SYLIN

Gemini Carbon

by Google · Model · Language & code models

mixed

Strong benchmarks and best at hard reasoning, but slow, token-hungry and uneven hands-on.

#1 chat#1 fewest hallucinations

Free Carbon access expanded to 10,000 scientists through Antigravity

AccessAPIPrice$2 in · $10 out per M tokensContext1,049K tokens

Why we say so

Artificial Analysis ranks it second on its Intelligence Index, MathArena put it first, and Vending-Bench and Blueprint-Bench rank it first. AI Search found it worse than Fable 5.2 on coding and computer use. Hacker News users report slow first tokens, OpenRouter timeouts and wasted tokens on simple tasks, though some prefer it for architecture and root-cause analysis. Price is $2/$10 per million tokens until the promo ends, then $4/$20.

Ranks #2 for knowledge work · #3 for image understanding · #3 for math reasoning

Evidence

  1. Nov 19
    Reviewer unimpressed: water-balloon physics code needed two iterations and looked poor; highest setting in Antigravity very slow, often over an hour per test
  2. Nov 18
    Argon preferred for most coding tasks; Carbon preferred for architecture and deep root-cause analysis of complex issues
  3. Nov 18
    On long complex tasks it did better than Hydrogen in similar time at about 50% more cost; on simple tasks too slow and costly
  4. Nov 18
    One commenter found it far ahead of Astra Ultra, Opus 5.5 and Fable 5.1 on their research questions (anecdotal)
  5. Nov 18
    Commenter says Sol makes more mistakes than Carbon and Opus 5.5, and Astra burns subscription limits faster than Carbon
Show all 19
  1. Nov 15
    Released for long-horizon agentic work via Gemini API, AI Studio, Android Studio; Ultra subscribers get it in Gemini app and Antigravity
  2. Nov 15
    Ranks #3 on Vals Index about 1pp behind Fable 5.2, at $10.51 vs $17.94 per test; #1 on Finance, Legal, Tax
  3. Nov 15
    In a procedural Unreal Engine ninja-game test, worse than Fable 5.2 (wobbly animation, weaker lighting) and slightly behind GPT-6 Astro
  4. Nov 15
    Ranked #1 on Vending-Bench (highest balance), and #1 on Agents' Last Exam when used with Google's harness
  5. Nov 15
    Ranked #1 on Blueprint-Bench, tied with the human baseline
  6. Nov 14
    Took #1 on MathArena, displacing GPT-6.1 Sol
  7. Nov 11
    Tops CursorBench at 60.4% on high effort
  8. Nov 11
    Cost advantage over GPT-6 Astra and Opus 5.5 widens on the most expensive agent sessions
  9. Nov 11
    #3 on Arena agent
  10. Nov 11
    #3 on Arena text (1532)
  11. Nov 11
    #1 on Arena text (1538)
  12. Nov 1
    Google's Gemini 4 Carbon, claimed frontier on agentic coding, computer use, science and cyber; only for an initial cohort of cyber defenders
  13. Nov 1
    Ranked first on Artificial Analysis index, one point above Claude Opus 5.5 and GPT-6 Astra, and cheaper than Opus
  14. Nov 1
    Top accuracy and lowest hallucination rate on the Omniscience benchmark

Cost to run

Alternatives

GPT AstraRebuilt an Airbnb property in Blender with more faithful details (pillows, robes) than Gemini Carbon and Claude Fable, though exterior was wrong
Claude FableGPT, Gemini and open-weight LLMs score near zero gap; on 56 post-cutoff proteins Fable 5.2 got +22.1, GPT-6.2 Sol -0.4
Gemini NeonAt high effort Neon matched Carbon's 60 on the Intelligence Index at $3.74 per task
Claude Opus#3 on Arena text (1543)
Gemini Argon#2 on Arena text (1532)
GPT SolIn Codex Ultra, its Yonkers-to-mech three.js animation was weaker at morphing the city into a mech than Fable 5.2 and Gemini 4 Carbon
Built on Gemini Carbon