Summarize this article with:
Gemini 4 Carbon scores highest on the Artificial Analysis Intelligence Index reported by Requesty. Argon has the stronger published workload result, and Google says Carbon spends more thinking tokens per turn. Carbon and Argon share a $2 / $10 price through November 30, 2026, then both move to $4 / $20, so the right tier depends on cost per accepted task, not token price alone.
Google made Gemini 4 Carbon available to all paying API developers and enterprises on November 10, 2026. Two weeks earlier, it had released Carbon to cybersecurity defenders through its Fairwind Program. Carbon and Gemini 4 Argon share an introductory price of $2 / $10 per 1M input / output tokens through November 30, then both rise to $4 / $20 on December 1.
Carbon scores higher on the Artificial Analysis Intelligence Index, as reported by Requesty. Argon has the stronger published workload result, and Google says Carbon spends more thinking tokens per turn. Equal token prices do not mean equal task costs.
This comparison covers Carbon, Argon and Gemini 3.8 Flash on price, evidence, thinking levels, caching and timeouts, then gives a starting tier per workload. For Carbon's full launch benchmarks, see our Gemini 4 Carbon release analysis.
Carbon, Argon and Flash compared
Flash's $0.75 / $3.75 per 1M input / output tokens is Google's listed price through December 31, 2026, according to Requesty's comparison. That is 37.5% of Carbon and Argon's introductory rates, and 18.75% of their rates from December 1.
| Dimension | Carbon | Argon | 3.8 Flash |
|---|---|---|---|
| Price through Nov. 30, 2026 | $2 / $10 | $2 / $10 | $0.75 / $3.75 |
| Next listed price | $4 / $20 from Dec. 1, 2026 | $4 / $20 from Dec. 1, 2026 | $1.50 / $7.50 from Jan. 1, 2027 |
| Intelligence Index v5.0.1 | 60 (high thinking) | 50 (high thinking) | 31 (high thinking) |
| Context window | 1M tokens | 1M tokens | 1M tokens |
| Starting role | Hardest, high-value tasks | Production baseline | Evaluated high-volume work |
Sources: Google's Gemini API documentation (Carbon and Argon prices, context), Requesty's comparison updated November 12, 2026 (Flash prices, Artificial Analysis scores) and Requesty's model page (Flash context). Prices per 1M input / output tokens, checked November 16, 2026.
All three index scores were measured at high thinking, while Carbon and Argon default to medium. In the same Artificial Analysis data relayed by Requesty, Carbon scored 58 at medium and 54 at low. Google's Carbon model card lists a different index version, AA Index v5.0, with 59 for Carbon and 50 for Argon.
To compare tiers on your own workload, keep the same prompts, thinking level and acceptance test for each model. Add up billable input, cached input and output tokens (thinking included), then divide total spend by the number of accepted results. This counts the failed attempts that a per-token comparison misses.
Gemini 4 Carbon: test it on your hardest tasks
Google calls Carbon its strongest reasoning model, built for sustained agentic workflows. The model card says it is based on Gemini 4 Argon. Its lead holds at the default setting: Carbon's 58 at medium thinking is above Argon's 50 at high, according to Requesty's figures.
That quality has a token cost. Google says Carbon spends more thinking tokens per turn than Argon to explore hypotheses, verify its work and check edge cases. Thinking tokens are billed at the output rate, so Carbon can cost more per task at the same listed price.
It also takes longer. At high thinking, Artificial Analysis measured 66.29 seconds to Carbon's first answer token, against 39.34 seconds for Argon and 16.90 seconds for Flash, as reported by Requesty. At medium, Carbon's wait was 23.19 seconds.
Moving from Argon only means changing the model ID from gemini-4-argon to gemini-4-carbon; the request schema stays the same. Google's Carbon documentation asks you to allow for longer reasoning in max_output_tokens and HTTP timeouts. It recommends at least 32,768 output tokens for apps with a lower limit, and background execution or timeouts of at least 15 minutes for hard prompts.
Test Carbon where a wrong answer or an extra review cycle costs more than extra tokens. Google's examples include complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and security auditing.
Gemini 4 Argon: the better-supported production baseline
Google positions Argon for most advanced software engineering and knowledge work, with faster responses and lower token spend per turn than Carbon. It also has the only published workload result of the three.
Requesty ran Argon on 129 private Go backend tasks with the same agent and an 80-turn budget. In its October 20, 2026 run, Argon solved 88.4% of tasks, passed 95.1% of tests and ranked third out of 25 models. Requesty reports $2.14 per task, $2.43 per solve and an 86.7% cache hit rate.
These costs reflect introductory pricing: the same token use would cost twice as much from December 1. The benchmark is private and Go-only, so it may not transfer to another language, agent design or acceptance test. Requesty's comparison, updated November 12, 2026, has no Carbon result on it.
Before moving an Argon workload to Carbon, test Argon at high thinking as well as its medium default. A higher thinking level may meet your quality bar without a model change, at the price of more tokens and a longer wait.
