Cheapest models by task: what each kind of work actually costs

The lowest multiplier is not the lowest price

A multiplier is applied to that model’s own official per-token rate, so two multipliers are only comparable when the underlying rates match. Across the catalog they rarely do, which makes multiplier-shopping actively misleading.

The clearest case: grok-composer-2.5-fast carries the lowest multiplier here, 0.05×. But its official rate is $3 per million input tokens, so it bills at $0.15. gemini-3-flash is 0.1× — twice the multiplier — on a $0.50 official rate, so it bills at $0.050. That is three times cheaper than the model with the lowest multiplier in the catalog.

Every number below is an effective rate: official rate × multiplier, the price you are actually billed. Multiply first, then compare.

High-volume, short answers

Classification, extraction, routing, tagging — work where you send a lot of requests and each answer is a few dozen tokens. Input dominates the bill, so the input rate is the number to sort on.

gemini-3-flash is the lowest input rate in the catalog at $0.050 per million (0.1× on $0.50), and it also has the lowest output rate at $0.30 per million. It carries a 1,000,000-token context and lists text, tools and thinking. Next up: gpt-5.6-luna at $0.06 input (0.3× on $0.20 — a higher multiplier than most, on a lower base), then gemini-3.7-flash-high and gemini-3.6-flash-high at $0.075, then gpt-5.4-mini at $0.075.

Those four are within 1.5× of each other on input, which at this volume is real money but not a category difference. What separates them is not in the price column: route (/v1/chat/completions for the Gemini ids, /v1/responses for the GPT ids) and whether the model gets your task right often enough that you are not paying twice. A cheap model that needs a retry costs more than the expensive one that did not.

# gemini-3-flash: $0.050 / M input, $0.30 / M output (0.1x on $0.50 / $3)
curl -sS https://token-share.app/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "gemini-3-flash",
    "messages": [
      {"role": "user", "content": "Classify this ticket as bug, feature, or question. Reply with one word.\n\nTicket: the export button downloads an empty CSV"}
    ]
  }' | jq '{output: .choices[0].message.content, usage}'

Long inputs, output-heavy work, and everything with tools

When the input is large, the context ceiling and the input rate decide together. Five ids hold 1,000,000 tokens at a Flash-tier price: gemini-3-flash at $0.050, gemini-3.7-flash-high and gemini-3.6-flash-high at $0.075, gemini-3.8-flash at $0.15 (0.2× on $0.75), gemini-3.1-pro-low at $0.20. Above them, claude-sonnet-5 holds 1,000,000 tokens at $0.30 and raises the output ceiling from 65,536 tokens to 128,000.

When the output is large, sort on the output rate instead, and check the output ceiling. The five Gemini ids cap at 65,536 output tokens (65,535 for gemini-3.1-pro-low). If you need more than that in one response, the cheapest ids that reach 128,000 output tokens are gpt-5.6-luna at $0.36 per million output, then gpt-5.6-terra at $1.20, then claude-sonnet-5 at $1.50.

Tools are not a differentiator: 27 of the 31 catalog entries list tools, and the four that do not are the image and video models. Thinking is narrower — 25 entries list it, and the text models without it are stable-claude-haiku-4-5 ($0.40/$2.00) and grok-composer-2.5-fast ($0.15/$0.75). If you want tools without paying for reasoning tokens, those two are the ids that cannot spend them.

Images and video are billed differently

Image models do not follow the per-token pattern, and they are also where the multiplier stops helping. grok-imagine-image is $0.01 per image and grok-imagine-image-quality is $0.04 per image, both at a 1× multiplier — no discount at all, the only two entries in the catalog priced that way. gpt-image-2 is a 0.1× model priced per token at $5/$30 official, so $0.50 and $3.00 per million.

grok-imagine-video-1.5 is billed per video second at 0.1×, so the length of the clip, not the length of a prompt, is what moves the bill.

Comparing the three image ids against each other therefore means comparing two different billing units. A per-image price is predictable in a way a per-token price is not; a per-token price scales with the size of what you send. Which is cheaper for you depends on how large your prompts are and how many images you generate per prompt — count both before assuming.

Common questions