What Gemini 4 Carbon is, and when it came out
Google calls Carbon "our strongest model". It builds on Argon and is meant for complex multi-step coding, debugging and open-ended research. Google says it uses more tokens than Argon to explore and verify its work. It's one of four Gemini 4 models, all named after elements: Argon, Carbon, Hydrogen and Neon. See all Gemini 4 models for the whole family.
- October 28, 2026: announced by Koray Kavukcuoglu as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities". It went first to a subset of cyber defenders, then to partners in Google's Fairwind Program.
- November 10, 2026: generally available in the paid Gemini API, Gemini for business and CodeMender, with Antigravity, the Gemini app and AI Mode rolling out to AI Ultra subscribers.
- Release date, two answers: Artificial Analysis dates Carbon October 28 (the limited announcement); public API access started November 10.
- Since then: Hydrogen launched November 12 and Neon November 18. Carbon is still the model Google names for the longest, hardest tasks.
Gemini 4 Carbon specs
| Gemini 4 Carbon | |
|---|---|
| Announced | October 28, 2026 (cyber defenders); GA November 10, 2026 |
| Model ID | gemini-4-carbon |
| Context window | 1,048,576 tokens, one limit shared by input and output |
| Output per segment | Up to 65,536 tokens streamed (32,768 non-streaming) |
| Thinking levels | LOW, MEDIUM (default), HIGH |
| Google Cloud | global, us and eu. An allowlisted gemini-4-carbon-cyber-permissive variant (default HIGH) is available through the Fairwind Program. |
| API price per 1M tokens | $2 input, $0.10 cached, $10 output through November 30, 2026; $4, $0.20 and $20 from December 1. No free tier. |
Gemini 4 Carbon price
Gemini API Standard prices per million tokens, from Google's pricing page on November 21, 2026. Carbon's GA post says the new price applies "From November 30"; the pricing page says the promo runs through November 30 and the new price starts December 1, so we follow the pricing page. Carbon "is not available on the Free tier". Argon has the same price and the same promo. Claude Opus 5.5 prices are Anthropic's.
| Model | Input | Cached input | Output | Free tier (API) |
|---|---|---|---|---|
| Gemini 4 Carbon, through Nov 30 | $2 | $0.10 | $10 | No |
| Gemini 4 Carbon, from Dec 1 | $4 | $0.20 | $20 | No |
| Gemini 4 Argon | $2, then $4 | $0.10, then $0.20 | $10, then $20 | No |
| Gemini 4 Neon | $3 | $0.15 | $15 | No |
| Gemini 4 Hydrogen | $1.50 | $0.075 | $7.50 | Yes |
| Claude Opus 5.5 | $4 | $0.20 | $20 | - |
Where you can use Gemini 4 Carbon
- Gemini app: AI Ultra only (rolling out). Argon and Neon are on AI Pro and Ultra; Hydrogen is for everyone.
- Google Antigravity: AI Ultra. Google offers scientists 6 months free on Antigravity's top tier, by application.
- Gemini API and Google AI Studio: paid tier only, model ID
gemini-4-carbon. - Also: Android Studio, Gemini for business, CodeMender, and AI Mode in Search for AI Ultra subscribers.
- Google Cloud:
global,usandeu.
Gemini 4 Carbon benchmarks (Google's numbers)
These are Google's own figures from its October 28 and November 10 posts. Several November 10 claims are ranks without scores, and we haven't checked them on the benchmark owners' sites. Google's model card exists but we didn't read it, so nothing here comes from it.
| Benchmark | Gemini 4 Carbon | Reported by |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59, #1 (as of Oct 28) | Google, Oct 28 |
| SEC-bench Pro | 88.2% | Google, Oct 28 |
| ExploitGym | 54.8% | Google, Oct 28 |
| CWE-bench | 71.0% | Google, Oct 28 |
| WebDev Arena | #4, 1768 Elo | Google, Oct 28 |
| Internal secure-code benchmark | More than 2x Argon | Google, Oct 28 |
| CursorBench 4.0 | #1 (no score given) | Google, Nov 10 |
| Perplexity WANDR | #1 (no score given) | Google, Nov 10 |
| Zapier AutomationBench | #1, more than 2x lower cost per task than the next best external model | Google, Nov 10 |
| Multimodal Knowledge Agent (with Artificial Analysis) | #2, at more than 6x lower cost than #1 | Google, Nov 10 |
| Balyasny General Financial Intelligence | Top of 23 configurations | Google, Nov 10 |
Safety numbers Google reports
- Gray Swan indirect prompt injection: 0.0% attack success, against 0.7% for Argon.
