Model pricing explained: how the multiplier produces your bill

One formula, applied per model

Consumer price = official per-token rate × that model’s rateMultiplier. The official rate is the upstream vendor’s published price for that model; the multiplier is the factor we apply to it. Both are recorded per model in the catalog, and both are visible on every model page.

A worked example. claude-sonnet-5 is $3 per million input tokens and $15 per million output at the official rate, with a 0.1× multiplier. A request that reads 50,000 tokens and writes 8,000 costs (50,000 ÷ 1,000,000 × $3) + (8,000 ÷ 1,000,000 × $15) = $0.27 official, and $0.027 after the multiplier.

The same request on gpt-5.6-sol — $5 and $30 official, also 0.1× — is $0.49 official and $0.049 billed. On gemini-3-flash, $0.50 and $3 official at 0.1×, it is $0.049 official and $0.0049 billed. Same arithmetic, three different results, because the base rates differ.

The multiplier is not a ranking

This is the mistake that costs money. Because each multiplier is applied to a different base rate, a lower multiplier can produce a higher bill. gpt-5.6-luna carries 0.5× on a $0.20 official input rate and bills at $0.10 per million. gpt-5.6-sol carries 0.1× — five times better-looking — on a $5 official rate, and bills at $0.50. The 0.5× model is the cheaper one.

It happens in the other direction too. stable-claude-haiku-4-5 is 0.4× on a $1 official input rate, so $0.40 per million. claude-sonnet-5 is 0.1× on $3, so $0.30. The model with the smaller-looking multiplier costs less than the one that looks four times more discounted.

And the case a multiplier column hides completely: every Grok id carries the same 0.2×, so sorted by multiplier they tie. On price they do not. grok-composer-2.5-fast lists at $3 and $15 per million, so it bills $0.60 and $3.00; grok-4.7, grok-4.6 and grok-4.5 list at $2 and $6 for prompts under 200K tokens and bill $0.40 and $1.20. Same multiplier, 1.5× the input price and 2.5× the output. Multiply first, then compare. Never sort a catalog by multiplier.

What the multipliers actually are

The fair summary is "as low as 0.1×; Claude at 0.25–0.35×; 0.2× for other mainstream pool models", and each part of it names a family rather than the whole catalog. The 0.1× floor is gpt-5.6-sol and gpt-image-2. Every Grok id is 0.2× — text, image and video alike — and so are gemini-3.8-flash and stable-gpt-6-astra. gpt-5.6-terra and stable-gpt-5.6-terra are 0.3×. Every stable-* Claude lane is 0.4×. gpt-5.6-luna is 0.5×, the highest in the catalog.

The Grok Imagine image models are no exception, even though they bill per image rather than per token: 0.2× of $0.01 and $0.04 official is $0.002 and $0.008 per image. Forty standard images come to $0.08 and forty quality images to $0.32. Any statement that "everything here is 0.1×" is wrong — wrong on every Grok id, and wrong most sharply on gpt-5.6-luna at five times that.

The stable-* lanes are the other place the multiplier rises. A stable-* id is a dedicated lane served on our own key rather than the shared account pool, and it does not participate in pool failover. Every Claude stable lane is 0.4×, and stable-gpt-6-astra is 0.2×. So claude-sonnet-5 at $0.30/$1.50 and stable-claude-sonnet-5 at $0.80/$4.00 are the same weights on different routing, priced accordingly.

What is counted, and how to check a bill

Input and output are billed at different rates, and the gap is large: across the catalog output runs 3× to 6× input. grok-4.7, grok-4.6 and grok-4.5 have the narrowest ratio at 3× ($0.40 against $1.20 after the multiplier); gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna are at 6×; the other 14 of the catalog’s 20 text ids, including every Claude id and grok-composer-2.5-fast, are at 5×. Anything that lengthens a response is charged on the expensive side.

Cached input is priced separately from fresh input, at its own per-token rate below the standard input rate, and the multiplier applies to it the same way. That is the lever for a long prefix that genuinely repeats across turns. Media models step outside the per-token model entirely: the two grok-imagine-image ids bill per image, and grok-imagine-video-1.5 bills per video second at 0.2×, so clip length rather than prompt length drives that bill.

To check any of this yourself, read the token counts back from the response and multiply. Every route reports usage — the Anthropic shape as input_tokens and output_tokens with cache_read_input_tokens and cache_creation_input_tokens broken out, the OpenAI shape as prompt_tokens and completion_tokens with cached_tokens inside prompt_tokens_details. Multiply each bucket by its own effective rate and the total should match the charge.

# claude-sonnet-5: official $3 / $15 per M, multiplier 0.1x
curl -sS https://token-share.app/v1/messages \
  -H "x-api-key: $TOKEN_SHARE_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Summarise this changelog in three bullets."}]
  }' \
| jq '.usage as $u
      | {usage: $u,
         billed_usd: (($u.input_tokens  / 1000000 * 3  * 0.1)
                    + ($u.output_tokens / 1000000 * 15 * 0.1))}'

# Every model publishes its own rate and multiplier — read them from the catalog:
curl -sS https://token-share.app/v1/models \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
| jq '[.data[] | {id, endpoint, context_window, max_output_tokens}]'

Common questions