DeepSeek V4 Flash, gpt-oss-120b and Llama 3.3 70B Pricing by Host: Up to 7.3× Apart on October 29, 2026

Short answer. Read on October 29, 2026, three open models cost more than 1.15× as much on one host as on another. DeepSeek V4 Flash 0731: $0.06 per million input tokens on DeepInfra, $0.44 on Novita AI (7.3×). gpt-oss-120b: $0.05 on Novita AI, $0.15 on Fireworks AI, Groq and Together AI (3.0×). Llama 3.3 70B: $0.10 against $0.135 (1.4×).

In plain words. Open models can be run by many companies. Each sets its own price per token. A token is a short piece of text, about three-quarters of a word. For most open models the price is the same everywhere. For three, one company charges more than seven times what another does.

An open-weights model is the same file wherever it runs: the weights a host downloads are the ones the maker published. What changes from host to host is the price list. On October 29, 2026 AI Signal added two hosts to its dataset, Novita AI and Baseten, and the comparison of host prices changed with them. Three models now show a gap above the 1.15× line that AI Signal uses to call two prices “the same”. This post sets out those three, host by host, with the row each host’s page printed that day.

Every figure here is per million tokens, in US dollars, as each page printed it. The “example request” column is our arithmetic on those figures: one request of 10,000 input tokens and 1,000 output tokens, with no cache hits. Nothing here says which host to use; the pages quoted below are the hosts’ own.

The three gaps above 1.15×

The ratio compares input prices, the same measure as the open-model prices by host page, rounded to one decimal. On gpt-oss-120b three hosts share the highest price, so the table names all three. On Llama 3.3 70B the two hosts do not serve the same context, which the section below shows.

DeepSeek V4 Flash 0731: $0.06 to $0.44

The rows as each page printed them on October 29, 2026:

• DeepInfra’s pricing page: “DeepSeek-V4-Flash-0731 | 1024k | $0.06 / $0.015 cached | $0.18 | View more”.
• Baseten’s pricing page, which prints each name and figure twice: “DeepSeek-V4-Flash-0731 $0.13 $0.13 $0.028 $0.028 $0.26”.
• Novita AI’s pricing page: “DeepSeek V4 Flash 0731 | 1M | $0.44 /Mt · Cache Read $0.028 /Mt | $1.32 /Mt | More |”.

The cached input prices are the same on Baseten and Novita AI, $0.028, while their input prices differ more than threefold. Output follows input: $1.32 on Novita AI is 7.3× DeepInfra’s $0.18, the same ratio as on input. In the example request the three hosts charge $0.00078, $0.00156 and $0.00572. OpenRouter also lists this model, but it has no list price there: no endpoint is the model’s maker, so it is shown apart on the DeepSeek V4 Flash 0731 page.

gpt-oss-120b: three hosts tied at $0.15

The rows as each page printed them on October 29, 2026:

• Novita AI’s pricing page: “OpenAI GPT OSS 120B | 128K | $0.05 /Mt | $0.25 /Mt | More |”.
• Baseten’s pricing page: “GPT OSS 120B $0.10 $0.10 - - $0.50”.
• Fireworks AI’s pricing page: “OpenAI GPT OSS 120B | $0.15 / $0.015 / $0.60 | $0.18 / $0.018 / $0.72 |”. The same page explains the cells: “In each Standard or Priority cell, prices are input / cached input / output (USD per 1M tokens), in that order.” The table above uses the Standard cell.
• Groq’s models page: “GPT OSS 120B openai/gpt-oss-120b | 500 | $0.15 input $0.60 output”. Its model page for gpt-oss-120b adds the cached price: “PRICING Input $0.15 6.7M / $1 Cached Input $0.075 13M / $1 Output $0.60 1.7M / $1”.
• Together AI’s serverless models page: “OpenAI | GPT-OSS 120B | openai/gpt-oss-120b | 131072 | $0.15 | - | $0.60 | MXFP4 | Yes | Yes |”.

Until October 29, 2026 the three hosts at $0.15 were the only ones priced in AI Signal’s dataset for this model, and the gpt-oss-120b page showed no spread at all. Novita AI’s $0.05 and Baseten’s $0.10 are what opened the 3.0× gap. In the example request the five hosts charge $0.00075, $0.0015 and, on the three tied hosts, $0.0021.

Llama 3.3 70B Instruct: 1.4×, on different contexts

The rows as each page printed them on October 29, 2026:

• DeepInfra’s pricing page: “Llama-3.3-70B-Instruct-Turbo | 128k | $0.10 | $0.32 | View more”.
• Novita AI’s pricing page: “Llama 3.3 70B Instruct | 12K | $0.135 /Mt | $0.4 /Mt | More |”. Its model list, read the same day, gives “context_size”: 12288 for meta-llama/llama-3.3-70b-instruct.

