Gemini 4 Argon is out of preview, and its $2 price now has an end date: November 30. Until then it is the cheapest frontier model for prompts over 272K tokens. On December 1 it costs what Claude Opus 5.5 costs.

Short version. $2 in, $10 out through November 30, then $4 and $20. That gives the public 49 days at the low price. Gemini 3.8 Flash got 121.

Argon has no long-context step, so for the next six weeks a 400K-token prompt costs about half what it does on GPT-6.1 Sol or Opus 5.5. From December 1 it costs the same as Opus 5.5 and a little more than Sol. Caching is implicit only.

The whole card, both dates

Google's launch post on September 30 gave two prices and the phrase "after the introductory period expires". The Gemini API pricing page never uses the word introductory. It just prints two columns of numbers with a date on each, and it fills in the lines the launch post skipped.

Gemini API pricing page, model ID gemini-4-argon, read October 20, 2026.

Every line doubles. Batch is half of standard on both dates, and Priority is 1.8x. The pattern from the launch post holds on the new lines too. Batch through November 30 is $1 and $5, which is GPT-6.1 Sol's batch card. From December 1 it is $2 and $10, which is Opus 5.5's batch card. Google borrowed both companies' prices twice.

Argon isn't on the free tier. Search grounding gets 5,000 free requests a month, shared with every Gemini 3 and newer model on the project, and then costs $14 per 1,000. Google bills each search query separately, and one request can make several.

No 200K step is the real discount

On short prompts, Argon at $2 and $10 is exactly Sol. Our test call is 30,000 tokens in and 3,000 out. It costs $0.09 on both, $0.096 on Gemini 3.1 Pro Preview and $0.18 on Opus 5.5. At 10,000 calls a month that is $900 against $1,800. That gap is real, but you already had it with Sol.

Long prompts are where Argon's card is different. Sol bills $4 and $15 once input passes 272,000 tokens, and we price the whole request at that rate. 3.1 Pro goes to $4 and $18 over 200,000. Argon's row has one rate and no step at all. Take a document job with 400,000 tokens in and 8,000 out, run 10,000 times a month with no cache hits:

• Argon, through Nov 30$8,800
• GPT-6.1 Sol (over 272K)$17,200
• Gemini 3.1 Pro (over 200K)$17,440
• Claude Opus 5.5$17,600
• Argon, from Dec 1$17,600

Monthly cost at standard rates. Per request: Argon $0.88 now and $1.76 from December 1, Sol $1.72, 3.1 Pro $1.744, Opus 5.5 $1.76.

Until November 30 Argon does that job for $8,800, about half of any other frontier model we list. On December 1 it lands on $17,600, the same as Opus 5.5 to the cent and $400 over Sol. After that date Argon never beats Sol on price. Below 272K it costs double, and above it the input matches while Argon's output costs $5 more per million.

Batch keeps the same order. The job is $4,400 on Argon now, $8,600 on Sol's long-context batch rate, and $8,800 on Argon or Opus 5.5 from December 1. Long prompts you can run overnight are the best use of the next six weeks.

You can't pin a cache

One line under the caching price changes how you plan an agent: "Implicit caching only (explicit caching not supported)." On Gemini 3 and 2.5 models you can create a cache with a TTL, pay an hourly storage fee, and know your 200K-token system prompt bills at the cached rate on every call. On Argon Google decides when a hit happens.

That cuts both ways. You pay no storage, and there's no write fee either. Opus 5.5 charges $5 per million to write a 5-minute cache. But if hits don't land, a long agent loop bills its full prefix at $2, or $4 from December 1, where you modelled 10 or 20 cents. That is a 20x difference on most of the bill.

So don't take a cache-hit rate from another model and assume Argon matches it. Read cachedContentTokenCount in the usage metadata of your first few hundred calls (or usage.total_cached_tokens on the Interactions API), then price the workload on the hit rate you actually see.

What the per-task numbers said

Nothing on the benchmark side moved with GA. Artificial Analysis still has Argon at high scoring 53 for $1.99 a task at the $2 price and $3.98 at $4. Opus 5.5 at high scored 54 for $1.82 when we wrote up the launch, and GPT-6.1 Sol at max scored 52 for $0.72. Argon writes a lot of output, so the doubling hurts it more than the card suggests.

Our view hasn't changed. On reasoning-heavy, output-heavy work, Opus 5.5 at high or Sol already costs less per task, and from December 1 the gap gets wider. If you send long documents and want short answers back, Argon is the cheapest frontier model you can buy today. It stays that way for 41 more days.

What to do before December 1

• Put a hard date in the budget. From December 1 every Argon line costs double, so any forecast past November still using $2 is wrong.
• Move long-prompt backfills and evals onto Batch now. At $1 in, a 400K prompt costs 40 cents.
• Measure your implicit cache-hit rate before you scale an agent loop on Argon.
• If most of your traffic is under 272K and output-heavy, run the same eval on GPT-6.1 Sol. From December 1 it costs half as much per token.

We've moved Argon from gated to generally available in the catalogue and attached the December 1 price, so the calculator shows both rates for any token mix.

Sources

• Gemini API pricing - Argon standard, batch, flex, priority and caching rates on both dates, free tier, grounding; Gemini 3.1 Pro and 3.8 Flash rows
• Gemini API changelog - October 13, 2026: gemini-4-argon generally available
• OpenAI API pricing and Anthropic pricing - GPT-6.1 Sol long-context and batch rates, Opus 5.5 card and cache write
• Artificial Analysis leaderboard - Intelligence Index score and cost per task, as cited in our October 1 post