Guides · 2026-10-29
Gemini 4 Argon pricing: the December doubling, costed
Google made Gemini 4 Argon generally available on 13 October 2026 at $2.00 per million input tokens and $10.00 per million output — the same list price as Claude Sonnet 5.5 and GPT-6.1 Sol. That price is introductory. Google’s pricing page already shows the second column: from 1 December 2026 every Argon rate doubles, to $4.00 and $20.00— exactly Claude Opus 5.5’s rate card. Below: the full price table, what one request costs at eight prompt sizes, three monthly workloads before and after the change, and what you can do about it before December.
The rate card, before and after
| Per million tokens | Through 30 Nov 2026 | From 1 Dec 2026 | Claude Opus 5.5 (for reference) |
|---|---|---|---|
| Input | $2.00 | $4.00 | $4.00 |
| Output (incl. thinking) | $10.00 | $20.00 | $20.00 |
| Cached input | $0.10 | $0.20 | $0.20 (+ $5.00 to write) |
| Batch / Flex input / output | $1.00 / $5.00 | $2.00 / $10.00 | $2.00 / $10.00 (batch) |
| Priority input / output | $3.60 / $18.00 | $7.20 / $36.00 | — |
| Free tier | none | none | none |
- No long-prompt surcharge. Argon has a 1,048,576-token context window and one price for any prompt length. Gemini 3.1 Pro, Google’s previous top model (still in preview), charges $4/$18 once a prompt passes 200K; GPT-6.1 Sol moves to $4/$15 above 272K.
- Implicit caching only. Google’s model page says explicit caching is not supported. Repeated prefixes are discounted automatically, with no write fee and no storage fee — but you cannot pin a document in the cache, so the hit rate is something you measure, not something you set.
- Thinking is output. Argon has low, medium and high thinking levels; whatever it thinks is billed at the output rate, which becomes $20.00 per million in December. On reasoning-heavy work the level you choose moves the bill more than the prompt does.
- Search grounding is extra. 5,000 grounded requests a month are free (shared across Gemini 3 and newer), then $14 per 1,000. That price is not changing on 1 December.
One request, six price lists
Cost of 1,000 requests at each prompt size, 2,000 output tokens each, no caching, standard tier. The December column is simply the November one times two, and it matches Opus 5.5 to the cent at every size.
| Prompt | Gemini 4 Argon (to 30 Nov) | Gemini 4 Argon (from 1 Dec) | Claude Opus 5.5 | Claude Sonnet 5.5 | GPT-6.1 Sol | Gemini 3.1 Pro |
|---|---|---|---|---|---|---|
| 10K | $40.00 | $80.00 | $80.00 | $40.00 | $40.00 | $44.00 |
| 50K | $120 | $240 | $240 | $120 | $120 | $124 |
| 150K | $320 | $640 | $640 | $320 | $320 | $324 |
| 200K | $420 | $840 | $840 | $420 | $420 | $424 |
| 250K | $520 | $1,040 | $1,040 | $520 | $520 | $1,036 |
| 300K | $620 | $1,240 | $1,240 | $620 | $1,230 | $1,236 |
| 600K | $1,220 | $2,440 | $2,440 | $1,220 | $2,430 | $2,436 |
| 900K | $1,820 | $3,640 | $3,640 | $1,820 | $3,630 | $3,636 |
Tier rules: Gemini 4 Argon (to 30 Nov) — flat, implicit cache only; Gemini 4 Argon (from 1 Dec) — flat, implicit cache only; Claude Opus 5.5 — flat to 1M; cache writes 1.25× input; Claude Sonnet 5.5 — flat to 1M; cache writes 1.25× input; GPT-6.1 Sol — $4/$15 above 272K; Gemini 3.1 Pro — preview; $4/$18 above 200K. Cheapest per row in green; ties all marked.
