Gemini 4 Carbon vs Gemini 4 Argon: Google Beats Its Own Model in Two Weeks
Google's latest flagship model, Gemini 4 Carbon, is here, just two weeks after Gemini 4 Argon became available to everyone. Carbon posts significant improvements across the board, and here is how it compares with Gemini 4 Argon.
Google benchmarks
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | Difference |
|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 50 | Carbon +9 |
| Vals Index v2.2 | 70.4% | 69.2% | Carbon +1.2 points |
| SWE-sweep | 9.6% | 4.3% | Carbon +5.3 points |
| FrontierSWE 2 | 61.2% | 55.0% | Carbon +6.2 points |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | Carbon +8.1 points |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | Carbon +8.4 points |
| OSWorld-2.1 (offline subset) | 81.5% | 74.7% | Carbon +6.8 points |
| AutomationBench (Zapier) | 53.9% | 51.3% | Carbon +2.6 points |
| Vending-Bench 2 (bank balance) | $16,484 | $13,718 | Carbon +$2,766 |
| Chartography | 72.5% | 71.6% | Carbon +0.9 points |
| Blueprint-Bench 2 | 58.6% | 54.4% | Carbon +4.2 points |
| GraphWalks 256K to 1M | 96.2% | 84.2% | Carbon +12.0 points |
| RiemannBench | 80.0% | 76.0% | Carbon +4.0 points |
| Terminal-Bench Science 0.1 (AA) | 64.8% | 52.9% | Carbon +11.9 points |
| CWE-bench v1 | 71.0% | 68.0% | Carbon +3.0 points |
| ExploitGym | 54.8% | 42.0% | Carbon +12.8 points |
| SEC-bench Pro | 88.2% | 83.4% | Carbon +4.8 points |
Artificial Analysis benchmarks
| Benchmark | Gemini 4 Carbon (High) | Gemini 4 Argon (High) | Difference |
|---|---|---|---|
| Intelligence Index v5.0 | 59 | 50 | Carbon +9 |
| AA-Briefcase v1.1 | 1609 Elo | 1483 Elo | Carbon +126 Elo |
| AutomationBench-AA | 78% | 78% | Tie |
| Terminal-Bench 4.0 | 65% | 57% | Carbon +8 points |
| FrontierCode 2.0 | 41% | 26% | Carbon +15 points |
| AA-Omniscience v1.1 | 54 | 46 | Carbon +8 |
| GDP.pdf | 28% | 22% | Carbon +6 points |
| HLE-Diamond | 69% | 57% | Carbon +12 points |
| CritPt v2 (Hard Subset) | 63% | 53% | Carbon +10 points |
| Terminal-Bench Science 0.1 | 65% | 53% | Carbon +12 points |
Cost and speed
| Metric | Gemini 4 Carbon (High) | Gemini 4 Argon (High) | Difference |
|---|---|---|---|
| Cost per Intelligence Index task | $7.76 | $7.68 | Argon about 1% cheaper |
| Cost to run the Intelligence Index | $8,240 | $6,103 | Argon about 26% cheaper |
| Output tokens per task | 123K | 116K | Argon about 6% fewer |
| Reasoning tokens per task | 72K | 68K | Argon about 6% fewer |
| Output tokens to run the Index | 205M | 113M | Argon about 45% fewer |
| Output speed | 92 tokens/s | 82 tokens/s | Carbon about 12% faster |
| Time to first token | 100.31 s | 44.96 s | Argon starts answering sooner |
Gemini 4 Carbon vs GPT-6 Astra vs Claude Opus 5.5
| Benchmark | Gemini 4 Carbon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 57 | 58 |
| Vals Index v2.2 | 70.4% | 65.0% | 70.2% |
| SWE-sweep | 9.6% | 9.5% | 3.6% |
| FrontierSWE 2 | 61.2% | 65.5% | 62.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 59.1% | 59.6% |
| Agents' Last Exam v1.1 | 77.4% | 73.1% | 71.6% |
| OSWorld-2.1 (offline subset) | 81.5% | 80.9% | 87.2% |
| AutomationBench (Zapier) | 53.9% | 41.4% | 42.5% |
| Vending-Bench 2 (bank balance) | $16,484 | $15,515 | $9,235 |
| Chartography | 72.5% | 71.0% | 66.3% |
| Blueprint-Bench 2 | 58.6% | 49.7% | 51.2% |
| GraphWalks 256K to 1M | 96.2% | 71.8% | 66.8% |
| RiemannBench | 80.0% | 72.0% | 69.6% |
| Terminal-Bench Science 0.1 (AA) | 64.8% | 63.3% | 59.0% |
| CWE-bench v1 | 71.0% | 68.0% | 67.0% |
| ExploitGym | 54.8% | 42.4% | 34.5% |
| SEC-bench Pro | 88.2% | 85.4% | - |
Leaderboard rankings
| Leaderboard | Gemini 4 Carbon | Rank |
|---|---|---|
| AA Intelligence Index v5.0 | 59 | #1 |
| Vals Index v2.2 | 70.36% | #2 of 29 |
| Vals Legal Research Bench | 55.77% | #1 of 87 |
| Vals Finance Index | 66.35% | #1 of 33 |
| Vals Terminal-Bench 4.0 | 63.13% | #4 of 57 |
| WebDev Arena | 1768 Elo | #4Argon: #11, 1679 Elo |
Pricing and specs
| Specification | Gemini 4 Carbon | Gemini 4 Argon | Difference |
|---|---|---|---|
| Input, per 1M tokens (introductory) | $2 | $2 | Same |
| Cached input, per 1M tokens (introductory) | $0.10 | $0.10 | Same |
| Output, per 1M tokens (introductory) | $10 | $10 | Same |
| Input, per 1M tokens (standard) | $4 | $4 | Same |
| Cached input, per 1M tokens (standard) | $0.20 | $0.20 | Same |
| Output, per 1M tokens (standard) | $20 | $20 | Same |
| Introductory price ends | Not announced | November 30, 2026 | Different |
| Context window | 1M tokens | 1M tokens | Same |
| Max output | Not published | 1M tokens | Different |
| Reasoning | Yes | Yes | Same |
| Public API | Not yetCyber defenders only | Yesgemini-4-argon | Argon |
| Announced | October 28, 2026 | September 30, 2026Available to all on October 13 | Carbon 15 days after Argon's wide release |
Who can use Gemini 4 Carbon
Not everyone yet. Google is giving Carbon to a small group of cyber defenders first and will slowly add more trusted partners through its Fairwind Program. Google says it is coming soon to developers, businesses and consumers, but it has not given a date. Gemini 4 Argon, on the other hand, is available now in the Gemini API, Google AI Studio and the Gemini app for Google AI Pro and Ultra subscribers.