Choose Argon or Carbon based on the reasoning depth the task requires: Argon is intended for most advanced software engineering and knowledge work, while Carbon is for tasks that call for especially deep reasoning and verification. For Argon, evaluate low, medium, and high on the same workload, then use the lowest level that meets your quality bar.
Suitable tasks: Choosing Argon or Carbon for coding, agent, math, science, and multimodal tasks in the Gemini API or AI Studio, and configuring Argon's thinking level.
Unsuitable tasks: Inferring fixed accuracy, latency, or token savings percentages from these qualitative guidelines, or assuming that Argon should be used for every task.
Applicable model versions: Gemini 4 Argon (gemini-4-argon) and Gemini 4 Carbon (gemini-4-carbon); the level settings apply only to Argon.
Applicable clients, agents, or APIs: Gemini API and Google AI Studio. The page does not limit this workflow to a specific SDK.
Recommended reasoning levels and parameters: Argon defaults to medium; low, medium, and high are available. Evaluate all three on the target workload and choose the lowest level that meets your quality bar.
The following editorial decision table is based on the official model-selection guidance. It is not a complete prompt from Google or an independent benchmark result:
| Task characteristics | Evaluate first | Official rationale |
|---|---|---|
| Everyday feature development, interactive coding, rapid production iterations; most advanced engineering or knowledge work | Argon | Intended for most advanced software engineering and knowledge work; faster per turn and uses fewer tokens than Carbon |
| Complex root-cause debugging, architecture design, dense multimodal synthesis, scientific proofs, security audits | Carbon | Intended for tasks requiring deeper reasoning; spends more thinking tokens on exploration and verification |
| Simple or cost-sensitive Argon tasks | Argon low | Lowest token use and latency; quality on hard problems may decline |
| Routine tasks, including complex coding and agent scenarios | Argon medium | Default level, recommended by Google for most tasks |
| The hardest math, science, coding, or agent tasks | Argon high | Uses the most tokens and has the longest time to first token; suited to the hardest tasks |
Classify the actual workload as routine delivery or a task requiring especially deep reasoning and verification. Evaluate Argon first for the former; for the latter, evaluate Carbon as shown in the official examples.
Define pass criteria for representative tasks, such as passing tests, a verifiable root-cause explanation, proof steps that satisfy the constraints, or complete audit coverage. Pass criteria should come from the application itself; this article does not set score thresholds for a task.
For Argon, keep the same inputs, tools, and acceptance criteria, run low, medium, and high separately, and record whether each passes, its latency, and token use.
Choose the lowest level that meets the quality bar. If the highest level still falls short, adjust task decomposition or the verification process, then reassess the model choice; the official page does not provide a numerical threshold for switching to Carbon.
For especially complex tasks that depend on deep reasoning, include Carbon in the same evaluation and compare actual results. Do not treat the model descriptions as test results on your team's tasks.
These evaluation steps are editorially developed operational recommendations. The official page explicitly recommends evaluating Argon's three levels and choosing the lowest one that meets the quality bar. The user determines the specific metrics to record, pass criteria, and whether to compare Carbon.
Google describes Argon as suitable for most advanced software engineering and knowledge work; compared with Carbon, it responds faster per turn and uses fewer tokens for interactive coding, everyday feature work, and rapid production cycles.
Google positions Carbon for tasks requiring deeper reasoning, citing complex root-cause debugging, architecture design, dense multimodal synthesis, scientific proofs, and security audits. It says Carbon uses more thinking tokens per turn for deeper exploration and verification, at the cost of higher token use and latency per turn.
Google says Argon and Carbon share the same unified 1M-token context window, built-in tools, and per-token pricing. The page also lists a limited-time introductory price for Argon; that price does not change the task-quality and latency evaluation in this selection workflow.
Google defines Argon's low level as having the lowest token use and latency, with potentially lower quality on hard problems; medium as the balanced default, recommended for most tasks; and high as using the most tokens and having the longest time to first token, suited to the hardest tasks.
Google recommends evaluating all three thinking levels and then choosing the lowest one that meets the quality threshold. The page provides no specific test data, quantitative thresholds, or task-level win rate for Argon versus Carbon.
The task positioning for Argon and Carbon is official qualitative guidance. Results need to be evaluated for the specific codebase, tools, data, and acceptance criteria.
The page says both models have the same per-token pricing, but does not claim that total cost per task is the same. The statement that Argon uses fewer tokens per turn is Google's description, not an independent measurement in this article.
The low, medium, and high level descriptions apply to Argon. This article does not infer available levels or defaults for Carbon.
The page specifies no numerical threshold for switching between Carbon and Argon, and does not say that choosing Carbon will necessarily improve the quality of any given task.
Output limits, background long decoding, sampling parameters, and prompting techniques are covered by other official materials; this article does not expand on those settings.
Fix a representative set of tasks, input versions, tool permissions, and acceptance criteria. For Argon, record pass status, latency, and token use at each of the three levels. Then run Carbon through the same process for representative tasks that require deeper reasoning. Report each task's results and environment; do not generalize from a single task to a universal model ranking.
Gemini 4 Argon