Same Open Model, Different Host: What a Month Costs

An open-weights model is sold by several companies at prices each of them sets. The provider comparison ranks those prices per million tokens, and when to compare hosts says which models are worth the effort. This post asks what one fixed month of traffic costs on the cheapest and the most expensive host of each model.

The month used throughout is 100 million input tokens and 20 million output tokens: a five-to-one mix, standard serverless prices, no prompt caching and no Batch discount. Every price below was read on the host’s own pricing page and verified on October 12, 2026. Nothing here says why a host charges what it charges; where a page explains something, it is quoted.

Six models, two kinds of answer

Six open models in this dataset have two or more hosts with a verified price and a page of their own here. For each, the bill for the same month:

Model	Hosts priced	Cheapest host, month	Most expensive host, month	Difference
Llama 3.3 70B Instruct	2	DeepInfra $16.40	Together AI $124.80	$108.40 (7.61×)
DeepSeek V4 Flash 0731	2	DeepInfra $9.60	Together AI $19.60	$10.00 (2.04×)
Kimi K3	4	DeepInfra $570.00	Together AI, Fireworks AI, OpenRouter $600.00	$30.00 (1.05×)
GLM-5.3	3	$228.00 on every host	$228.00 on every host	none
gpt-oss-120b	3	$27.00 on every host	$27.00 on every host	none
MiniMax M3	3	$54.00 on every host	$54.00 on every host	none

Most of them do not move. Three cost exactly the same wherever they run, and Kimi K3 differs by five per cent. Two move a lot, and one of them accounts for almost all the money: run one month of each of the six on its cheapest host and the total is $905.00; on its most expensive host it is $1,053.40. Of the $148.40 between the two, $108.40 is Llama 3.3 70B Instruct alone.

The site-wide count says the same thing from further away: of the 8 models in this dataset that two or more hosts price, 6 stay within 1.15× on input price from their cheapest to their most expensive host. The two outside that group are Llama 3.3 70B Instruct and DeepSeek V4 Flash 0731, the two that move here.

Llama 3.3 70B Instruct: the one that matters

The two pages, as read on October 12, 2026:

• DeepInfra: Llama-3.3-70B-Instruct-Turbo | 128k | $0.10 | $0.32 | View more
• Together AI: Llama 3.3 70B | $1.04 | $1.04 |

Together AI charges the same $1.04 for a million tokens in and a million out; DeepInfra charges $0.10 in and $0.32 out. On input alone that is the 10.4× spread this site reports for the model. The bill does not grow by 10.4×, because the output gap is smaller: $1.04 against $0.32 is 3.25×. The month works out like this:

Host	Input, 100M tokens	Output, 20M tokens	Month
DeepInfra	$10.00	$6.40	$16.40
Together AI	$104.00	$20.80	$124.80

That is 7.61 times the bill, or $108.40 a month for a workload this size, and the ratio moves with the mix: the more output in the month, the closer the two hosts get, because output is where their prices are closest. At ten times the volume, a billion input tokens and 200 million output, the difference is $1,084.00 a month.

DeepSeek V4 Flash 0731: a real ratio, a small bill

• DeepInfra: DeepSeek-V4-Flash-0731 | 1024k | $0.06 / $0.015 cached | $0.18 | View more
• Together AI: DeepSeek V4 Flash 0731 | $0.14 $0.03 (cached) | $0.28 |

Input differs by 2.33× and output by 1.56×, so the month costs $9.60 on DeepInfra ($6.00 in, $3.60 out) and $19.60 on Together AI ($14.00 in, $5.60 out): 2.04 times as much, but $10.00 in absolute terms. Whether ten dollars a month is worth a second integration depends on the volume, not on the ratio. OpenRouter also lists this model, but DeepSeek itself is not one of its providers for it, so there is no maker’s list price there to quote, and this site shows OpenRouter as “no list price” for it.

