Molmo2-ER API pricing: no host sells the ER variant, so here is the only live rate and the real math

Molmo2-ER API pricing has no public number to look up. As of 26 October 2026, no marketplace sells the ER variant at a published rate. The price-shaped results in search belong to the base Molmo 2 models, and the number they quote is a stale database record. Behind it, the listings are not orderable today. This page shows the dated census behind all that. It then gives the Google rate that anchors the embodied reasoning market, the one listed live rate for the molmo2-er model id here, and the worked math for a real perception workload.

Last verified: 2026-10-26

What Molmo2-ER is, and why its price is hard to find

Molmo2-ER is a 4 billion parameter vision-language model from the Allen Institute for AI. Its job is embodied reasoning. The model card lists the skills: scene understanding, pixel-accurate pointing, multi-image and video match-ups, and time reasoning across frames. These are the perception skills a robot action model leans on. The model sits on the Molmo2-4B base. That base pairs a Qwen3-4B language backbone with a SigLIP 2 vision encoder. It arrived in May 2026 and carries an Apache-2.0 license.

On the publisher's own tests it beats every open-weight baseline, plus Gemini Robot-ER 1.5 Thinking and GPT-5, on 9 of 13 embodied reasoning runs. The overall average is 63.8 percent, 17 points over its starting checkpoint. Those are the publisher's numbers, quoted as such. The model is also the vision-language backbone of MolmoAct2, the action model in the paper the card itself cites. For the capability record on this platform, the model's own page is molmo2-er.

Four names trip up every pricing search. The Molmo2 base family, in 4B, 8B, and a fully open 7B O variant, are general vision-language models. Databases attach rates to those. Molmo2-ER is the perception and reasoning model trained from the 4B base. MolmoAct2 is the action model that uses ER as its eyes and words. And Molmo, the 2024 first family, is the only one with an old listing on Replicate. The prices you can find attach to the wrong three names. That ER checkpoint has one public repository, no fine-tunes, and no seller.

The host census, checked today

People search "molmo2-er api pricing", "molmo 2 api pricing", "molmo er api cost", and "embodied reasoning api price". Every one of those wants a rate row. Here is what each place that looks like a seller actually holds, checked on 26 October 2026.

Three notes on this table. First, every row carries its date because host states move. All of it was re-read on the day above. Second, the last rows are new specialist platforms. They did not exist when most absence claims were written, and even they do not carry the ER variant. Any sentence that says nobody will ever host this model is a prediction, not a fact. This page states what is true today, with dates. Third, the absence is not a license wall. Apache-2.0 lets anyone serve the model commercially, royalty free. What is missing is a seller who has done the serving work and published a rate. The wider census of which big hosts serve which robot models has its own guide: do the big hosts serve robot models.

The number search results show you, decoded

Search the family queries and the answer that comes back quotes about 0.20 per million tokens in and 0.20 per million out for Molmo 2. That figure belongs to the base 8B model, not the ER checkpoint. Trace it and it gets weaker. In the capture on 26 October 2026, one AI overview answered an ER-phrased query with the base model's figure. It named two aggregators as sources. Then it cut off mid sentence the moment it reached the ER clause. There is no ER price to finish with. One of the two named aggregators carries no molmo row in its public model list at all. The other shows the base-model page with a no-longer-available banner over its question-section rate. Directory pages repeat the figure with no date. So the number exists, and the listing behind it, today, does not. That is the gap this page fills: an ER-specific answer with dates.

What priced embodied reasoning costs: the Google reference

One embodied reasoning model does carry a full public rate card from a major vendor: Google's Gemini Robotics ER 2. On Google's pricing page, read 26 October 2026, the standard tier meters 1.00 US dollars per million input tokens and 5.00 per million output tokens. Thinking tokens count as output. Those columns hold through 31 December 2026. The same page states in each cell that both figures double on 1 January 2027, to 2.00 and 10.00. Batch runs at half the standard columns. Cached input sits at 0.10 per million through the end of 2026. A free tier exists, with a data-use tradeoff. Free tier inputs are used to improve Google's products. Paid tier inputs are not. The line-by-line read of that card, with its limits and request shapes, is the Gemini Robotics rate card guide. That card anchors what anyone can charge for embodied reasoning today.

The one listed live rate for Molmo2-ER

This platform lists the molmo2-er model id on its chat route. The facts of that row, as of 26 October 2026: the id is listed in the platform's model data and serves today, and it has been listed since 6 October 2026. Metering is per million tokens, the same unit family as Google's card. The served context window is 8,192 tokens. A listed chat row publishes the window it enforces. The models index is the live catalog, and the current rate always renders on the pricing page.

One relation worth writing down. The fenced pair equals Google's standard pair above through 31 December 2026. After Google's stated doubling date, the fenced pair is half the doubled Google pair. Equal today, half after, with both dates printed on Google's own page.

For robot perception loops there is a second lever. The same camera scene recurs across queries. A repeated input prefix bills at the cached rate in the fence, not the fresh rate. In a loop that grounds the same scene fifty times an hour, the fresh input rate is paid once and the cached rate covers the repeats, to the extent the prefix repeats. Spend runs against a prepaid wallet, and the wallet balance is the hard stop. A refused request bills nothing. The argument for that shape under unattended robot loops is the API bill shock guide.

