Cosmos Reason 2 API pricing: three sizes, three different answers

Cosmos Reason 2 API pricing has three answers, one per size, and search mixes them up. As of 27 October 2026 the 8B model runs on NVIDIA's own hosted tier at no charge, inside rate limits. The 32B carries published per-token rates at one host and an hourly dedicated rate at another. The 2B has no third-party hosted row on any surface this page checked, and the one listed per-token rate for it sits on this platform's chat route. This page shows the dated census behind those three sentences, decodes the numbers search actually returns, and works the token math for a grounding loop.

Last verified: 2026-10-27

What Cosmos Reason 2 is, and why one price does not exist

Cosmos Reason 2 is NVIDIA's reasoning vision-language family for physical AI. The official documentation describes an open, customizable model that understands space, time, and basic physics, reasons in chains of thought, and can act as a planning layer for an embodied agent. The family ships in three sizes. The 2B card is the small one, with 566,091 downloads in the last month on the model host. The 8B sits in the middle at 85,403 downloads in that window. The 32B is the flagship at 2,482 downloads in that window. All three accept images and video and return text. All three carry the NVIDIA Open Model License, which permits commercial use and derivative models.

Three sizes, three serving stories, and that is why "the Cosmos Reason 2 price" is not one number. The sizes land on different hosts because their hardware needs differ by more than an order of magnitude. The 32B page at one EU host lists a 160 GB minimum memory requirement. The 2B runs on a laptop-class GPU. A market that serves one size cheaply has no reason to serve all three. The census below shows that nobody checked today serves all three.

The platform side of this guide is the 2B. Its model page here, cosmos-reason2-2b, carries the capability record: what the model scores, what hardware it fits, and its measured behavior. This page is about money.

The host census, checked today

Here is what each surface that looks like a seller actually holds. Every cell was read on 27 October 2026.

Three notes on the table. First, the 2B line. On every surface above that could carry a hosted row, the 2B did not appear on this date. That is an observation with a date, not a prediction: new platforms appear, and any claim that the 2B will never be hosted elsewhere would be a guess. Second, the absence is not a license wall. The NVIDIA Open Model License lets anyone serve the model commercially, and the AWS listing itself states users are free to create and distribute derivative models. What is missing is a seller who has done the work and published a rate. Third, free is not the same as priced. The 8B row at NVIDIA is a free tier with limits, and the same 8B bills by the host hour at AWS. Neither is a per-token price.

The numbers search results show you, decoded

The money queries return a mix of three models and two billing units. For "cosmos reason 2 api pricing" on 27 October 2026, the top rows told a story. The Featherless 32B per-token page ranked first. The AWS 8B host-hour listing came second. Then came an API example page for Reason 1, the previous generation, a Jetson tutorial, a container page for the 2B (software to run yourself, not a hosted row), the GitHub repository, the 8B card, the pi.dev wrapper, and the HostYourAI page. Not one row was a per-token price for the 2B.

The neighboring queries drift further. "Cosmos reason api cost" mostly returns Reason 1 pages. One is a 2024-era AWS listing for the 7B. "Nvidia cosmos api price" returns the Cosmos world-model platform, which generates video and is a different product line. An AI overview for that query states that NVIDIA publishes no per-token or per-call price for Cosmos on its official platform. It describes free prototyping credits instead. An earlier AI overview for the pricing query mixed three things in one answer. One was a rate for Cosmos 3 Super Reasoner from a gateway. One was the 32B per-token pair at Featherless, with the model mislabeled as 33.4 billion parameters. One was the zero-rate agent wrapper at pi.dev. All three statements were about real pages. None was a 2B price.

So the practical decode: when a search result shows you a Cosmos number, first ask which model, then ask which unit. Per-token rates you can divide by a million belong to the 32B at one third-party host. Host-hour rates that rent you a whole instance belong to the 8B at AWS. Free-with-limits belongs to the 8B at NVIDIA's own tier. The 2B column is empty on every one of those surfaces.