Until 30 November Argon is the cheapest flagship-class model in this table at every prompt size (tied with Sonnet 5.5, and with GPT-6.1 Sol up to 272K). Above 200K it costs about half of Gemini 3.1 Pro. From 1 December it is twice Sonnet 5.5 everywhere and twice GPT-6.1 Sol up to 272K, and above 200K it is level with or a little above 3.1 Pro’s long-prompt rate, depending on how much of the bill is output.
Three workloads, a month each
Cache hit rates are assumed equal across models (all of the reused context hits). Cache-write fees on Claude ($2.50–$8 per million written, depending on model and cache lifetime) and GPT-6.1 Sol ($2.50) are left out, and so is Gemini 3.1 Pro’s explicit-cache storage fee — all of these favour the non-Argon columns slightly.
| Workload | Gemini 4 Argon (to 30 Nov) | Gemini 4 Argon (from 1 Dec) | Claude Opus 5.5 | Claude Sonnet 5.5 | GPT-6.1 Sol | Gemini 3.1 Pro |
|---|---|---|---|---|---|---|
| Coding agent 80K prompt per step (64K repo context + history reused), 4K out incl. thinking; 30,000 steps | $2,352 | $4,704 | $4,704 | $2,352 | $2,352 | $2,784 |
| RAG answers 25K prompt (3K shared instructions), 1.2K out; 100,000 questions | $5,630 | $11,260 | $11,260 | $5,630 | $5,630 | $5,900 |
| Contract extraction 150K prompt (one document, nothing shared), 3K out; 5,000 documents | $1,650 | $3,300 | $3,300 | $1,650 | $1,650 | $1,680 |
The three workloads add up to $9,632 a month on Argon in November and $19,264 in December. The same token counts cost $9,632 on Sonnet 5.5 and $9,632 on GPT-6.1 Sol. Whether Argon is worth the gap is a quality question your own evals have to answer; the price side is now simple — from December you are paying Opus money and should compare it with Opus-class results.
When the cache misses
The coding agent row assumes every one of its 64K reused tokens is a cache hit. With implicit-only caching that is the best case, not a setting. On December prices the same month costs $4,704 with every hit, $8,352 if only half the reused prefix hits, and $12,000 with no hits at all. Keeping the stable part of the prompt (system instructions, tool definitions, repository context) at the very start and appending the changing part at the end is what moves an implicit cache toward the first number.
Batch and Flex: the November price, kept
Batch and Flex are half of standard on every model here, and Argon’s December batch rate ($2/$10) is its November standard rate. Work that can wait — the contract extraction above, nightly evals, back-fills — loses nothing from the change if it moves off the synchronous API. The extraction workload through each model’s batch tier:
- Gemini 4 Argon (to 30 Nov): $825 a month
- Gemini 4 Argon (from 1 Dec): $1,650 a month (vs $3,300 standard)
- Claude Opus 5.5: $1,650 a month
- Claude Sonnet 5.5: $825 a month
- GPT-6.1 Sol: $825 a month
- Gemini 3.1 Pro: $840 a month
At $1,650 Argon-in-December batch matches its own November standard price for this job ($1,650). Batch results arrive within hours rather than seconds; Flex trades latency and availability for the same discount on synchronous calls.
What to do before 1 December
- Re-cost at the December price now. If your budget or your own customer pricing was set on $2/$10, it is half of what it will be in five weeks. Double every Argon line in the forecast, not just the input.
- Run one eval against Sonnet 5.5 or GPT-6.1 Sol. They list at Argon’s current price and will not change on 1 December. If they are good enough for a route, that route halves in cost compared with staying on Argon.
- Measure tokens per task, not per million. The tables above use the same token counts for every model, but tokenizers differ: Anthropic says its Claude 4.7-and-later tokenizer turns the same text into about 30% more tokens. Thinking length differs too. Count real tokens on a sample of your own requests with each vendor before comparing rate cards.
- Move batchable work to Batch or Flex and lower the thinking level on routes that do not need it — Batch alone takes a job back to the November price.
