Gemini 3 Flash: the lowest per-token price in the catalog

The number, and who comes next

gemini-3-flash carries an official rate of $0.50 per million input tokens and $3 per million output, billed at the 0.1x multiplier. That works out to $0.050 per million input and $0.30 per million output. Across every text id in this catalog, no other model is lower on either figure.

The runners-up put that in proportion. gpt-5.6-luna is next at $0.06 and $0.36 per million — 1.2x on both, from a $0.20/$1.20 official rate at 0.3x. Then gpt-5.4-mini at $0.075 and $0.45 (1.5x on both), and gemini-3.7-flash-high and gemini-3.6-flash-high at $0.075 and $0.375 (1.5x input, 1.25x output).

For scale against the mainstream ids: claude-sonnet-5 costs $0.30 and $1.50 per million, six and five times gemini-3-flash. grok-4.6 costs $0.20 and $0.60, four and two times. Those margins are narrow enough on the output side that a job which writes a lot should be priced rather than assumed.

The cheap id is not a cut-down id

Low-priced models often trade away specifications, and this one mostly does not. gemini-3-flash takes a 1,000,000-token context — the same window as claude-opus-5 and every other Gemini id here — and emits up to 65,536 tokens. It lists text, tools and thinking, so tool-calling agents and reasoning prompts are both on the table.

It is the oldest Gemini id in the catalog, released 2025-12-17 against 2026-03-01 for gemini-3.1-pro-low and 2026-09-02 for gemini-3.8-flash. Age is what the catalog records; it does not record a quality figure, so treat the date as a fact about the release and not as a proxy for how the model performs.

Cached reads are $0.05 per million against the official rate, which is $0.005 per million at 0.1x — a tenth of the fresh-input rate. On a repeated prefix that is the difference between paying for the same system prompt every turn and paying a tenth of it.

curl https://token-share.app/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "gemini-3-flash",
    "max_tokens": 256,
    "messages": [
      {"role": "system", "content": "Classify each ticket as bug, feature, or question. Reply with one word."},
      {"role": "user", "content": "The export button does nothing on Safari."}
    ]
  }' | jq '{content: .choices[0].message.content, usage}'

When the lowest price is the wrong choice

A price advantage is only realised if the cheaper id finishes the job. If a task needs two attempts on gemini-3-flash and one on a model at 4x the rate, the cheap id has cost you half of the expensive one — still cheaper, but not by the ratio on the rate card, and that is before the latency of the retry and the code that has to detect it.

Where the ratio flips entirely is work where a wrong answer costs more than the tokens do: a migration script, a billing calculation, anything a human then has to check. At $0.30 per million output, the token cost of a whole day of requests can be less than one engineer-hour spent unpicking a bad one. That is not an argument for the expensive id either — it is an argument for measuring instead of assuming.

The way to settle it is to run your own work through it. Take a sample of real tasks with known-good answers, run them through gemini-3-flash, and count how many come back acceptable. That accuracy figure against the 4x or 6x price gap is the actual decision; a general ranking is not.

Common questions