Is Gemini Flash the Cheap Gemini? On Our November Battery, Both Flash Models Cost More Than Gemini 4 Argon
By Promptster Team · 2026-11-13
Is Flash the cheap Gemini? On list price, yes: Gemini 3.8 Flash and 3.6 Flash list at $0.75 input / $3.75 output per million tokens, and Gemini 4 Argon lists at $2 / $10, 2.7× more on both rates. On the 20 calls we ran, no. Gemini 3.6 Flash cost $0.1048, 3.8 Flash $0.0742 and Gemini 4 Argon $0.0667, and all three passed 20/20. The cheapest Gemini on this battery is the one with the highest rate card.
The reason is thinking tokens. Both Flash models reported thoughtsTokenCount on every one of their 20 calls, 15,928 tokens in total for 3.8 Flash and 23,613 for 3.6 Flash. Argon reported it on 6 of 20, 2,161 tokens in total. Google bills thought tokens at the output rate, so at these volumes the Flash discount is used up before the visible answer starts. That holds until 2026-11-30; the price calendar further down changes the order twice.
Quality: no separation
| Task | Gemini 4 Argon | Gemini 3.8 Flash | Gemini 3.6 Flash |
|---|---|---|---|
Strict JSON with traps (json-order) |
5/5 | 5/5 | 5/5 |
Ten-step round-down arithmetic (stock-floor, answer 587) |
5/5 | 5/5 | 5/5 |
Three taglines, exactly 7 words (seven-word-taglines) |
5/5 | 5/5 | 5/5 |
Final value among superseded notes (launch-notes) |
5/5 | 5/5 | 5/5 |
| Total | 20/20 | 20/20 | 20/20 |
All three also passed every diagnostic: no code fences, a literal Answer: last line, and the bare Nov 21, $37500 the launch-notes prompt asks for. On this battery the question is only what each model costs and how long it takes.
What the 20 calls cost
cost_usd from summary.json, at the rates in shared/pricing.ts:
| Task | Gemini 4 Argon | Gemini 3.8 Flash | Gemini 3.6 Flash |
|---|---|---|---|
json-order |
$0.0062 | $0.0105 | $0.0176 |
stock-floor |
$0.0487 | $0.0361 | $0.0506 |
seven-word-taglines |
$0.0083 | $0.0213 | $0.0220 |
launch-notes |
$0.0036 | $0.0063 | $0.0146 |
| All 20 calls | $0.0667 | $0.0742 | $0.1048 |
One task goes the other way. On stock-floor, 3.8 Flash was cheaper than Argon, because Argon wrote about 700 visible tokens of working per call at $10 per million. On the three short-answer tasks Argon was cheaper than both Flash models, by up to 4.1× (3.6 Flash on launch-notes).
Where the money went
Every token the three models used over the 20 calls, summed from the raw_usage object in each run file, priced at today's rates:
| Gemini 4 Argon | Gemini 3.8 Flash | Gemini 3.6 Flash | |
|---|---|---|---|
| Input tokens | 3,030 | 2,410 | 2,410 |
Visible output (candidatesTokenCount) |
3,898 | 3,369 | 3,855 |
Thinking (thoughtsTokenCount) |
2,161 | 15,928 | 23,613 |
| Calls that reported thinking | 6 of 20 | 20 of 20 | 20 of 20 |
| Cost of the thinking | $0.0216 | $0.0597 | $0.0885 |
| Cost of everything else | $0.0450 | $0.0144 | $0.0163 |
The arithmetic for 3.6 Flash: 23,613 thought tokens × $3.75 per million = $0.0885, which is 84% of its $0.1048. Its visible answers and prompts cost $0.0163 together, about a third of the $0.0450 Argon spent on the same. Flash is cheaper per token; it used far more tokens.
How Gemini reports thinking
Gemini's usageMetadata has candidatesTokenCount for the visible answer and a separate thoughtsTokenCount for reasoning. totalTokenCount includes both. Here is one 3.6 Flash call on the tagline task, verbatim:
{"promptTokenCount": 65, "candidatesTokenCount": 25, "totalTokenCount": 1225, "promptTokensDetails": [{"modality": "TEXT", "tokenCount": 65}], "thoughtsTokenCount": 1135, "serviceTier": "standard"}
1,135 thought tokens to write three seven-word lines. If your cost dashboard prices candidatesTokenCount only, it reports this call at about 3% of what Google bills. Promptster's displayed cost has counted thought tokens since 2026-10-22, so the figures above match what the app shows.