- Watch Gemini Flash too. Gemini 3.8 and 3.6 Flash ($0.75/$3.75) double on 1 January 2027. A plan to route easy requests from Argon down to Flex-tier Flash still works, at the 2027 price.
The app around the model
At these prices the model is the largest line on most AI features’ bills. Our hosting data puts a Next.js front end at startup / small business traffic on these cheapest fitting plans:
- Next.js on Orkestr: Starter, free tier
- Next.js on DigitalOcean Droplets: Basic 512MB, about $4.00/month
- Next.js on OVHcloud VPS: VPS-1, about $4.54/month
Against that, the RAG workload alone is $11,260 a month on Argon from December. If the chat route runs on Vercel, the function time spent waiting for a streamed answer is billed separately — see what a streaming AI route costs in Fluid compute, and compare a container host with Vercel vs Railway or Vercel alternatives by cost. Embeddings and chat history need a database: see managed Postgres pricing for pgvector-ready plans.
Should you use Argon?
- Through November: on price, yes for testing — it costs the same as Sonnet 5.5 and GPT-6.1 Sol per token and has no long-prompt tier. Just do not build a business case on that price.
- From December, against Opus 5.5: same rate card, so it is a straight quality and token-count comparison. Argon has no cache-write fee; Opus lets you pin a cache entry for an hour.
- From December, against Sonnet 5.5 or GPT-6.1 Sol: Argon is twice the price per token. It has to do the job in half the tokens or do a job they cannot.
- On Gemini 3.1 Pro today: below 200K, Argon in December costs more (3.1 Pro is $2/$12); above 200K they are close. 3.1 Pro is still a preview model. The wider field is in LLM API pricing in 2026; for the budget end, see Claude Haiku 5.5’s 100K-token cliff.
Cost your own token volumes
The LLM API cost calculator uses the same data set. For the hosting side, the hosting cost calculator ranks every provider for your stack and traffic.
Prices verified 2026-10-29on Google’s Gemini API pricing page and Gemini 4 Argon model page, Anthropic’s pricing page and OpenAI’s API pricing page; GPT-6.1 Sol’s long-prompt rate from our LLM data set (2026-11-02). Base input/output prices come from that data set; hosting estimates from our hosting data set (2026-11-02). Token counts in the workloads are illustrative, not measured. No AI provider pays us anything, and none of the links in this article are affiliate links.
FAQ
How much does Gemini 4 Argon cost?
Through 30 November 2026: $2 per million input tokens, $10 per million output tokens (thinking tokens are billed as output) and $0.10 per million cached tokens. From 1 December 2026 every rate doubles: $4 input, $20 output, $0.20 cached. There is no free tier and no higher rate for long prompts. Batch and Flex are half the standard price ($1/$5, then $2/$10); Priority is $3.60/$18, then $7.20/$36.
Is Gemini 4 Argon more expensive than Claude Opus 5.5 from December?
At the same token counts it costs the same: from 1 December Argon is $4 input, $20 output and $0.20 for cached input per million, which is exactly Opus 5.5’s standard rate card. Differences come from elsewhere — Anthropic charges extra to write a cache entry while Argon’s implicit caching has no write fee, and the two tokenizers turn the same text into different token counts.
Can I use explicit context caching with Gemini 4 Argon?
No. Google’s model page says explicit caching is not supported for Gemini 4 Argon; only implicit caching is. Repeated prompt prefixes are discounted automatically when they hit the cache ($0.10 per million through November, $0.20 from December), with no storage fee, but hits are not guaranteed — you cannot pin a document in the cache the way explicit caching allows on Gemini 3.1 Pro.
What is the cheapest way to run Gemini 4 Argon after 1 December?
Batch or Flex, at $2 input and $10 output per million — half the standard price, which is the same as the standard November price. Work that can wait for results (document extraction, evaluations, back-fills) loses nothing from the price change if it moves to Batch. Keeping shared instructions at the start of the prompt also raises the implicit cache hit rate.