- Gray Swan automated cyberattacks: 2.17%, against 9.86% for Argon. Per-attempt breach rate 0.8%, against 3.1%.
- Agentic honesty: 92%, against 70% for Argon.
- Terminal-Bench 4.0: 0.0% successful cheating and 0.4% attempted.
Gemini 4 Carbon on Artificial Analysis and Arena
Rankings move fast. Google quoted Carbon as #1 on Artificial Analysis with 59 on October 28; three weeks later the same index shows 60 and #8. Every rank below carries the date we read it.
- Artificial Analysis Intelligence Index, read November 21: 60 at high, 58 at medium, 54 at low; #8 of 95 models. Of the Gemini 4 models, only Neon scores higher (61 at xhigh).
- Artificial Analysis Coding Agent Index v2.1 (AA's Neon article, November 18): Carbon 55 at high, Neon 59 at xhigh.
- Cost per Intelligence Index task (AA): $7.76, the highest of the four Gemini 4 models. AA prices Carbon at the post-promo $4 / $20.
- Arena Text leaderboard, page dated November 20: Carbon (High) is #1 with 1527, ahead of Argon (1520) and Neon (1513). Arena was formerly LMArena.
Is Gemini 4 Carbon slow?
Yes, by design and by measurement. Google's model guide rates Carbon's speed "Deliberate", against "Moderate" for Neon and "Fast" for Hydrogen, and points it at multi-hour background runs. Google's models page describes Argon as having "faster responses and lower token spend per turn than Carbon".
- Output speed (Artificial Analysis): 55 tokens per second, the slowest of the four. Argon runs at 69, Neon at about 110, Hydrogen at about 161–166.
- Time to first token (AA, high): 71.3 seconds. Argon's is 44.5.
- Time per task (AA): 37.5 minutes per Intelligence Index task.
- More tokens: Google says Carbon uses more tokens than Argon to explore and verify.
- What to do about it: Google suggests tuning
thinking_levelbefore switching models. Carbon defaults to MEDIUM; LOW is faster and scores 54 on AA against 60 at HIGH. Google also suggests planning with Carbon or Neon and implementing with Hydrogen, passing context withprevious_interaction_id.
Gemini 4 Carbon vs Neon vs Argon
Prices are Gemini API Standard rates per million tokens. AA figures are from Artificial Analysis model pages read November 21 (Neon's rank from November 19). Arena from the Text leaderboard dated November 20.
| Carbon | Neon | Argon | |
|---|---|---|---|
| Google's one-liner | "our strongest model" | "balances depth and efficiency" | "Built to solve your hardest problems" |
| Announced / GA | Oct 28 / Nov 10, 2026 | Nov 18, 2026 | Sep 30 / Oct 13, 2026 |
| API price in / out | $2 / $10 to Nov 30, then $4 / $20 | $3 / $15 | $2 / $10 to Nov 30, then $4 / $20 |
| Thinking levels | low, medium (default), high | low, medium, high (default), xhigh | low, medium (default), high |
| Speed (Google's guide) | Deliberate | Moderate | Not on the guide |
| AA Intelligence Index | 60 (high), #8 of 95 | 61 (xhigh), #6 of 96 | 50 (high), #33 of 95 |
| AA Coding Agent Index | 55 (high) | 59 (xhigh) | - |
| AA cost per Index task | $7.76 | $6.44 | $4.97 |
| AA output speed | 55 tok/s | about 110 tok/s | 69 tok/s |
| Arena Text (Nov 20) | #1, 1527 | #3, 1513 | #2, 1520 |
| Gemini app | AI Ultra | AI Pro, Ultra | AI Pro, Ultra |
Carbon, Neon or Argon: which to pick
Google's choosing-a-model guide covers Hydrogen, Neon and Carbon; Argon isn't on it.
