Gemini Robotics API pricing: the rate card, line by line

Google publishes exactly two robotics rows on its Gemini API pricing page. They are Gemini Robotics ER 2 Preview and ER 2 Streaming Preview, and this is every line of both, read on October 26 2026. Input costs 1.00 US dollars per million tokens through December 31 2026. It rises to 2.00 on January 1 2027. Output, with thinking tokens, costs 5.00 per million tokens through the end of 2026, and 10.00 after. Batch, caching, and grounding carry the same doubling date. The guide also covers the request shapes, the token limits, our robot-day estimate, and the other public robot rate cards of late October 2026.

Last verified: 2026-10-26

What the card prices

Gemini Robotics ER 2 is Google's embodied reasoning model for robots. Google's pricing page describes it as a vision-language model endpoint that enables robots to understand their environments precisely, supporting agentic orchestration of robots, video progress understanding, multi-robot collaboration, and advanced spatial reasoning. The announcement post calls it a high-level brain for robots. It plans, watches video, tracks whether a task is going well, and picks the next step.

The two rows on the card are one model in two shapes. The standard row takes a request and returns a response. Streaming, the second shape, runs on the Live API. It takes text, image, video, and audio, and it supports two-way streaming with function calling. That is the shape a live robot loop needs.

What the card does not price matters just as much. ER 2 is not a motor control model. Google's announcement says it hands off motor execution to a lower level vision-language-action model, a VLA. The VLA tier from Google has no public price and no row on the card. If your question is what it costs to move an arm, this card does not answer it. The action model is a separate item. You host it yourself, or you buy it from someone else.

The lineage explains the shape. ER 1.5 arrived on September 25 2025. It was the first model in the family open to all developers. ER 1.6 followed on April 14 2026. Google retired it on August 31 2026. ER 2 shipped on July 30 2026 and is the current row. Moving between them was a model string swap, not a rewrite. On quality, Google's model card charts ER 2 at 78.5 percent on its ERQA spatial reasoning measure. ER 1.6 scored 72.5. The model sits in public preview on Google AI Studio and the Gemini API. It sits in private preview on Google's enterprise agent platform, and that platform's public pricing page has no robotics rows at all as of October 26 2026. The public rate card lives in one place: the Gemini API pricing page.

The rate card, line by line

The paid column is headed, in full, Paid Tier, per 1M tokens in US dollars. Every row prints two dated prices. Here is the standard tier of Gemini Robotics ER 2 Preview, complete.

The batch tier exists for ER 2 only. It costs half the standard columns, with the same doubling date.

ER 2 Streaming Preview carries the standard input and output columns at the same numbers. It has nothing else: no caching row, no batch tier. Three more lines complete the register. Grounding with Google Search costs 14 US dollars per 1,000 requests on the paid tier, after 5,000 free requests per month. The free quota is shared across all Gemini 3 and newer models. The free tier row says input and output are free of charge on both models, with grounding marked not available. Then comes the data row. Free tier inputs are used to improve Google's products. Paid tier inputs are not. A robot streaming video of a customer site should read that line before any rate.

Read the batch row with its shape in mind. Batch is not real time. You submit work and collect results later. That fits offline video review, dataset labeling, and overnight reports, and it halves both columns. It does not fit a control loop that needs an answer in the next fraction of a second. The streaming row is the real-time one, and it has no discount tier. Caching helps when the same context repeats: a fixed prompt, a map, a long instruction set. The cache read rate is a tenth of fresh input through 2026. But cached storage bills by the hour, so a cache held over an idle weekend still costs money.

The January 1 2027 doubling

Every paid column on the card doubles on January 1 2027. Input goes from 1.00 to 2.00 per million tokens. Output goes from 5.00 to 10.00. Batch goes from 0.50 and 2.50 to 1.00 and 5.00. Caching goes from 0.10 to 0.20. Cached storage goes from 0.50 to 1.00 per million tokens per hour. The grounding overage holds at 14 per 1,000 requests. Free tier rows stay free. And the doubling is printed in each cell, not buried in a terms page. Budget any late-2026 robot project against the doubled columns. The date sits inside most deployment lifetimes.