Novita AI serves it with a 12K-token context; DeepInfra with 131K. The two prices are for the same weights, but not for the same longest prompt: a request longer than 12,288 tokens fits on DeepInfra’s endpoint and not on Novita AI’s, by the figures each host publishes. The Llama 3.3 70B Instruct page shows the context column next to each price.

Seven models priced the same on every host

The other seven models that two or more hosts price in the dataset stayed within 1.15× on October 29, 2026: DeepSeek V4 Pro 0813, GLM-5.3, GLM-5.3 Flash, Kimi K3, MiniMax M3, Muse Spark Open and Qwen3.8 Max. Six of them cost exactly the same on every host; Kimi K3 spans 1.1×, $2.85 to $3.00 per million input tokens. The full tables are on open-model prices by host.

What changed, and when

Two things moved this comparison in October 2026, one on a host’s side and one on ours.

On October 22, 2026, Together AI removed Llama 3.3 70B Instruct Turbo from serverless inference. Its deprecations page, read on October 26, 2026, lists it as “2026-10-22 | meta-llama/Llama-3.3-70B-Instruct-Turbo | meta-models/Muse-Glimmer-30B | Yes |”. Until then Together AI priced it at $1.04 for input and output; the weekly detector last confirmed the row “Llama 3.3 70B | $1.04 | $1.04 |” on October 19, 2026. Against DeepInfra’s $0.10, that was a 10× gap, and it closed with the removal. The same day Together AI removed DeepSeek V4 Flash 0731 too (“2026-10-22 | deepseek-ai/DeepSeek-V4-Flash-0731 | deepseek-ai/DeepSeek-V4.1-Flash | Yes |”), which it priced at $0.14 for input, 2.3× DeepInfra’s $0.06. From October 22 to October 28 no model in the dataset had two current hosts more than 1.15× apart.

On October 29, 2026, AI Signal added Novita AI and Baseten to the hosts it reads: five offerings from Novita AI and seven from Baseten, each with its page’s row and the day it was read. The three gaps in this post come from those additions. They are a change in what the dataset covers, not a price change by any host: the prices of DeepInfra, Fireworks AI, Groq and Together AI read that day are the ones the dataset already had. When Novita AI and Baseten set theirs is not known.

An earlier snapshot of the same question, with the hosts of October 12, 2026 and a monthly bill instead of one request, is in what a month costs on each host. What the hosts charge for besides tokens (prepaid credits, minimum purchases, regional premiums) is in what open-model hosts say besides the token price.

How we checked

On October 29, 2026 AI Signal read DeepInfra’s, Novita AI’s, Baseten’s and Fireworks AI’s pricing pages, Together AI’s serverless models page, Groq’s models page and its gpt-oss-120b model page, and Novita AI’s model list, with the same script the weekly price detector uses, which turns each page into plain text. Together AI’s deprecations line comes from the reading of October 26, 2026, and its old Llama 3.3 70B price from the detector’s check of October 19, 2026. Each reading is saved with its date, and a test checks every quoted row in this post against those saved readings, and every figure in the tables against the rows. The ratios and the example requests are computed from those figures. The cached prices of DeepInfra and Fireworks AI are quoted from their rows; AI Signal’s dataset does not store them yet.

Update, October 29, 2026: Later the same day the dataset began storing those two cached prices, together with DeepInfra’s $0.285 for Kimi K3 and Fireworks AI’s $0.06 for MiniMax M3, as a correction of its own reading: each of those rows already printed the figure.

Frequently asked questions

How much does DeepSeek V4 Flash 0731 cost per million tokens?

Read on October 29, 2026: $0.06 input and $0.18 output on DeepInfra, $0.13 and $0.26 on Baseten, $0.44 and $1.32 on Novita AI, per million tokens. The highest input price is 7.3× the lowest.

What does gpt-oss-120b cost on Together AI, Fireworks AI and Groq?

$0.15 per million input tokens and $0.60 per million output tokens on all three, read on October 29, 2026. Baseten prices it at $0.10 and $0.50, and Novita AI at $0.05 and $0.25.

Is Llama 3.3 70B served with the same context on every host?

No. On October 29, 2026 DeepInfra's pricing page printed '128k' for Llama-3.3-70B-Instruct-Turbo, and Novita AI's model list gave a context_size of 12288 for meta-llama/llama-3.3-70b-instruct, which its pricing page prints as '12K'.

Which open models cost the same on every host?

On October 29, 2026, seven of the ten models that two or more hosts price in AI Signal's dataset stayed within 1.15× on input: DeepSeek V4 Pro 0813, GLM-5.3, GLM-5.3 Flash, Kimi K3, MiniMax M3, Muse Spark Open and Qwen3.8 Max.