- Carbon: "multi-hour background runs, deep research, formal proofs, and security audits". Google says to escalate here when Neon at high or xhigh gets stuck.
- Neon: "multi-file refactoring, repository-wide debugging, and analytical tasks", when someone is waiting on the answer. On AA, Neon scores a point higher than Carbon, runs about twice as fast and costs less per task. From December 1 it's also 25% cheaper per token.
- Hydrogen: interactive web apps, routine code edits, document pipelines and parallel worker subagents. It's the cheapest and the only one with a free API tier. See Gemini 4 Hydrogen.
- Gemini 2.5 Pro and 3.1 Pro preview users: Google says to move to Neon or Carbon.
Gemini 4 Carbon and Argon
Argon came first and Carbon builds on it. They share a price, a promo, a context window and the same three thinking levels, so the choice comes down to depth against speed and cost.
- Google's own advice on November 10: Argon "remains the ideal choice for the majority of software engineering and knowledge work"; Carbon is for harder, open-ended tasks.
- Scores: Carbon leads Argon on AA's Intelligence Index (60 vs 50) and on Arena Text (1527 vs 1520, Nov 20).
- Cost and speed: Argon is cheaper per AA task ($4.97 vs $7.76) and faster (69 vs 55 tokens per second).
- Safety, per Google: Carbon's agentic honesty is 92% against Argon's 70%, and prompt-injection success is 0.0% against 0.7%.
- Access: Argon is in the Gemini app for AI Pro and Ultra and in Antigravity for Ultra; Carbon is Ultra-only in both. Google Cloud's docs name Neon as Argon's upgrade. See Gemini 4 Argon.
Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2
Only figures from one referee for all models, plus list prices. We have no Carbon numbers on benchmarks Anthropic reports for its own models, so there are no vendor-to-vendor rows. Artificial Analysis's Intelligence Index score for Opus 5.5 isn't in our sources, so it's left out. For Neon against Opus 5.5 on more benchmarks, see Gemini 4 Neon.
| Gemini 4 Carbon | Claude Opus 5.5 | Claude Fable 5.2 | |
|---|---|---|---|
| Arena Text (Nov 20) | #1, 1527 (High) | #5, 1506 (High) | #7, 1503 (Max) |
| AA Coding Agent Index v2.1 (Nov 18) | 55 (high) | 58 (max) | 58 (xhigh, in Claude Code) |
| AA Intelligence Index (Nov 18 order) | 60 (high) | - | Ranked above Neon (61) and Carbon |
| API price in / out per 1M | $2 / $10 to Nov 30, then $4 / $20 | $4 / $20 | - |
| Cached input | $0.10, then $0.20 | $0.20 | - |
| Context | 1,048,576 tokens, shared by input and output | 1M input, 128K output | - |
What that comparison shows
- Chat preference: on Arena's Text board, built from people's votes, Carbon is rated 21 points above Opus 5.5 and 24 above Fable 5.2 as of November 20. Arena's confidence intervals weren't readable when we checked, so treat small gaps with care.
- Agentic coding: on AA's Coding Agent Index, Opus 5.5 and Claude Code with Fable 5.2 both score 58, three points above Carbon's 55. Neon scores 59.
- Price: until November 30, Carbon costs half of Opus 5.5 per token. From December 1 the two cost the same, $4 / $20.
- The agent matters too. Opus 5.5 and Fable 5.2 run in Claude Code; Carbon runs in Google Antigravity and through the Gemini API. See Claude Code vs Codex for how the two main coding agents compare.
How to use Gemini 4 Carbon
Try it in Google AI Studio
Pick Gemini 4 Carbon in AI Studio. It isn't on the free tier, so you need a paid API key.
