biology safeguards, available through Anthropic's Cyber Verification and Life Sciences Verification programs. Google keeps a cyber-specialized Carbon with fewer guardrails behind its Fairwind Program. Neither is available to ordinary developers.
Last updated: