Google’s Gemini 4 Carbon: From AI Assistant to Teammate
“I tossed out the requirements on my way home from work, and by the next morning a new library and a demo app were finished.”
This one developer’s account shows the new stage AI has reached. If artificial intelligence has so far stayed in the role of an assistant that answers short questions or writes fragments of code, it is now eyeing the place of a “teammate” that plans long-term projects on its own and sees them through to the end.
That is the core of Gemini 4 Carbon, the new large language model unveiled by Google. Going beyond one-off Q&A, it is optimized for agentic workflows that autonomously carry out long, complex tasks such as software engineering, demanding financial analysis and scientific reasoning.
Carbon differs in character from the previous model, Argon, which handles everyday work. In human terms, it is closer to a special-missions troubleshooter that takes on only the hardest problems. Carbon swept first place on major development and automation benchmarks, including CursorBench 4.0, WANDR and AutomationBench. Most striking is the cost per task. While posting top scores, it cut task costs to less than half those of competing models. In other words, it became smarter and cheaper at the same time.
To maximize this performance, Google also introduced /swarm, a technology in which multiple agents collaborate simultaneously. Given just a single line of instruction, Carbon takes charge of overall design and supervision, while lighter models are deployed in parallel to test and verify the output.
The capabilities confirmed in actual tests go beyond simple calculation. It interpreted more than 200 urban-planning drawings and PDF documents on its own to reconstruct a 3D spatial model based on more than 900,000 polygons, and in a complex codebase tangled with tens of thousands of tests, it caught 94% of bugs without breaking existing functionality. It also showed industry-leading quality on financial due-diligence reports full of complex tables and formulas and on sophisticated Excel modeling.
| Category | Benchmark | Notes | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|---|---|
| Knowledge work | Vals Index | 68.9% | 63.1% | 65.8% | 67.0% | |
| AutomationBench | Score | 51.3% | 41.4% | 31.4% | 42.5% | |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% | ||
| Harvey’s Legal Agent Benchmark | 19.6% | 5.4% | 6.7% | 3.8% | ||
| Agentic coding | DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% | |
| FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% | ||
| Vibe Code Bench | 91.9% | 89.6% | 90.3% | 90.3% | ||
| Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% | ||
| ML engineering | PostTrainBench | 45.3% | 44.3% | 40.2% | 49.3% | |
| Science and math | Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% | |
| LABBench 2 | 88.8% | 85.4% | 68.6% | 73.1% | ||
| RiemannBench | 76.0% | 72.0% | 65.6% | 69.6% | ||
| Long context | GraphWalks | Up to 128k, BFS (F1) | 99.7% | 98.7% | 91.4% | 90.6% |
| 256k to 1M, BFS (F1) | 84.2% | 71.8% | 65.0% | 66.8% | ||
| Computer use | Agent’s Last Exam | Pass rate | 39.5% | 34.2% | — | 38.2% |
| OSWorld-2.0 | Offline subset Partial score | 69.2% | 72.6% | — | — | |
| Multimodal understanding | Chartography | 71.6% | 71.0% | 46.2% | 66.3% | |
| LVBench | 91.7% | 87.5% | 79.7% | 83.7% | ||
| Cybersecurity | CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% |
*Methodology: deepmind.google/models/evals-methodology/gemini-4-argon
Industry is responding immediately to Carbon’s contextual understanding and execution, because it has reached into areas that once required heavy human effort, such as producing reliable recommendations in complex regulatory environments and catching subtle errors in 3D physics simulations on its own.
Gemini 4 Carbon will roll out first on the platform for paid developers and enterprise customers, followed by Google AI Ultra subscriber services across the board.
The era of wondering “how to make use of” AI is passing. Digital colleagues to whom complex projects can be entrusted wholesale are right in front of us. Now the question narrows to one: which problem will we hand to this powerful work partner first?