Gemini 3.8 Flash: use it where evaluations pass
Flash has the lowest token rate here, at $0.75 / $3.75 per 1M input / output tokens through December 31, 2026. Requesty reports that this doubles to $1.50 / $7.50 on January 1, 2027, so it is not a lasting price.
Its index score is 31 at high thinking, against 50 for Argon and 60 for Carbon. The lower rate only helps where Flash meets your acceptance threshold on representative prompts. That is the same logic as in our article on rethinking your default LLM provider: the cheapest token is not automatically the right default.
How to choose a Gemini tier before December 1
Carbon and Argon cost the same per token through November 30, 2026, which makes this the cleanest window to compare their token use and completion rates. Set December routing from measured cost per accepted result, not from the temporary price.
Recommended starting points
Choose by workload and acceptance rate
Hardest tasks
Test Carbon
Use Carbon for high-value work where its deeper reasoning passes your evaluation and saves enough rework to cover the extra thinking tokens.
Complex production
Start with Argon
Use Argon as the baseline for advanced coding and knowledge work. Try high thinking before moving a task to Carbon.
High volume
Try Flash
Use Flash where representative evaluations pass consistently, and count rejected answers when you compare its cost.
Caching is the other big lever when you reuse long system prompts or documents. Carbon does not support explicit context caching. Google says implicit caching is on by default for prompts of at least 4,096 tokens and cuts cached input by 95%: $0.10 per 1M tokens through November 30, then $0.20 from December 1.
Google's lineup also includes Gemini 4 Hydrogen, which sits outside this three-way comparison. Our Gemini 4 Hydrogen analysis covers its benchmarks and pricing.
Fallback needs Carbon-specific timeouts
Operational fallback and quality escalation solve different problems. A router can react to provider errors, health failures or time limits. It cannot tell that a valid response failed your acceptance test unless your application checks it.
A safer Carbon request path
Send the Carbon request
Use Carbon-specific output limits and a timeout that allows long reasoning.
Check operational status
Fall back only on provider errors, health failures or deliberate limits.
Evaluate the answer
Apply task tests, schema checks or human review.
Escalate deliberately
Send a new request only when the acceptance rule requires it.
Record total task cost
Count every attempt and billable token behind the accepted result.
The model card warns of occasional slowness or timeout issues with Carbon, especially at higher thinking levels. A timeout copied from Argon can cut off normal Carbon reasoning and start a backup request, so you may pay for both calls. Set timeouts and output limits per model, not one global value.
Where Eden AI fits
A gateway sits between your application and the model providers, so you can switch tiers by changing a model ID. Requesty, which published the Go benchmark above, is one option. It serves all three models through an OpenAI-compatible endpoint with automatic failover, and lists a 5% fee on pay-as-you-go keys and 0% when you bring your own provider keys.
Eden AI gives you one API key and one OpenAI-compatible endpoint for more than 1,000 AI models, from LLMs to OCR, speech and image. With provider routing, you choose the order of providers, or sort them by price, speed or latency. Automatic fallback to backup models covers provider failures.
Check Eden AI's catalog for Carbon, Argon and Flash and their exact model IDs before planning a three-tier test there; the price of every model is listed in the catalog. Whichever gateway you use, apply the timeout rule above to its fallback settings. Routing does not make a weaker model suitable for every task.
What to keep in mind
- Prices are dated: Carbon and Argon double on December 1, 2026, and Flash's introductory price ends December 31, 2026.
- Benchmark settings differ from defaults: the 60, 50 and 31 index scores use high thinking, while Carbon and Argon default to medium.
- Carbon has no matching workload result: Requesty's November 12, 2026 comparison reports the Go benchmark for Argon only.
- Retries can erase token savings: rejected outputs, repeated requests and escalations on a cheaper tier add to its task cost.
- Your workload decides: a private Go benchmark may not transfer to your language, agent or acceptance rules.
FAQ
Is Gemini 4 Carbon better than Gemini 4 Argon?
Should teams switch to Carbon before December 1, 2026?
Why can Carbon and Argon have different task costs at the same token price?
Does moving from Argon to Carbon require API schema changes?
How should Carbon timeouts be configured?
Can I use Gemini 4 Carbon on the free tier?
How does caching affect Carbon and Argon costs?
Does automatic fallback check answer quality?
Further reading
- Google's Carbon announcement: availability, positioning and introductory pricing.
- What's new in Gemini 4 Carbon: thinking levels, caching, migration steps and timeout guidance.
- Gemini 4 Carbon model card: architecture, evaluations and known limitations.
- Requesty's three-model comparison: index scores, latency figures, the Go benchmark and a worked cost-per-solve example.
- Our Gemini 4 Carbon release analysis: launch benchmarks against Claude Opus 5.5 and GPT-6 Astra.
- Our Gemini 4 Hydrogen analysis: the other new Gemini 4 tier, with benchmarks and migration notes.


