Gemini 4 Argon's Price Doubles on December 1: On Our Battery It Lands Level With Claude Opus 5.5
By Promptster Team · 2026-11-15
What does Gemini 4 Argon cost after its promo ends? Twice as much. shared/pricing.ts lists it at $2.00 input / $10.00 output per million tokens with the comment "Promo until 2026-11-30, then $4.00/$20.00". Both rates double on 2026-12-01, so every Argon bill doubles, whatever the mix of input, output and thinking tokens.
On our November battery that takes Argon from $0.0667 to $0.1333 for 20 calls. From December 1, Argon costs about the same as Claude Opus 5.5 ($0.1341) on this battery, and several other models that passed 20/20 cost less than either, including GPT-6.1 Sol, GPT-6 Sol and DeepSeek V4 Pro. If you chose Argon because it was cheap, December is the time to check that again.
The arithmetic
The battery is four script-graded tasks, five runs each: strict JSON with traps (json-order), a ten-step round-down arithmetic chain (stock-floor), three taglines of exactly seven words (seven-word-taglines), and a final-value extraction from superseded meeting notes (launch-notes). Argon passed all 20 calls. Its cost_usd from summary.json, at today's rate and doubled:
| Task | Argon now ($2 / $10) | Argon from 12-01 ($4 / $20) | Claude Opus 5.5 ($4 / $20) |
|---|---|---|---|
json-order |
$0.0062 | $0.0123 | $0.0156 |
stock-floor |
$0.0487 | $0.0973 | $0.0689 |
seven-word-taglines |
$0.0083 | $0.0165 | $0.0389 |
launch-notes |
$0.0036 | $0.0072 | $0.0108 |
| All 20 calls | $0.0667 | $0.1333 | $0.1341 |
$0.06665 × 2 = $0.1333. Opus 5.5 lists at $4 / $20 too, so from December the two have the same rate card and the difference is tokens. Argon still comes out cheaper on three of the four tasks at December rates. It loses on stock-floor, where it wrote 593–785 visible tokens of working per call and on two of five runs also reported 508 and 723 thought tokens. That one task is 73% of Argon's bill, and it cancels out the savings on the other three.
Where Argon's $0.0667 went over the 20 calls (summed from the run files' raw_usage): 3,030 input tokens ($0.0061), 3,898 visible output tokens ($0.0390) and 2,161 thought tokens ($0.0216). Google bills thoughtsTokenCount at the output rate, and Argon reported it on 6 of 20 calls. Thinking is a third of the bill, and from December each thought token costs $20 per million.
In our October head-to-head, on an easier battery at max_tokens 1,500, Argon's December estimate came out at 1.47× Opus 5.5. On this battery the two are within $0.001 of each other. The tasks are different, so the ratio moved. The conclusion did not: from December, Argon is not the cheaper of the two.
Some of the models that passed 20/20 for less
Every row below passed all 20 calls, along with every format diagnostic (no fences, a literal Answer: line, the bare Nov 21, $37500):
| Model | Spend, 20 calls | vs Argon from 12-01 | Median latency, s (json / stock / tag / launch) |
|---|---|---|---|
| GPT-6.1 Sol | $0.0325 | 0.24× | 2.0 / 7.7 / 4.7 / 2.0 |
| GPT-6 Sol | $0.0370 | 0.28× | 1.6 / 3.2 / 4.7 / 1.3 |
| DeepSeek V4 Pro | $0.0498 | 0.37× | 3.6 / 5.7 / 5.3 / 3.1 |
| Gemini 4 Argon, now | $0.0667 | 0.50× | 4.9 / 11.0 / 3.8 / 2.2 |
| Gemini 4 Argon, from 12-01 | $0.1333 | 1× | same |
| Claude Opus 5.5 | $0.1341 | 1.01× | 2.6 / 5.5 / 4.2 / 2.1 |
GPT-6.1 Sol and GPT-6 Sol list at $2 / $10, Argon's current rate, and cost about half of what Argon costs today on the same 20 calls. They also report reasoning inside their output tokens, so there is no separate thinking line to watch. DeepSeek V4 Pro is priced at its peak-hour rate ($1.32 / $3.96); summary.json does not model DeepSeek's off-peak discount, so the off-peak figure would be lower.
Latency adds to the case. Argon's slowest call took 35.2 s (a stock-floor run that also reported 508 thought tokens), and its 20 calls ranged from 1.9 s to 35.2 s. Opus 5.5 ranged from 1.9 s to 5.8 s. Argon had the slowest median of these five models on json-order and stock-floor.
The cheapest model to pass all 20 was GPT-6 Luna at $0.0019. That is a different price tier; see its October match-up with Claude Haiku 5.5.
What to do before December 1
- Find your Argon traffic. Anything with
gemini-4-argonin a saved test, schedule or API call is billed at the new rate from 12-01. Promptster's cost display usesshared/pricing.ts, so it will show the doubled figure once the rate there changes. - Price your own prompts, not the rate card. Argon was cheaper than Opus 5.5 on three of our four tasks at December rates and still tied overall, because one long-output task was most of its bill. Your prompt mix decides which way it goes.
- Re-run your prompts on GPT-6.1 Sol or GPT-6 Sol. On this battery either one does the same graded work for under a third of Argon's December price.
- Check
thoughtsTokenCount. Argon reported thinking on 6 of 20 calls here, and those six cost $0.0216 of thinking at promo rates. At $20 per million they cost twice that. - Don't switch to Flash on the name. Gemini 3.8 Flash ($0.0742) and 3.6 Flash ($0.1048) cost more than Argon today on this battery, because they thought on every call (the Flash post). Their promos run to 2026-12-31, so in December only they are cheaper than Argon, and from 2027-01-01 they double too.
What this does not show
n=5 per task. Every model here passed every call, so this is a cost comparison, not a quality ranking: four single-turn tasks cannot show what Argon or Opus 5.5 do better on long agentic work or hard coding. Promo dates can move. We read them from the comments in shared/pricing.ts on draft day, and if Google changes them, the doubling date changes with them. The arithmetic stays the same.
Method
Four script-graded tasks (no LLM judge), five runs per task per model, max_tokens 4,000, run 2026-10-23 through Promptster's own provider handlers. Temperature 0.7 was requested on every call; GPT-6-family and Claude 5-family models reject a custom temperature, so Promptster sends them none. cost_usd is calculateCost() from shared/pricing.ts at list rate, including Gemini thought tokens at the output rate, with no cache, batch or off-peak discounts. December figures are the measured totals × 2. Latency is wall-clock from one machine in one session, not a provider SLA. For the open-weight side of this battery, see Benchmarking open-source vs. closed-source models.
Run this comparison yourself — no account, no API key
- Run the ten-step stock-floor task, the task that was 73% of Argon's bill (answer: 587)
To price Argon, Opus 5.5 and the GPT-6 Sol models on your own prompts before the rate changes, that's Promptster.
All figures from content/blog/data/q4b-2026-battery/ (summary.json plus one raw file per call, with each provider's usage object verbatim), run 2026-10-23. December figures double the measured cost_usd per the shared/pricing.ts comment for gemini-4-argon: promo $2 / $10 until 2026-11-30, then $4 / $20.