The doubling is also the most misquoted fact on this query. In the October 26 2026 capture, Google's own AI Overview quotes input with its through-2026 tier. Then it quotes output at 10.00 with no tier at all. That states the 2027 column as if it were today's. Today's output price is 5.00, so the flat 10.00 doubles it. One aggregator page presents 2.00 and 10.00 as the standard rates. Two more list the 2026 columns with no dates, which reads as a permanent price. All four errors flatten a two-column card into one number. The defense is mechanical. Read the cell on Google's pricing page itself, both halves of it. Carry the date next to every number you write down.

Request shapes and limits

The robotics models are called through Google's Interactions API. Its endpoint is a POST to the /v1beta/interactions path on Google's generativelanguage API host. The two model strings are gemini-robotics-er-2-preview and gemini-robotics-er-2-streaming-preview. A request carries a model, a list of input parts, and a generation config. Images go in as base64 inline data or as a Files API reference. Cloud storage URIs in the gs:// form are not read directly. The generation config takes a thinking level, and the docs example sets it to high. That matters for cost: thinking tokens bill at the output rate. Pointing queries return JSON. Points come back in y then x order, scaled 0 to 1000, with labels. Function calling is supported, and it is the point of the model. You declare the robot's own interfaces as tools: VLA models, navigation, arm control. The model then drives them. Google's demo has it running a Boston Dynamics Spot through the Spot APIs.

Two limits are printed and one is not. The token limits are 131,072 in and 65,536 out. The access limit: a plain, unrestricted API key gets a 403 Forbidden on these models. Set key limits before the first call, not after. The missing one is a per-model rate limit. Google's rate limits page prints no requests-per-minute or per-day numbers for the robotics models. It says limits are per project and can be viewed in AI Studio. The per-day counter resets at midnight Pacific time. And it warns that specified rate limits are not guaranteed and actual capacity may vary. For a multi-robot plan, treat the AI Studio dashboard as the live source. The request contract details are in Google's robotics documentation.

What a robot-day costs

The card meters tokens. A robot consumes frames, audio, and instructions, and it produces reasoning. Most budget errors start in that gap. The arithmetic is simple once the inputs are named.

tokens per day   T  =  3600 x H x f x k  +  O

H: hours the robot runs per day
f: camera frames sent per second
k: tokens per frame
O: output and thinking tokens per day

cost per day     C  =  T / 1,000,000 x P

P: the blended per-million rate from the card, both date tiers

Our working estimate uses an image-heavy stream at 1 to 2 frames per second and about 258 tokens per frame, plus outputs. Those are the assumptions from our October 2026 research record. That lands at roughly 10 to 30 million tokens per robot-day. At the through-2026 columns, a robot-day of streamed reasoning costs in the band of 10 to 50 US dollars. This is our estimate, not Google's number. Google publishes rates, not robot-day figures. One sensitivity line: long sessions at high thinking level push past the top of the band, because output bills at five times the input rate through 2026. And the whole band doubles on January 1 2027 with the card. For more on robot pricing units, read the pricing units guide and the per-robot-hour glossary.

The other public rate cards, October 2026

An honest reading in late October 2026: Google's is the only public rate card for a robot model from a major first-party vendor. It is no longer the only public robot rate card. A specialist host, abstract.inc, publishes one too. A second platform looks exactly like a rate-card seller and publishes none. Here is the state, checked October 26 2026.

abstract.inc serves a menu of robot models behind one chat-style endpoint. It prints a rate table on its pricing page. The units are the interesting part. Input and cached input are metered per 1M tokens. Output is metered per 1M actions. The hour column states its own basis: one hour of a 30 Hz control loop, 108,000 actions. The rows, in US dollars.

