Feature · Gemini 4 Carbon

Cheaper than Opus and GPT, and No. 1? Gemini 4 Carbon's price list and scorecard side by side: "half the input price, but you can't use it yet"

Google's Gemini 4 Carbon, released Oct. 28, ranks first on the Artificial Analysis Intelligence Index v5.0 with 59 points. Introductory pricing is $2 per million input tokens and $10 per million output tokens, half of Claude Opus 5.5 ($4 and $20). After the introductory period the prices match Opus, and there is no general release date. Using each company's official price page and only like-for-like benchmarks, here are eight comparisons. Rival scores follow Google's own comparison table, and cells we could not confirm are left blank.

1. What shipped: a frontier model two weeks after Argon, no general release date

   Google released Carbon two weeks after Gemini 4 Argon on Oct. 13, citing gains in coding, agentic tasks and finding security vulnerabilities. Access starts this week with some cyber-defense organizations and widens to partners in its Fairwind program. Developers, businesses and consumers get "coming soon," with no date.

2. Pricing ①: Carbon intro $2/$10; Opus 5.5 is $4/$20

   All figures are dollars per million input/output tokens from each company's official price page. Gemini 4 Carbon: 2/10 introductory, then 4/20, no public API. Gemini 4 Argon: 2/10 through Nov. 30, 4/20 from Dec. 1. Claude Opus 5.5: 4/20 (fast mode 8/40). Claude Fable 5.1: 10/50. At introductory prices Carbon's input rate is half of Opus 5.5's; at the later price it is equal.

3. Pricing ②: OpenAI and xAI range from $1 to $10

   GPT-6.1 Sol 2/10, GPT-6 Astra 10/50, GPT-6.1 Terra 1/5 (OpenAI developer pricing, inputs up to 272K tokens). Grok 4.7 is 2/6 (xAI docs, under 200K tokens). Simply adding one million input and one million output tokens gives: Carbon $12 introductory and $24 later, Opus 5.5 $24, Fable 5.1 $60, Astra $60, Sol $12, Terra $6, Grok 4.7 $8. That is token price only, not the cost of doing the same job.

4. Intelligence ranking: Carbon 59, Opus 5.5 58, Astra 57, Argon 50, Fable 5.1 49

   Artificial Analysis Intelligence Index v5.0: Carbon (High) 59, Opus 5.5 (max) 58, GPT-6 Astra (max) 57, Argon 50 and Fable 5.1 49 (the last two from Google's own comparison table). We could not confirm v5.0 scores for GPT-6.1 Sol or Grok 4.7. Arena puts Carbon first in Text Arena; on Vals Index v2.2 it is 70.4% to Opus 5.5's 70.2%, Argon's 69.2% and Astra's 65.0%. Artificial Analysis revised the index to v5 the same day, adding two benchmarks and upgrading three.

5. Coding and agents: where Carbon leads and trails

   In Google's table, Carbon scored 65.2% on Terminal-Bench 4.0 against Opus 5.5's 59.6%, Astra's 59.1%, Argon's 57.1% and Fable 5.1's 52.0%. On FrontierSWE 2, Astra led at 65.5%, then Opus 62.3%, then Carbon 61.2%; on OSWorld (offline) Opus led 87.2% to Carbon's 81.5% and Astra's 80.9%. SWE-sweep is a near tie (Carbon 9.6%, Astra 9.5%). On Artificial Analysis's Coding Agent Index, Claude Code with Opus 5.5 scored 62 to Carbon's 59 (VentureBeat). WebDev Arena has Carbon fourth at 1769.

6. Security benchmarks: scores only, all from Google's table

   From Google's comparison table. SEC-bench Pro (vulnerability finding): Carbon 88.2%, Astra 85.4%, Argon 83.4% (no entries for Opus or Fable). ExploitGym: Carbon 54.8%, Astra 42.4%, Argon 42.0%, Opus 5.5 34.5%. CWE-bench (security patches): Carbon 71.0%, Astra and Argon 68.0%, Opus 5.5 67.0%. Google said it ran the external security evaluations without guardrails, so results under real restrictions may differ. Some outlets report Google claiming more than 10 times better security patches than any publicly available model, which the table's margins do not show; it needs outside verification.

7. Price for performance: token price alone does not set the order

   Wccftech, citing Artificial Analysis, puts the cost of one index task at $7.76 for Carbon (about 123,000 output tokens), $10.04 for Opus 5.5 at max (about 173,000) and $6.25 for GPT-6 Astra at max. Astra, at five times Opus's token price, has the lowest cost per task. Carbon scores 59 at 22% less per task than Opus but 24% more than Astra, and that cost rises when $4/$20 applies. Artificial Analysis's hallucination-rate metric went from 4% for Argon to 11% for Carbon.

8. When and for whom: models you can use today are different

   Carbon opens first to cyber defenders, and Artificial Analysis lists it without a public API. Models usable at the listed prices today include Argon (public since Oct. 13), Opus 5.5, GPT-6.1 Sol and Grok 4.7. A post on X saying Pro, Ultra and API users can't use it drew 784 likes. Google has not said how long the introductory price lasts.

9. Developer reaction: "announcement-first" vs. "Google is back"

   Cline, the coding-agent company, claimed Carbon beats Opus 5.5 on nearly every benchmark at half the cost (628 likes). Andon Labs said Carbon is first on Vending-Bench 2 at $16,484 against Astra's $15,515 (974). The other side was loud too: one user wrote that they hate announcements weeks out from use and Fable 5.5 would arrive first (1,426), and another noted Google shipped before OpenAI and Anthropic's next models (1,033). The Fable 5.5 rumor is unconfirmed.