Gemini 4 Argon migration means more than a model swap: temperature, top_p and top_k are ignored, thinking budget gives way to thinking level
Until November 30, Google's Gemini 4 Argon bills input at $2 per million tokens and output at $10 per million, then $4 and $20. Thinking tokens bill at the output rate, according to a dev.to walkthrough of Google's pricing page.
The Engineer · Build desk
Key figures from an illustrative example in a dev.to post: the feature's Argon bill is $240 a month from December, against $120 a month at current rates, for the same workload of 20 million input tokens a month.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Illustrative bill from December, per dev.to | 240 dollars a month | 16 | |
| Same traffic, current rates | 120 dollars a month | 19 | |
| Example input tokens | 20 million a month | 18 |
What happened
- Google opened Argon to everyone on the paid Gemini API on October 13, and the Vercel AI SDK's Google provider added the gemini-4-argon model ID the same day.
- Custom temperature, top_p and top_k values are ignored on Argon and return a deprecation warning, so the request still succeeds.
- The thinking_budget setting is deprecated in favour of thinking_level, which takes low, medium or high, and explicit context caching is not supported.
- Google's guide says hard reasoning prompts can take several minutes and recommends an HTTP timeout of at least 15 to 20 minutes.
Why it matters
- cost A budget built on this month's rate covers only half the bill from December 1 for identical traffic.
- constraint Anyone who tuned temperature for stable extraction output loses that control on Argon, and a passing test suite will not show it because the request still succeeds.
- exposure Serverless routes with a maximum function duration shorter than Google's 15 to 20 minute timeout advice are exposed when they run Argon at the high thinking level.
- decision Each call site now needs a deliberate thinking level. Google's guidance puts classification, routing and extraction at low, and multi-file debugging and architecture work at high.
The dev.to post's before-and-after for the Vercel AI SDK shows how small the diff is. A call on gemini-2.5-pro that passes temperature 0.2 and a thinking budget of 2048 becomes a call on gemini-4-argon with thinkingLevel set to 'low' and no temperature at all [17]. The thinkingLevel option and the new model ID both exist in @ai-sdk/google 4.0.95, released October 13. Update the package first. Any cachedContent in providerOptions has to come out [14].
One trap sits in the types. The SDK's type also accepts 'minimal', because older Gemini models do, but Argon's docs list only low, medium and high [15].
The billing detail is where the default matters. Thinking sits inside the output line [4], and the default thinking level is medium, so every call thinks, short answers included [5]. Argon is also "not available on the Free tier or on the Google AI Pro plan," so testing starts on the paid tier [6]. Batch and Flex run at half the standard rates through November 30 and double on the same date [3].
The post gives an illustration and labels it as one, not a benchmark. In it, a feature makes 10,000 monthly calls at 2,000 input tokens and 800 output tokens apiece, with thinking counted in the output [16]. That is 20 million input tokens and 8 million output tokens a month [18]. At $2 and $10 per million, the bill is $40 plus $80, or $120. At $4 and $20 from December, it is $80 plus $160, or $240 [19]. Output is two-thirds of the bill in both periods [20].
The number transfers only if your calls average near 800 output tokens once thinking is counted. The post does not say how much of those 800 is thinking.
For slow calls, the post points to the background flag on the Interactions API. It also says the AI SDK's changelog now has generateText polling background calls until they finish [13].
Every figure here is the post's reading of Google's pricing page and developer guide. The post adds that Google reports large gains on its own codebase, that independent rankings disagree on whether Argon is the best model, and that the honest test is your own prompts [21]. We would cost those tests at the December rates.
What to watch
- Whether the $4 and $20 rates take effect on December 1 as the pricing page states.
- Whether independent evaluations settle the disagreement over whether Argon beats rival models.
- Whether Google turns the sampling-settings deprecation warning into a hard error.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence70
- Adoption
- Insufficient
- Hype gap−10
- Incentives
- Insufficient
- Confidence68
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Google opened Gemini 4 Argon, its new flagship model, to everyone on the paid Gemini API on October 13, with model ID gemini-4-argon, and the Vercel AI SDK's Google provider added it the same day.
- [2]
According to Google's pricing page, Argon costs $2 per million input tokens and $10 per million output tokens through November 30; from December 1 it becomes $4 and $20.
- [3]
Batch and Flex pricing are half of the standard Argon numbers, and they double on the same date (December 1).
- [4]
Google's pricing table labels the output rate 'output price (including thinking tokens)', so thinking is billed as output.
- [5]
The default thinking level is medium, so every call does some thinking, even for short answers.
- [6]
Google's pricing page says Argon is 'not available on the Free tier or on the Google AI Pro plan.'
- [7]
temperature, top_p and top_k are ignored on Argon. In Google's words: 'Custom sampling values are ignored by gemini-4-argon and return a deprecation warning.' The request succeeds; it is a warning, not an error.
- [8]
thinking_budget is deprecated on Argon, which uses thinking_level with three values: low, medium (the default) and high.
- [9]
Google suggests low thinking for classification, routing and extraction, and high for multi-file debugging and architecture work.
- [10]
Explicit context caching (creating a cache and passing cached_content) is not supported on Argon. Caching is implicit, and cached input tokens are 95% off.
- [11]
Google's guide says hard reasoning prompts 'can take several minutes' and recommends an HTTP client timeout of at least 15 to 20 minutes, or running the request in the background (background: true on its Interactions API).
- [12]
A typical serverless request handler is not built to wait 20 minutes; anyone planning thinkingLevel 'high' behind an API route should check the platform's maximum function duration or move the work to a background job and poll.
- [13]
The AI SDK's changelog says generateText now polls background calls until they finish when you use the Interactions API.
- [14]
The thinkingLevel option and the gemini-4-argon model ID both exist in @ai-sdk/google 4.0.95 (released October 13); users should update the package first and remove cachedContent from providerOptions if used.
- [15]
The SDK's type also allows 'minimal' because older Gemini models accept it, but Argon's docs list only low, medium and high.
- [16]
The post's illustrative example (labelled not a benchmark): a feature makes 10,000 calls a month, each with 2,000 input tokens and 800 output tokens including thinking, which the post puts at $120 a month now and $240 from December.
- [17]
The post's Vercel AI SDK 'before' uses google('gemini-2.5-pro') with temperature 0.2 and thinkingBudget 2048; the Argon 'after' uses google('gemini-4-argon') with thinkingLevel 'low' and no temperature.
- [18]
The example workload is 20 million input tokens and 8 million output tokens a month.
- [19]
At the current rates the example costs $120 a month; at the December rates it costs $240 for the same traffic.
- [20]
Output tokens, thinking included, are two-thirds of the example bill in both periods.
- [21]
The post says Google reports big wins on its own codebase, independent rankings disagree about whether Argon is the best, and the honest test is your own prompts, not a leaderboard.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toGemini 4 Argon is half price until November 30. Three of your settings already do nothing
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- LLM API PricingFollow
- Reasoning ModelsFollow
- Model migrationFollow