Google announced Gemini 4 Carbon on Wednesday, October 28, at about 1 PM ET, two weeks after Gemini 4 Argon went on general release, and the Gemini 4 Carbon price is already set: $2 per million input tokens and $10 per million output tokens at launch, the same introductory rate as Argon. After an introductory period Google hasn’t dated, it rises to $4 and $20. Google calls Carbon its strongest model, with “state-of-the-art cybersecurity and long horizon capabilities.”
The catch is that almost nobody can use it yet. Carbon isn’t on Google’s Gemini API price list or models page, and the Gemini app release notes don’t mention it. This week it goes to “a subset of cyber defenders” in Google’s vetted Fairwind Program, and developers, businesses and consumers get it “as soon as possible,” with no date.
- Price: $2 per million input tokens and $10 output at launch, cached input 95% off. Then $4/$20 once the introductory period ends (no end date given).
- Who can use it: only selected cyber defenders in Google’s Fairwind Program, starting this week.
- Free or in the Gemini app? Not yet. There’s no Carbon entry on the Gemini API pricing page, and Google hasn’t named plans or a date.
- The catch: Google’s comparison table mixes its own runs with rivals’ self-reported and leaderboard scores. One independent tester, Artificial Analysis, does put it at No. 1, but not in coding.
Gemini 4 Carbon price, and how it compares
Google’s launch post, written by Google DeepMind SVP Koray Kavukcuoglu, says Carbon “will launch at the same introductory price as Argon at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off.” A footnote adds that “after the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.” Artificial Analysis, which tested the model, describes it the same way: a 50% launch discount “with no confirmed end date.”
Argon’s discount does have a date. Google’s Gemini API pricing page lists Argon at $2/$10 through November 30, 2026, and $4/$20 from December 1. Don’t assume Carbon follows the same calendar until Google posts its price row. Here’s how the standard rates line up, per million tokens, from each company’s official price page:
| Model (standard, per 1M tokens) | Input | Output |
|---|---|---|
| Gemini 4 Carbon (introductory) | $2 | $10 |
| Gemini 4 Carbon (after intro period) | $4 | $20 |
| Gemini 4 Argon (to Nov 30 / from Dec 1) | $2 / $4 | $10 / $20 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.1 | $10 | $50 |
| GPT-6 Astra (short context) | $10 | $50 |
| GPT-6.1 Sol | $2 | $10 |
So at its full price Carbon costs the same per token as Claude Opus 5.5 and two-fifths of GPT-6 Astra. Per-token prices aren’t the whole bill, because models use very different numbers of tokens for the same job. Artificial Analysis measured $7.76 per task on its Intelligence Index at Carbon’s $4/$20 rate, about 23% less than Opus 5.5 at max settings ($10.04) but more than GPT-6 Astra ($6.25).
Who can use Gemini 4 Carbon right now
Google says it’s “taking a deliberate approach to rollout” and is “actively engaged in the U.S. government’s voluntary process for pre-release model testing.” Access starts with a subset of cyber defenders and then phases to “additional trusted partners” through Fairwind, while Google keeps iterating “on guardrails.” Sundar Pichai told people on X to “hang tight.”
Fairwind isn’t a waitlist you can join as an individual developer. Google’s Fairwind page says it prioritises governments and national cyber authorities, critical infrastructure operators such as healthcare, telecoms, energy and finance, and “core technology platforms.” Applicants get background checks. Members have to use phishing-resistant multi-factor authentication, can only give access to internal security, incident response or penetration testing teams, and can’t share or resell it. Google says it works with more than 650 partners.
The only timing clue is Argon. Google announced Argon for Fairwind defenders on September 30 and released it widely on October 13, 13 days later. That’s a precedent, not a promise.
The benchmarks: Google’s table and one independent check
Google’s table compares Carbon with Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 18 tests. Its methodology page says Carbon was run at its highest thinking setting and that rival scores are “sourced from providers’ self-reported numbers” unless noted, with several taken from public leaderboards. So this is Google’s compilation, not a same-lab test. Carbon has the best score in every row but two:
- FrontierSWE 2 (agentic coding): GPT-6 Astra 65.5%, Opus 5.5 62.3%, Carbon 61.2%.
- OSWorld-2.1, offline subset (computer use): Opus 5.5 87.2%, Astra 80.9%, Carbon 81.5%, which still puts it second.
