Two changes reach Gemini API bills on December 1, 2026. The Standard price of Gemini 4 Carbon and Gemini 4 Argon doubles, from $2.00 input and $10.00 output per 1M tokens to $4.00 and $20.00. And every request addressed to gemini-2.5-pro starts being answered by gemini-3.8-flash and billed at its rates, while gemini-2.5-flash goes to gemini-3.5-flash-lite. Carbon reached paying developers only on November 10, so its launch price lasts three weeks. Gemini API pricing · Gemini API changelog
S5 Labs ran none of these models. Prices, dates and migration rules come from Google’s pages as read on November 12. Dollar totals are our arithmetic from those rates, and the benchmark and safety figures are Google’s claims. The launch itself and the Fairwind cyber-defender program are in our Carbon access article, updated today.
Carbon and Argon double in every paid tier
Carbon and Argon carry identical prices and the same dated step. Neither is on the Free tier, and thinking tokens bill as output. Gemini API pricing
| Per 1M tokens, input / output | Through Nov 30, 2026 | From Dec 1, 2026 |
|---|---|---|
| Standard | $2.00 / $10.00 | $4.00 / $20.00 |
| Batch and Flex | $1.00 / $5.00 | $2.00 / $10.00 |
| Priority | $3.60 / $18.00 | $7.20 / $36.00 |
| Cached input: Standard, Batch and Flex, Priority | $0.10, $0.05, $0.18 | $0.20, $0.10, $0.36 |
The Argon figures match what we published when Argon launched in October. Google’s launch post gives Carbon the same introductory price as Argon and puts the November 30 end date in a footnote. Gemini 4 Carbon launch post
These are AI Studio prices. The pricing page says Gemini Enterprise Agent Platform prices may differ, and that platform’s model list shows Carbon as generally available with no prices. No page we opened says whether the December 1 step applies there. Gemini Enterprise Agent Platform model list
The same tokens cost double unless the tier or model changes
Take 100 million input tokens and 20 million output tokens a month, with no cache hits, grounding or storage charges. This is S5 Labs arithmetic from the list rates:
| Model and tier | At rates through Nov 30 | At rates from Dec 1 |
|---|---|---|
| Carbon or Argon, Standard | $400 | $800 |
| Carbon or Argon, Batch or Flex | $200 | $400 |
| Carbon or Argon, Priority | $720 | $1,440 |
| Hydrogen, Standard | $300 | $300 |
From December 1, Batch and Flex cost what Standard costs today, so a Carbon or Argon job that can run on either tier keeps its November budget by moving.
The other option is Gemini 4 Hydrogen, generally available since November 12 at $1.50 input and $7.50 output, with a free tier and no dated step on the pricing page. Gemini API changelog On this workload it costs $300 against Carbon’s $800.
A fixed token count is not a fixed amount of work, though. Google’s Carbon migration guide says moving from Argon requires no request schema changes, but that Carbon uses more thinking tokens per turn on complex tasks, and it advises keeping Argon for faster, lower-token interactive coding. What’s new in Gemini 4 Carbon Hydrogen defaults to high thinking, where Carbon and Argon default to medium. Gemini thinking docs To compare costs instead of rates, measure billed output tokens per finished task on your own prompts at each thinking level.
2.5 Pro calls will be answered by a Flash model
The November 2 changelog entry deprecates gemini-2.5-pro and says that from December 1 all API requests to that ID route automatically to gemini-3.8-flash and bill at 3.8 Flash rates. Google’s instruction is to change the model string before then. gemini-2.5-flash gets the same notice with gemini-3.5-flash-lite as the target. Gemini API changelog This supersedes the September 18 entry, which said the 2.5 models were not deprecated. gemini-2.5-flash-lite still is not, and has no shutdown date.
The deprecations page gives both IDs a shutdown of no earlier than December 1, 2026. The reroute is the fixed event, and it works like the October 8 reroutes of 3.7 and 3.5 Flash: the old ID keeps answering, from a different model. Our article on those reroutes was written before these notices.
3.8 Flash, the target, has an introductory price of its own: $0.75 input and $3.75 output through December 31, 2026, then $1.50 and $7.50 from January 1, 2027. Gemini API pricing Against 2.5 Pro’s Standard rates of $1.25 and $10.00 for prompts up to 200k tokens, our arithmetic gives:
| Monthly tokens, uncached | 2.5 Pro today | 3.8 Flash, December | 3.8 Flash from Jan 1, 2027 |
|---|---|---|---|
| 100M input, 20M output | $325 | $150 | $300 |
| 500M input, 10M output | $725 | $412.50 | $825 |
From January, input costs 20% more than 2.5 Pro and output 25% less, which breaks even at 10 input tokens per output token. The input-heavy second row pays 13.8% more than now. Prompts above 200k tokens, which 2.5 Pro bills at $2.50 and $15.00, stay cheaper, because 3.8 Flash has one price tier. 3.8 Flash also has a free tier, but Google does not say whether free-tier 2.5 Pro use carries over to it.