Call it from the Gemini API
Use the model ID
gemini-4-carbon. Thinking defaults to MEDIUM. Use HIGH for the hardest tasks and LOW when speed matters more.# Gemini API: model ID gemini-4-carbon (paid tier only) curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-carbon:generateContent" \ -H "x-goog-api-key: $GEMINI_API_KEY" \ -H "Content-Type: application/json" \ -d '{"contents": [{"parts": [{"text": "Audit this repository for injection bugs"}]}]}' # thinking levels: LOW | MEDIUM (default) | HIGHRun long jobs in Google Antigravity
On AI Ultra, choose Carbon in Antigravity for multi-hour background runs, and hand routine implementation to Hydrogen.
A new model is half of it. The other half is a job.
Models keep getting better and cheaper, and the agents around them still wait for the next prompt. Crewmark turns coding agent sessions into AI employees with a role, a goal in numbers, working hours and a heartbeat, so they take the next issue from a board on their own. Claude Code and Codex employees run natively; other agents, such as OpenCode, DeepSeek Harness and Gemini, connect through ACP (acpx). An employee uses whatever model its agent runs.
Pick a role
From the AI employee roles, such as an AI software engineer.
Put the work on a board
Assign issues, make the regular ones repeat, and let the heartbeat wake the employee to take the next one.
See what it can do
Jobs with setup and rules for each agent are in Claude Code use cases and Codex use cases.
FAQ
What is Gemini 4 Carbon?
Google's strongest Gemini 4 model, built on Argon for complex multi-step coding, debugging, open-ended research and security work. It has a 1M-token context and three thinking levels (MEDIUM by default). The model ID is gemini-4-carbon.
When was Gemini 4 Carbon released?
Google announced it on October 28, 2026, for a subset of cyber defenders, and made it generally available on November 10, 2026, in the paid Gemini API, with the Gemini app and Antigravity rolling out to AI Ultra subscribers.
How much does Gemini 4 Carbon cost?
In the Gemini API, $2 per million input tokens, $0.10 cached and $10 output through November 30, 2026, then $4, $0.20 and $20 from December 1. There is no free API tier. In the Gemini app and Antigravity it needs AI Ultra.
What are Gemini 4 Carbon's benchmark results?
Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench and #1 on CursorBench 4.0 and Zapier AutomationBench. Independently, Artificial Analysis shows 60 on its Intelligence Index (#8 of 95 on November 21), and Carbon is #1 on Arena's Text leaderboard with 1527 (November 20).
Why is Gemini 4 Carbon slow?
Google built it for multi-hour background runs and rates its speed "Deliberate". It uses more tokens than Argon to explore and verify. Artificial Analysis measures 55 tokens per second, a 71.3-second time to first token at high, and 37.5 minutes per Intelligence Index task. Lower thinking_level or use Neon or Hydrogen when speed matters.
Gemini 4 Carbon vs Neon: which is better?
On Artificial Analysis, Neon scores slightly higher (61 vs 60 on the Intelligence Index, 59 vs 55 on the Coding Agent Index), is about twice as fast and costs less per task. Carbon ranks higher on Arena Text (#1 vs #3). Google keeps Carbon for multi-hour runs and for tasks where Neon gets stuck.
Gemini 4 Carbon vs Argon?
Same price and context. Carbon scores higher (60 vs 50 on AA's Intelligence Index) but is slower and costs more per task. Google says Argon remains the right choice for most software engineering and knowledge work, with Carbon for harder, open-ended tasks.
Is Gemini 4 Carbon better than Claude Opus 5.5 or Fable 5.2?
On Arena's Text leaderboard (November 20), Carbon is #1 at 1527, Opus 5.5 #5 at 1506 and Fable 5.2 #7 at 1503. On Artificial Analysis's Coding Agent Index, Opus 5.5 and Claude Code with Fable 5.2 score 58 against Carbon's 55. Carbon costs half of Opus 5.5 per token until November 30 and the same from December 1.
Checked against Google's Carbon posts, the Gemini API pricing, models and choosing-a-model pages, Google Cloud's Carbon docs, Artificial Analysis and Arena on November 21, 2026.