The free tier, read closely

The free 8B tier is real. It is worth reading closely, because "free" carries two limits that matter to a robot builder. NVIDIA's Build copy pins the request limits. Up to 40 requests per minute. Up to 10,000 requests per day. A note adds that limits may vary by model, and that traffic from other users may throttle you. A directory that tracks these listings, read the same day, agrees: the 8B is free at NVIDIA Build as of 27 October 2026. An agent wrapper over NVIDIA's own endpoint shows a rate table of zero across the board, with a cost calculator built on cache reuse.

Now the two limits. First, rate. 10,000 requests a day sounds large until a perception loop polls once a second. That is 86,400 a day, far past the cap. Second, shape. A free prototype tier is not a published per-token rate. There is no per-token number to budget against, no cache column, and no commitment that the free shape holds. For evaluation and demos it is generous. For an unattended loop it is a ceiling you hit on day one. The line between free and priced is the whole reason this page exists. Priced things can be budgeted, and budgeting is the real job when robots run unattended.

The priced sizes, at their published rates

Two sizes carry published numbers today, in two different units.

Per token, the 32B at Featherless: 0.416 US dollars per million input tokens and 1.664 per million output, read 27 October 2026 on the model page linked in the census. Both columns of that pair sit at more than four times the fenced 2B columns below. The same host's own card data lists a 32,768-token context and FP8 serving.

Per host hour, the 32B at HostYourAI: a dedicated vLLM instance on European cloud. The memory floor is 160 GB. The hourly rate read just over six euro on 27 October 2026. The shared per-token route for the same model reads on request. The dedicated rate had moved from about four euro an hour when the market was last sampled on 25 October. Hourly GPU rates drift weekly.

Per host hour, the 8B at AWS: 1.00 US dollars per host hour on the small g6e instance sizes, 4.00 on the mid sizes, and 8.00 on the largest. That covers the software layer. AWS infrastructure bills separately. One host hour is a whole instance, and it serves as many requests as that instance can carry. For scale across units, one small-size host hour equals more than three thousand queries at the fenced rate below.

Google runs the one full public rate card for embodied reasoning from a major vendor. The line-by-line read of it is the Gemini Robotics rate card guide.

The one listed per-token rate for the 2B

This platform lists the cosmos-reason2-2b model id on its chat route, and it has been listed since 6 October 2026. Metering is per million tokens, the same unit family as the 32B pair above. The served context window is 4,096 tokens per request. The models index is the live catalog, and the current rate always renders on the pricing page.

Two platform facts that matter to a grounding loop. The same scene recurs across queries, and a repeated input prefix bills at the cached rate in the fence, not the fresh rate. And spend runs against a prepaid wallet whose balance is the hard stop: a refused request bills nothing, so an unattended loop cannot dig a hole. The argument for that shape is the API bill shock guide.

The request shape is the OpenAI chat form. The strings you type:

POST /v1/chat/completions
Authorization: Bearer <your key>

{
  "model": "cosmos-reason2-2b",
  "messages": [
    {"role": "user", "content": [
      {"type": "text", "text": "Describe the scene and the next safe action."},
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,<frame bytes>"}}
    ]}
  ],
  "max_tokens": 512
}

The response carries the answer in choices[0].message.content. The accounting sits in usage, which reports token counts plus usage.cost. That is the charge for that exact call at published retail. GET /v1/models lists every public alias with its pricing and context fields, so the model string never has to be guessed. The full compatibility contract, including what happens to unknown parameters, is the OpenAI-compatible API guide.

The worked math, tokens first

The budget is math on four named inputs:

cost per query  =  ( I x Pin + O x Pout ) / 1,000,000

I:     input tokens per query, image tokens plus text, capped by the 4,096 window
O:     output tokens per query, reasoning included
Pin:   input rate from the live table above
Pout:  output rate from the live table above

A labeled example: one camera frame plus a task sentence, about 1,000 input tokens, and an answer with reasoning, about 500 output tokens. Those two figures are examples, not measurements; put your own counts in. The math: 1,000 times the fenced input rate per million is 0.0001. Add 500 times the fenced output rate per million, 0.0002. The total is 0.0003 per query. That is 0.30 per 1,000 queries, and about 9 US dollars for a 30 day month at a thousand queries a day. With the full input prefix cached at the cached rate, the same query costs 0.00005 plus 0.0002, so 0.00025.