The worked math, tokens first

Robot action APIs price per call. This model returns text, so it meters tokens. 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
O:     output tokens per query, thinking included
Pin:   input rate from the live table above
Pout:  output rate from the live table above

A labeled example query: one camera image plus a task sentence, about 1,500 input tokens, and an answer with reasoning, about 400 output tokens. Those two figures are examples, not measurements. Put your own counts in. The math: 1,500 times the fenced input rate per million is 0.0015. Add 400 times the fenced output rate per million, 0.0020. The total is 0.0035 per query. That is 3.50 per 1,000 queries. With the full input prefix cached at the cached rate, the same query costs 0.00075 plus 0.0020, so 0.00275. That is 2.75 per 1,000. A perception loop doing 200 grounded queries a day for a 30 day month runs about 6,000 queries. That is about 21 at the uncached figure, less with caching.

One sensitivity line. Output bills at five times the input rate in the fenced row. The output token count is the lever that matters most. Cut the answer length in half and the example query cost falls from 0.0035 to 0.0025. The same math at Google's through-2026 standard columns lands on the same figures, because the two pairs are equal until the doubling date. For the conversion between token budgets, per-call pricing, and per-robot-hour figures, read the pricing units guide and the measured robot-hour cost report.

Running the open weights yourself

The Apache-2.0 license makes self-hosting the open alternative. Its official repository supports that directly. Molmo2 is officially supported in vLLM from version 0.15.0, with a Docker path and a serving demo in the repo. The shipped checkpoint stores 32 bit weights, about 19.4 GB of files. A 16 bit conversion roughly halves that. The result fits the 16 GB card class. For that class, the site's measured card economics, in weights and in revenue terms, live in two guides: the RTX 3060 budget economics guide and the RTX 3090 account-rate guide.

Card hours, named party, read 26 October 2026 on the RunPod pricing page: the 24 GB tier rents at 0.27 per hour on the A5000, 0.50 on the RTX 3090, and 0.89 on the RTX 4090. Rates move. Treat every figure as a snapshot with a date. To compare a card hour with the fenced rate you need one more input: seconds per query on your card, from a benchmark you run yourself. Keep the comparison symbolic until you have it:

card cost per 1,000 queries  =  R x t x 1,000 / 3,600

R: hourly rate for the card, from a dated price page
t: seconds per query, from your own benchmark

One sensitivity line. Idle time, restarts, and your own engineering sit on top of the raw card time. Treat the formula's output as a floor at full duty, never a bill.

What to do now

1. 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 26 October 2026.
2. Write the client before you spend. The chat route is OpenAI compatible and the model id is molmo2-er. The request and response shapes are documented today. Start at the quickstart and the OpenAI-compatible guide. Your driver can be finished and tested before any spend begins.
3. Use the cached-input lever. Put the stable scene prefix first so it repeats. Repeats then bill at the cached rate in the fence.
4. Re-check the census dates before you cite a host. If a seller appears, the census table above is the part that goes stale first. 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 Molmo2-ER API?

No public priced one from any marketplace, checked 26 October 2026. Hugging Face lists zero hosted providers for the checkpoint. One multi-provider router carries a base-model page with no orderable rate. Puter's base-model page sits under a no-longer-available banner. The specialist robot platforms serve MolmoAct2, the action model, not the ER model. One exception exists: this platform's own row. The molmo2-er model id serves on the chat route today, and its rate renders in the fence above and on the pricing page.

What does a Molmo2-ER query cost?

In tokens, at the example workload of 1,500 input and 400 output: 0.0035 per query, 3.50 per 1,000, about 21 for a 6,000 query month, less when the scene prefix caches. Those come from the fenced rate read on 26 October 2026. Change either token count and the math follows the formula in the worked section. The pricing page carries the live figure.

Is Molmo2-ER the same model as MolmoAct2?

No. Molmo2-ER is a vision-language model that reads scenes and points at things. MolmoAct2 is the action model built on top of it, the one that drives arms. The ER card cites the MolmoAct2 paper as its downstream user. Pricing units differ too. ER-class models meter tokens when anyone prices them at all. Action models price per call or per action. The action model's own page is the MolmoAct2 pricing guide.

Why does search show a Molmo 2 price but never an ER price?

Because the visible price belongs to the base Molmo 2 models. It survives in databases after the listings behind it stopped being orderable. On 26 October 2026 the figure most often quoted, about 0.20 per million tokens in and out, traces to a delisted page and to directories that repeat it without dates. No database carries an ER row, because no host has ever published one. That is why an ER query gets a base-model answer that cannot finish its own sentence.

Can I run Molmo2-ER myself?

Yes, and the license is the easy part. Apache-2.0 permits commercial serving with no royalties. The practical path is the official repository, with vLLM support from version 0.15.0 and a Docker route. A 16 GB card class fits after converting the shipped 32 bit weights to 16 bit. Card hours on the 24 GB tier rented at 0.27 to 0.89 per hour on 26 October 2026. Your costs are card hours plus your engineering time. The site's measured card economics guides carry the fit tables.

The short version: the ER variant has no price anywhere else, the base-model figure in search is a stale record, and the one listed rate meters tokens with the live figure on the pricing page. Sign up to create a key and budget metered queries.

Last verified: 2026-10-26