Per-run thinking tokens
thoughtsTokenCount per run, runs 1–5 in order ("—" means the field was absent):
| Task | Gemini 4 Argon | Gemini 3.8 Flash | Gemini 3.6 Flash |
|---|---|---|---|
json-order |
—, 248, —, —, — | 377, 544, 640, 408, 506 | 872, 756, 1,056, 791, 877 |
stock-floor |
—, 508, —, 723, — | 1,446, 833, 1,449, 1,415, 1,354 | 2,149, 2,272, 1,802, 1,789, 1,877 |
seven-word-taglines |
2, 590, —, —, — | 1,360, 892, 1,152, 1,271, 819 | 1,387, 1,135, 1,304, 1,012, 853 |
launch-notes |
—, —, 90, —, — | 285, 305, 297, 351, 224 | 724, 684, 741, 768, 764 |
The Flash models thought on every call, and 3.6 Flash thought more in total than 3.8 Flash on every task. Argon's usage object omitted the field on 14 of 20 calls. On the tagline task, Argon's run 2 reported 590 thought tokens and cost $0.0064; its other four runs cost $0.0005 or less each. Which prompts make Argon think is not something five runs per task can predict, so treat its low average as an average.
The Flash models' thinking also sits inside maxOutputTokens. This battery ran at max_tokens 4,000. In a separate run of stock-floor at 1,500 (content/blog/data/q4b-2026-maxtokens-1500/), both Flash models went 0/5, every call stopped with MAX_TOKENS after 1,326 (3.8 Flash) and 1,438 (3.6 Flash) median thought tokens. A Flash model needs room for its thinking as well as its answer.
The price calendar changes the order twice
From shared/pricing.ts: Argon's $2 / $10 is a promo "until 2026-11-30, then $4.00/$20.00". Both Flash promos run "until 2026-12-31, then $1.50/$7.50". Every rate doubles, so each model's total doubles exactly, whatever its mix of tokens:
| Period | Gemini 4 Argon | Gemini 3.8 Flash | Gemini 3.6 Flash | Cheapest |
|---|---|---|---|---|
| Now to 2026-11-30 | $0.0667 | $0.0742 | $0.1048 | Argon |
| 2026-12-01 to 12-31 | $0.1333 | $0.0742 | $0.1048 | 3.8 Flash |
| From 2027-01-01 | $0.1333 | $0.1483 | $0.2096 | Argon |
For December only, both Flash models are cheaper than Argon on this battery. After that Argon is cheapest again, by the same margins as today. We compared Argon with Claude Opus 5.5 at both rates in October.
Latency
Median wall-clock seconds per call:
| Task | Gemini 4 Argon | Gemini 3.8 Flash | Gemini 3.6 Flash |
|---|---|---|---|
json-order |
4.9 | 5.6 | 4.4 |
stock-floor |
11.0 | 12.5 | 10.4 |
seven-word-taglines |
3.8 | 9.1 | 4.9 |
launch-notes |
2.2 | 3.3 | 3.7 |
"Flash" does not mean faster here either. 3.6 Flash had the lowest median on two tasks, Argon on the other two, and 3.8 Flash was slowest on three. Argon also had the widest spread: 1.9 s to 35.2 s across its 20 calls, against 2.8–13.5 s for 3.8 Flash and 3.3–11.7 s for 3.6 Flash.
Which Gemini to use
- Until November 30, on prompts like these: Argon. It was cheapest overall and on three of four tasks.
- Long worked answers: check your own prompts. On
stock-floor3.8 Flash beat Argon, because visible output is where Argon's higher rate shows. - In December: 3.8 Flash, if the Flash promo is still on. Then re-test in January.
- Always: log
thoughtsTokenCount. It was 84% of 3.6 Flash's bill here, and a cost estimate from list prices and visible output will be wrong for any Gemini model that thinks.
In October, Gemini 3.8 Flash cost 31× GPT-6 Luna on our billed estimate for the same reason. We explained how reasoning tokens are billed in May.
Method
Four script-graded tasks (no LLM judge), five runs per task per model, max_tokens 4,000, temperature 0.7, run 2026-10-23 through Promptster's own provider handlers. cost_usd is calculateCost() from shared/pricing.ts at list rate, including Gemini's thoughtsTokenCount at the output rate, with no cache or batch discounts. December and January figures double the measured totals. Latency is wall-clock from one machine in one session, not a Google SLA. For how these compare with open-weight models, see Benchmarking open-source vs. closed-source models.
Run this comparison yourself — no account, no API key
- Run the seven-word tagline task, where 3.8 Flash spent 819–1,360 thought tokens to write about 25 visible ones
To run the three Gemini models side by side on your own prompts and see thoughtsTokenCount per call, that's Promptster.
All figures from content/blog/data/q4b-2026-battery/ (summary.json plus one raw file per call, with Google's usageMetadata verbatim), run 2026-10-23. Token sums and the thinking-cost split are computed from those files at the shared/pricing.ts rates; they add up to each model's cost_usd total. Promo end dates are from the comments in shared/pricing.ts.