# Dream-1 Robot API: Pricing, Compatible Robots and Research Beta Limits

Dream Labs offers a shared action model for two-arm robots, but integration and safety work still sit outside the advertised hourly price.

By Paras Madan | 2026-10-20


Dream Labs announced public API access to Dream-1 on October 19. The research-beta model predicts video and robot actions together, with Dream-1 priced at $3 per hour and Dream-1 Flash at $2. Its initial scope is bimanual robots with parallel grippers, rather than every kind of robot.

The [official announcement on X](https://x.com/dreamlabs_ai/status/2112230668805828801) says the API is live. The company's [technical release](https://dreamlabs.ai/research/dream-1/) explains how it transfers actions between different robot bodies, and where that approach still breaks down. This is hosted model access, not a release of open model weights.

## Dream-1 API pricing

Dream Labs says each inference call returns 1.6 seconds of actions. Its hourly equivalence is therefore 2,250 model calls.

| Model | Advertised hourly cost | Approximate cost per inference call |
| --- | --- | --- |
| Dream-1 | $3 | $0.00133 |
| Dream-1 Flash | $2 | $0.00089 |

These are charges for running the model. They do not include the robot, cameras, networking, supervision or the work needed to integrate and evaluate a particular setup. The hourly equivalence also should not be treated as a complete description of how idle time or an application's billing is handled.

For a robotics team, the more useful number will be cost per completed task. A task that needs several retries, manual resets or frequent intervention can cost substantially more than its model calls suggest. That does not make the API price irrelevant; it makes it one component of the calculation.

A [public Python SDK](https://pypi.org/project/dreamlabs/) provides the client interface for policy selection, observations and action requests. API and package availability establish a route to integration, but this article does not claim an account-level or physical-robot test.

## Which robots can Dream-1 control?

Dream-1 targets two-arm robots with parallel grippers. Dream Labs says it can generalize to many such robots without collecting embodiment-specific demonstrations first. That is the company's capability claim, with important conditions around the robot's geometry and controller.

The model uses three camera views: one from the head and one from each wrist. For supported native embodiments it can produce joint-angle commands. For other robots it can predict end-effector poses: where each gripper should be and whether it should be open or closed.

The robot's inverse-kinematics solver translates those desired poses into joint commands. This creates a shared interface across different arm designs, but the solver and the physical robot still determine whether the requested motion is feasible.

**A shared action interface reduces one integration barrier without eliminating robot-specific engineering.** Calibration, reachable space and motion constraints still matter when the same requested pose is sent to a different body.

## How the World Action Model works

Dream Labs calls Dream-1 a World Action Model because it learns to predict future video and actions together. The idea is to connect what a robot should do with how the scene is expected to change as it does it.

That differs from treating every new robot as a completely separate action system. Video provides a common observation format, while end-effector poses provide a common command format for the supported class of robots.

However, a pose specifies a destination more directly than the route to reach it. Dream Labs explicitly warns that end-effector prediction gives the model less control over the path around obstacles or near joint limits. Unfamiliar kinematics or dynamics can also lead to predicted states the robot cannot achieve.

These are reasons to evaluate the complete control stack. A convincing model output is insufficient if the downstream motion is unsuitable for the physical environment.

## What the external benchmark measures

The public [Robocurve evaluation](https://robocurve.org/) lists Dream-1 with a 59.3% mean task-progress score across 120 attempts per configuration. It estimates hosted inference cost at $0.15 per attempt.

Mean progress is not a full-task success rate. Partial completion contributes to that score, so it would be misleading to say Dream-1 successfully finishes 59.3% of tasks based on this number alone. The evaluation's attempt cost also belongs to its setup, rather than being a universal price for a robot job.

The external result is useful evidence that goes beyond the launch demos. It does not establish reliability across every compatible robot, task or environment, and it cannot justify unsupervised deployment near people.

## Research beta limits and deployment decisions

Dream Labs says Dream-1 will make mistakes and has not been validated for safety-critical use. It asks users to test in controlled settings and take extra care around people. That warning belongs in the decision to try the system, rather than in a footnote after the performance claims.

Fine-tuning can improve adaptation, but the company presents specific examples rather than a universal five-minute recipe. Some of its more different embodiments required much more data. Teams should not assume that dexterous hands or mobile manipulation inherit the same out-of-the-box support as the initial parallel-gripper scope.

The useful change is that a team can start evaluating a reusable robot policy through an API. Whether it becomes a viable product then depends on the remaining work: motion feasibility, controlled evaluation, intervention rates and consistent task completion. Lower model-access cost makes those experiments easier to start; it does not answer them in advance.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn technical launches, integration details and product limitations into clear stories that developers and founders can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).