Kimi K3: four hosts, five per cent

• DeepInfra: Kimi-K3 | 1024k | $2.85 / $0.285 cached | $14.25 | View more
• Together AI: Kimi K3 | $3.00 $0.27 (cached) | $15.00 |
• Fireworks AI: Kimi K3 | $3.00 / $0.30 / $15.00 | $3.75 / $0.375 / $18.75 |

OpenRouter publishes Moonshot AI’s own list price for it, $3.00 / $15.00. The month costs $570.00 on DeepInfra and $600.00 on the other three. Two things on these pages change that number more than the choice of host does. Together AI’s page still shows a lower price labelled “Promotion valid until October 11, 2026”: at its $2.70 / $13.50 the same month would have cost $540.00, and that date has passed. Fireworks AI prints a separate row, Kimi K3 (US) | $4.50 / $0.45 / $22.50 |, at which the month costs $900.00.

GLM-5.3 and MiniMax M3: identical everywhere

GLM-5.3 is $1.40 input and $4.40 output on Together AI (GLM-5.3 | $1.40 $0.26 (cached) | $4.40 |), on Fireworks AI (GLM 5.3 | $1.40 / $0.26 / $4.40 |) and at Z.ai’s list price on OpenRouter: $228.00 for the month on all three. MiniMax M3 is $0.30 and $1.20 on Together AI (MiniMax M3 | $0.30 $0.06 (cached) | $1.20 |), on Fireworks AI (MiniMax M3 | $0.30 / $0.06 / $1.20 |) and at MiniMax’s list price on OpenRouter: $54.00 on all three.

gpt-oss-120b: three hosts, one price

OpenAI’s open-weights model has three priced hosts in this dataset, and all three print the same figures:

• Together AI: OpenAI | GPT-OSS 120B | openai/gpt-oss-120b | 131072 | $0.15 | - | $0.60 | MXFP4 | Yes | Yes
• Fireworks AI: OpenAI GPT OSS 120B | $0.15 / $0.015 / $0.60 | $0.18 / $0.018 / $0.72 |
• Groq: GPT OSS 120B openai/gpt-oss-120b | 500 | $0.15 input $0.60 output

At $0.15 input and $0.60 output per million tokens, the month is $15.00 in and $12.00 out, $27.00 on each host. The three bills are identical to the cent.

What changes the bill for gpt-oss-120b is not the host but the price list within one host. Fireworks AI explains its two columns: “In each Standard or Priority cell, prices are input / cached input / output (USD per 1M tokens), in that order.” At the Priority price, $0.18 / $0.72, the same month costs $32.40. The same page says “Batch inference is billed at 50% of serverless pricing on both input and output”, which would make it $13.50. Its Standard cell also prints a cached input price, $0.015; this month assumes no cache hits.

What to do with this

• For most open models, pick the host on something other than price. On four of the six, the host changes the bill by nothing or by five per cent.
• Check the two that move. Llama 3.3 70B Instruct is where a default host costs real money: $108.40 a month at this size.
• Price your own mix, not the input ratio. The 10.4× input spread on Llama 3.3 70B became 7.61× on a five-to-one month; an output-heavy workload would narrow it further.
• Read the whole row. A promotion, a US-only row or a Priority column can move the bill more than switching hosts.

How this was read

Every host price comes from this site’s dataset, where each offering links to the host page it was read on, verified on October 12, 2026; the quoted lines were read that day (Together AI’s Kimi K3 line by hand, after its promotion had ended). OpenRouter prices are the model maker’s list price on OpenRouter’s providers page. The monthly figures are tokens divided by a million, times the price, with no caching, Batch or Priority unless the paragraph says so, in US dollars.

Frequently asked questions

How much does the host change the monthly bill for an open model?

For four of the six open models with two or more priced hosts on October 12, 2026, very little: GLM-5.3, gpt-oss-120b and MiniMax M3 cost the same on every host, and Kimi K3 varies by $30.00 on a $600.00 bill. For Llama 3.3 70B Instruct the same 100 million input and 20 million output tokens cost $16.40 on DeepInfra and $124.80 on Together AI.

Why is the gap on the bill smaller than the gap on input price for Llama 3.3 70B?

Because input and output spread differently: Together AI charges 10.4 times DeepInfra's input price but 3.25 times its output price, so a month with 20 million output tokens costs 7.61 times as much, not 10.4 times.

Does gpt-oss-120b cost less on Groq, Fireworks AI or Together AI?

None of them: all three print $0.15 input and $0.60 output per million tokens, so the example month costs $27.00 on each. Fireworks AI also prints a Priority price of $0.18 / $0.72, which makes the same month $32.40, and says Batch inference is billed at 50% of serverless pricing.