› On the day Gemini 4 Carbon became available to everyone (10 November 2026) I ran it through more than 50 tasks and put the same tasks to Claude Fable 5.2, GPT-6.1 Sol and Gemini 4 Argon.
› Every call went through OpenRouter with the same settings, and every answer was graded by unit tests, an answer key or an automatic checker - not by my impression.
› Here is what Carbon can do, what it costs and where it trips up.
› Prices first Gemini 4 CarbonGemini 4 ArgonClaude Fable 5.2GPT-6.1 Sol API price per 1M tokens (in / out)2 / 10 USD until 30 Nov, then 4 / 202 / 10 USD until 30 Nov, then 4 / 208 / 40 USD2 / 10 USD Context window1,048,5761,048,5761,000,0001,050,000 Results on the same tasks TaskCarbonArgonFable 5.2GPT-6.1 Sol 23 practical tasks (low effort)99.5%*, 0.23 USD, 264 s99.5%*, 0.22 USD, 135 s100%, 1.37 USD, 190 s100%, 0.22 USD, 180 s Agent: 6 documents with a trap, table, Python check10/10, 54 s10/10, 38 s10/10, 31 s10/10, 20 s Spreadsheet engine from scratch, 34 unit tests34/34, 0.47 USD, 518 s34/34, 0.05 USD, 64 s34/34, 0.82 USD, 201 s34/34, 0.05 USD, 77 s Security review of a Flask app, 8 planted flaws8/88/88/88/8 Web server log after a break-in (SOC analysis)7/77/77/77/7 Patching a vulnerable module, 14 tests (medium)0/14 (indentation error)14/1414/143/14 (Windows-only issue) 6 coding tasks, 76 unit tests (medium)63/7676/7676/7676/76 Made-up sources: 16 scientific papers0 invented, 2 minor errors2 invented0 invented, 0 errors0 invented, 1 error Vision: 10 questions on a chart, a table and a receipt10/1010/1010/1010/10 * The grader deducted one point for a correct grammatical form - its own mistake, the same for both Gemini models.