Gemini 4 Carbon vs Gemini 4 Argon vs Gemini 3.8 Flash

In November 2026, Gemini 4 Carbon is Google's model for tasks that need more reasoning, while Gemini 4 Argon is its recommendation for most coding and knowledge work. Through November 30, Carbon and Argon cost $2 per million input tokens and $10 per million output tokens; Gemini 3.8 Flash costs $0.75 and $3.75. Google opened Carbon to paying API users on November 10.

Carbon scored higher than both models on an independent intelligence benchmark. Whether it is the right choice depends on its cost per successful task, which includes the extra tokens it uses.

Where each model fits

Carbon is based on Argon, according to Google's model card. Google positions Carbon for tasks that test the limits of a model's intelligence, while its launch post recommends Argon for most software engineering and knowledge work.

The scores come from Artificial Analysis: its Carbon vs Argon comparison and Argon vs Flash comparison. They measure a benchmark suite, so test your own tasks before moving traffic.

Google's developer guide suggests using Carbon to plan and verify work, then sending parallel implementation slices to Argon or Flash. Google also reported an 88.2% score for Carbon on SEC-bench Pro, a benchmark for finding software vulnerabilities.

What do the APIs cost?

These are Google's listed prices per million tokens in November 2026. The later rows show the price after each introductory window.

On Requesty's model list, Carbon's cached input costs $0.10 per million tokens. Google Cloud's Carbon documentation says Carbon supports implicit caching with a 4,096-token minimum; explicit context caching is not supported.

Google's sources cover the Carbon price change, Argon's matching prices and Flash pricing. Our guide to LLM list prices looks at how promotional windows affect comparisons.

Token price is only one part of the request bill. Google says Carbon uses more tokens than Argon, so compare the bill for completed work alongside the price list.

Requesty's pricing adds 5% on top of model costs. The worked example below uses Google's listed model prices alone.

What does a successful task cost?

Take a batch of 1,000 tasks. Say every request uses 10,000 uncached input tokens and 2,000 billable output tokens, regardless of model or outcome. That gives the batch 10 million input tokens and 2 million output tokens.

At November prices, Carbon and Argon each cost $40 for the batch: $20 for input and $20 for output. Flash costs $15, split into $7.50 for input and $7.50 for output. From December 1, the same Carbon or Argon batch costs $80.

Now say Argon solves 80% of the tasks and Carbon solves 90%. Argon's November cost per solve is $40 / 800, or $0.05, while Carbon's is $40 / 900, or about $0.044. Carbon wins this example even though both models cost the same per request.

Equal token usage is the weak point of this example. If Carbon uses twice as many tokens for the same batch, its cost rises to $80 and its cost per solve to about $0.089. Matching Argon's $0.05 per solve at that cost would require 1,600 solves out of 1,000 tasks, which is impossible within this batch.

The success rates and token counts here are examples, not benchmark results. Run the numbers with your own accepted answers, retries and token usage. The December price increase changes the budget for both Carbon and Argon, even if their relative price stays the same.

Argon has a measured coding result

Requesty's coding benchmark put Argon through 129 private Go backend tasks with the same agent and an 80-turn budget. In the October 20, 2026 run, it solved 88.4% of tasks, ranked third out of 25 models and passed 95.1% of tests.

Cost per solve includes the effect of unsuccessful tasks. Use Argon's result as the baseline and measure Carbon against it on your own coding tasks before moving traffic; our coding model comparison has more benchmarks across providers.

Thinking level changes the wait

Artificial Analysis timed the first answer token for all three models at high thinking, and for Carbon at medium and low thinking. Carbon's setting changes both the wait and the benchmark score.

The figures come from Artificial Analysis comparisons of Carbon vs Argon and Argon vs Flash. These are benchmark measurements, so your prompts and output lengths can produce different waits.

Carbon supports LOW, MEDIUM and HIGH thinking levels, with MEDIUM as the default. Google Cloud's documentation says thinking_budget and MINIMAL return HTTP 400, while custom temperature, top_p and top_k settings are ignored.

Carbon has a unified limit of 1,048,576 tokens across input and output, including thinking tokens. A single foreground segment can generate up to 65,536 output tokens when streaming, or 32,768 when not streaming.

Which model should you choose?

Try Carbon behind the same API

Requesty's model list lists all three models behind its OpenAI-compatible API. Point your client at https://router.requesty.ai/v1, then choose the model ID:

Change the model ID to run the same workload on each model. Compare accepted results, token cost and wait time before deciding how much traffic to move.

Requesty supports fallback policies, so Argon or Flash can serve as a fallback for Carbon. Test the fallback against your quality threshold too: a successful fallback response still needs to solve the task.