One sensitivity line. Output bills at four times the input rate in the fenced row, so the output token count is the lever that matters most. Cut the answer length in half, to 250 output tokens, and the example query cost falls from 0.0003 to 0.0002, a third off. For the conversion between token budgets, per-call figures, and per-robot-hour figures, read the pricing units guide.

Running the open weights yourself

The license lets you self-host, and the official reference gives a one-line serve command:

uv run vllm serve nvidia/Cosmos-Reason2-2B \
  --max-model-len 8192 \
  --reasoning-parser qwen3 \
  --port 8000

The model is also in transformers from version 4.57.0. What self-hosting costs is a card that fits the weights and a person who keeps it running. The memory ladder, with each number owned by its source, is on the model page. For the measured economics of consumer cards, in watts and in revenue terms, see the RTX 3060 budget economics guide and the RTX 3090 account-rate guide. The 2B fits that class. The 32B, at its 160 GB floor, never will; that is why its only hosted shapes are a whole rented GPU.

For the cloud-versus-local tradeoff, read the Jetson versus cloud guide.

What to do now

1. Name the size before you budget. The three sizes have three different answers. This page's fenced row is the 2B; the 8B and 32B rows live at their hosts, dated above.
2. Budget from the fence, not from search results. The current listed rate always renders on the pricing page. The rate this page fenced is the read of 27 October 2026.
3. Write the client before you spend. The chat route is OpenAI compatible and the model id is cosmos-reason2-2b. Start at the quickstart and the OpenAI-compatible guide. Your driver can be finished and tested before any spend begins.
4. Re-check the census dates before you cite a host. If a seller appears, the census table is the part that goes stale first. Hourly GPU rates moved inside two days this month.
5. Start small on purpose. You can sign up and create a key without talking to anyone. The prepaid balance is the spend stop.

Frequently asked questions

Is there a Cosmos Reason 2 API?

Yes, more than one, and they are different products. NVIDIA's own hosted tier serves the 8B model id at no charge inside request limits. A third-party host serves the 32B per token, and another rents it by the GPU hour. AWS lists the 8B as a host-hour deployment. This platform's chat route serves the 2B per token under the model id cosmos-reason2-2b. All verified 27 October 2026.

What does a Cosmos Reason 2 query cost?

For the 2B on this platform: the fenced row above, worked through the formula section, lands at 0.0003 for the labeled 1,000-in, 500-out example, less with caching. The live figure always renders on the pricing page. For other sizes, the unit differs by host: per-token at one, per host hour at two others, free-with-limits at NVIDIA.

Is the 8B model really free?

At NVIDIA's own tier, yes, within limits: up to 40 requests per minute and 10,000 per day, per NVIDIA's Build copy read 27 October 2026. It is a prototype tier, not a published per-token rate, and there is no per-token number to budget against. The same 8B bills 1.00 US dollars per host hour and up at AWS. Free and priced are different products with different shapes.

Why does search show a Cosmos price but never a 2B price?

Because the query space is polluted in both directions. Older Reason 1 pages, the Cosmos 3 generation, and the world-model video platform all rank for these queries, and the pages that do carry Reason 2 numbers are about the 8B or the 32B. No surface checked on 27 October 2026 carried a third-party per-token row for the 2B. The census table is the part to re-check.

Can I run Cosmos Reason 2 myself?

Yes. The weights are open under the NVIDIA Open Model License, which permits commercial use, and the official reference gives a one-line vLLM serve command. The 2B fits consumer cards; the 32B wants 160 GB. What you trade for the card hour is the person who keeps it running: the measured card economics are linked above.

The short version: three sizes, three answers, all dated. The 8B is free with limits at its maker's tier. The 32B is priced per token at one host and by the hour at two more. The 2B has one listed per-token rate, it is fenced above, and the live figure is on the pricing page. Sign up to create a key and budget metered queries.

Last verified: 2026-10-27