Two honesty notes on that table. The pi0.5 row is a tune trained on recordings published by the Allen Institute for AI. The host itself notes it is not a Physical Intelligence release. And the action-metered output column is a different unit than Google's. One million actions at a 30 Hz loop is about 9.26 hours of motion. The two cards do not compare row to row without that conversion. The per-action split is worked through in the pricing units guide.

roboticinference.com is the platform-shaped entrant. One API covers 24 models from 13 families. The families come from NVIDIA, Physical Intelligence, Hugging Face, and others, and the site is independent of all of them. It runs a LIBERO leaderboard measured in-house. Several base models can be tuned side by side and then tested in simulation. Each model carries a machine-readable contract: state layout, units, coordinate frames, control mode, gripper convention, zero-action safety, control frequency, chunk size. That is the contracts idea taken seriously. No rates appear anywhere on its site as of October 26 2026. Access is by registration or a pilot.

The first parties around Google stay unpriced. Physical Intelligence runs its API beta as a waitlist, with batch onboarding. The application covers team, robots, and teleop hours. There are no public rates. The waitlist mechanics have their own guide. AY-Robots publishes prices but not a rate card. Its tiers cost 49.99 and 299.99 euro per month. The full fee returns as training credit for operator hours and GPU time. Inference is bundled into plans, not metered per call. So the October 2026 state, dated and specific: two public rate cards exist. One is from a major first-party vendor and meters tokens. One is from a specialist host and meters tokens in, actions out. Every other name is unpriced, waitlisted, or wrapped in a subscription. The embodied reasoning term guide covers how this model class got its name.

Reading this card next to this platform

This platform meters its chat route per million tokens, the same unit family as Google's card. The budget math above translates directly. The platform keeps a written capability record for its own robot reasoning model, molmo2-er. Its model page carries the use cases and call shapes. This guide quotes no rate of ours. Live numbers render on the pricing page, the single source for them. The molmo2-er cost story has its own guide. Spend runs against a prepaid wallet, and the wallet balance is the hard stop. That is the discipline the doubling date argues for: a cap that holds in January as well as October. The API bill shock guide documents why that matters for unattended robot loops. The measured robot-hour cost report puts a number on the self-hosted path. When you are ready to budget metered calls, you can sign up and create a key without talking to anyone.

Frequently asked questions

Is there a free tier for Gemini Robotics?

Yes. Both rows list input and output as free of charge on the standard tier. Two caveats sit on the card itself. Grounding with Google Search is not available on the free tier. And free-tier inputs are used to improve Google's products, while paid-tier inputs are not. A robot streaming customer sites should read the second caveat first.

What exactly changes on January 1 2027?

Every paid column doubles. Input goes from 1.00 to 2.00 per million tokens, and output from 5.00 to 10.00. Batch goes from 0.50 and 2.50 to 1.00 and 5.00. Caching goes from 0.10 to 0.20, and cached storage from 0.50 to 1.00 per million tokens per hour. The grounding overage and the free tier do not change. Carry the doubled columns forward for anything still running next year.

What is the lowest-cost way to run ER 2?

Batch, where the workload allows it: offline video review, labeling, overnight analysis, at half the columns. Caching, where context repeats, at a tenth of fresh input through 2026. Neither fits a live control loop. The streaming endpoint has no discount tier. A real-time robot pays the standard columns for the life of the deployment, doubled from 2027.

Is Gemini Robotics ER 2 a vision-language-action model?

No. It is an embodied reasoning model. It perceives, plans, tracks progress, and calls tools. Motor execution goes to a separate, lower-level VLA that you supply. Google's own VLA tier has no public price. The action side of the stack is covered in the VLA models explained guide.

Who else publishes robot model rates?

As of October 26 2026, abstract.inc is the other public rate card: six models, token-metered input, action-metered output, hour estimates at a 30 Hz loop. roboticinference.com runs a 24-model platform with no public rates. Physical Intelligence is waitlisted and unpriced. AY-Robots sells subscriptions that return as credit. Among major first-party vendors, Google is the only one publishing.

The short version: two rows, two date tiers, one doubling date. Input 1.00 and output 5.00 per million tokens through December 31 2026. Then 2.00 and 10.00 from January 1 2027. Batch sits at half. Streaming has no discounts. The free tier's data-use row deserves a read. Live rates for this platform are on the pricing page, and you can sign up to budget metered calls.

Last verified: 2026-10-26