The cybersecurity scores Google leads with, 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on Collinear’s CWE-bench for patching, were “evaluated without guardrails,” Google says. Google ran SEC-bench Pro and ExploitGym itself, in its own Antigravity harness, with ExploitGym’s mitigations off. CWE-bench numbers come from the benchmark’s public leaderboard. Google also says Carbon jumped 89 points to fourth place on WebDev Arena, at 1768.
The independent Artificial Analysis results back the headline and complicate the rest. Carbon scores 59 on its Intelligence Index v5, ahead of Opus 5.5 (58) and GPT-6 Astra (57), and leads 4 of the 9 tests in the index. But it’s second to Claude Code with Opus 5.5 on the firm’s Coding Agent Index (59 vs. 62), and behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on its AA-Briefcase knowledge-work test. Its hallucination rate on AA-Omniscience rose from Argon’s 4% to 11%, though Artificial Analysis says that’s still the second lowest of any model it has tested.
Google’s safety claims
Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0% (attempts to cheat fell to 0.4%), made about half as many unsupported claims as Argon in internal testing, and drew almost 80% fewer high-severity misalignment alerts in live coding with Google’s own teams. On Gray Swan’s indirect prompt injection benchmark, Google reports a 0.0% attack success rate, down from Argon’s 0.7%. Apart from Gray Swan’s benchmark, these are numbers from Google’s own testing. The launch post links an evaluation methodology document and points back to Argon’s safety and security report. Google has not posted a Carbon model card; the address that follows the pattern of Argon’s card returned a 404 on Thursday morning.
What to watch next
Three things decide when this matters to you. First, a gemini-4-carbon row on the Gemini API pricing page, which will show the real end date of the discount and any Batch or Priority rates. Second, which Google AI plans get Carbon in the Gemini app; Argon went to AI Pro and Ultra. Third, a model card, with the context window and knowledge cutoff. Artificial Analysis lists a 1 million-token context window and text, image, video and speech input, but Google hasn’t published those specs itself.
Frequently asked questions
How much does Gemini 4 Carbon cost?
Google says Carbon launches at $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After an introductory period, which Google hasn’t dated, the price becomes $4 per million input tokens and $20 per million output tokens.
Can I use Gemini 4 Carbon yet?
Probably not. As of October 29, 2026, Carbon is limited to selected cyber defenders in Google’s Fairwind Program. It isn’t listed on the Gemini API pricing page or available in the Gemini app, and Google has only said wider access is coming “as soon as possible.”
Is Gemini 4 Carbon better than Claude Opus 5.5 and GPT-6 Astra?
On Artificial Analysis’s independent Intelligence Index v5, Carbon scores 59 to Opus 5.5’s 58 and Astra’s 57. In Google’s own comparison table it trails Astra and Opus 5.5 on FrontierSWE 2 and Opus 5.5 on OSWorld-2.1, and Artificial Analysis ranks it second to Opus 5.5 on coding agents.
What is the difference between Gemini 4 Carbon and Gemini 4 Argon?
Carbon is Google’s newer frontier model, built on Argon and aimed at long-running agentic work and cybersecurity. Argon is generally available now in the API and Gemini app. Carbon has the same introductory API price but limited access, and it scores 9 points higher on Artificial Analysis’s Intelligence Index.
Sources
- Google, Koray Kavukcuoglu, “Gemini 4 Carbon: our next frontier model” (October 28, 2026; price, footnote, rollout, benchmarks, safety claims): blog.google
- Google DeepMind, Gemini 4 Carbon evaluation methodology: deepmind.google
- Google DeepMind, Fairwind Program (eligibility, terms, partner count): deepmind.google
- Google AI for Developers, Gemini API pricing (Argon introductory dates; no Carbon entry as of October 29): ai.google.dev
- Artificial Analysis, “Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5” (October 28, 2026): artificialanalysis.ai
- Anthropic, Claude API pricing (Opus 5.5, Fable 5.1): platform.claude.com
- OpenAI, API pricing (GPT-6 Astra, GPT-6.1 Sol): developers.openai.com
- Google, “Gemini 4 Argon: our next era of frontier intelligence” (September 30, 2026; Argon’s Fairwind-first rollout): blog.google
- Google, “Gemini 4 Argon: Built to solve your hardest problems” (October 13, 2026; general release): blog.google
- Google DeepMind and Sundar Pichai on X, Carbon launch posts (October 28, 2026): @GoogleDeepMind, @sundarpichai
All sources checked on the morning of October 29, 2026 (ET). Benchmark figures are Google’s or Artificial Analysis’s, as labelled.
Last updated: October 29, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