Those totals assume token counts carry over, which a different model does not guarantee. Google’s models page calls 3.8 Flash its previous-generation Flash model. Gemini models Whether it holds quality on prompts tuned for a Pro model is yours to test; our 3.8 Flash cost article has the task-cost caveats.
For gemini-2.5-flash, the list rates barely move. Flash-Lite’s text rates equal what 2.5 Flash charges now, $0.30 input and $2.50 output, and by our comparison of the two pricing rows audio input falls from $1.00 to $0.30.
Google lists the request changes, not what the router does with old requests
For both reroutes the changelog names the same breaking changes: function calling and thought_signature validation, thinking_level replacing thinking_budget, and removed sampling parameters and prefilled model turns. It does not name the sampling parameters or describe any error behavior. The 3.8 Flash migration guide, written for calls made directly to that model, is more specific. What’s new in Gemini 3.8
| 2.5-era pattern | Guide’s instruction for 3.8 Flash |
|---|---|
thinking_budget | Replace with thinking_level: low, medium (default) or high; minimal is not supported |
temperature, top_p, top_k, candidate_count | Remove |
| Prefilled model turns | Remove |
| Multi-turn state | Standardize on server-side previous_interaction_id |
FunctionResponse with generateContent | Include call_id and name |
The Flash-Lite guide says sampling parameters are deprecated and currently ignored, with a 400 error reserved for future model generations, while a request whose last non-empty turn is a model turn returns HTTP 400 today. Flash-Lite’s default thinking level is minimal, and Google suggests medium or high for tool-calling subagents. Gemini 3.6 and 3.5 Flash-Lite guide
All of that describes direct calls. Google does not say what the December 1 router does with a thinking_budget, a temperature or a prefilled turn in a request still addressed to a 2.5 ID. The only way to avoid finding out in production is to change the model string and the request shape now.
For these two IDs Google names one target each. Sending a 2.5 Pro workload to Hydrogen, Argon or Carbon instead is your own model-string change, with its own migration. For Carbon the guide says explicit caching is not supported and sampling parameters are ignored, and it recommends a max_output_tokens of at least 32,768 and, for hard prompts, HTTP timeouts of at least 15 minutes. What’s new in Gemini 4 Carbon
Carbon’s benchmark and safety numbers are Google’s
Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at less than half the cost per task of the next best external model. It also reports a per-attempt breach rate in simulated cyberattack environments of 0.8% with Carbon’s safeguards on, against 3.1% for Argon, and 92% against 70% on a 375-task agentic honesty benchmark. It repeats a 0% attack success rate on the Gray Swan IPI benchmark, a result it says it shared two weeks earlier, alongside the admission that prompt injection robustness “remains an area for continued improvement”. Gemini 4 Carbon launch post We opened no independent reproduction of any of these.
The one outside measurement we opened is Artificial Analysis, which lists Carbon at medium thinking, with Google as the only provider, at $4.00 and $20.00, the rates that apply from December 1. As read on November 12, its headline figures are $6.59 per Intelligence Index task, 24.0 minutes per task and 23.34 seconds to the first answer token, while the provider table on the same page shows 23.3 minutes, 23.16 seconds and a 5.9% endpoint error rate. Artificial Analysis, Gemini 4 Carbon The speed figures are five-day averages on a page that keeps updating. Next to Google’s timeout advice, they argue for long timeouts and a retry budget.
Before December 1
- Search code, configs and environment variables for
gemini-2.5-proandgemini-2.5-flash. Each hit becomes a different model on December 1. - Choose the replacement yourself: the Flash target Google names, or a Gemini 4 model if that target fails your evaluation.
- Fix the request before you switch:
thinking_levelin place ofthinking_budget, no sampling parameters, no prefilled model turns, andcall_idandnameon generateContent function responses. - Set a thinking level explicitly on Flash-Lite, whose default is
minimal. - Rerun your evaluation on the new ID and record output tokens per task.
- Put the December Carbon and Argon rates in every budget and customer quote, move what tolerates it to Batch or Flex, and price Hydrogen on the same tasks.
- Check two neighboring dates. The TTS and Live previews in our preview shutdown article can end no earlier than November 17. Requests to
gemini-omni-1.1-flashroute to 1.2 Flash, at its rates, from December 3, per the November 5 changelog; our Omni 1.2 Flash article does not mention it. - Put the January 1, 2027 rate for 3.8 Flash in next year’s budget now. At the same token counts, the reroute saves money for one month. After that, workloads above 10 input tokens per output token, on prompts up to 200k tokens, cost more than they did on 2.5 Pro.
