Abstract is building the inference platform for robotics.
Introducing AbstractRead the blog↗
Abstract is building the inference platform for robotics.
Available through our API and harness.
Build with Abstract A rack on your floor, or Thor
Copy setup promptSetup prompt copied
Open the setup prompt in an assistant
- Open in
- Claude
- ChatGPT
- Cursor
- Reserved before it runs
- Settled to measured usage
- Keys stored as digests
Works with
The path to useful robots runs through inference. Open-source and custom robot foundation models are improving fast, but serving them in production still means GPUs, queues, retries, and bills nobody can predict.
Serving is the constraint.
Real robots need low latency, high reliability, and a budget they cannot overrun. Training more capable models will not fix the path between a policy and the hardware it drives.
Abstract serves and scales those models behind one OpenAI-compatible API, with hard spend limits enforced before the first token.
Every lane is priced at provider cost, with no markup. Agents connect through our MCP server, and robots stream actions with the same keys, in our cloud or yours.
abstract
Camera Joints Task Observe Policy Action chunk 50 Hz Pack a boxFold this towelPlug this cableLace this harness Closed loop, every chunkClosed loop Stream via API Your robot or
Figure 1. Camera frames, joint state and a task enter Abstract. A policy model infers an action chunk, which streams back to your robot through the API, or to your agent through MCP.
What we do
Robot models, served.Open-source and custom robot foundation models behind one API, with the low latency and high reliability real robots demand.
In our cloud or yours.Run on our GPUs, or self-host the same stack in your own cloud when the data cannot leave it.
Hard spend limits.Every request reserves its maximum cost before it runs and settles to measured usage, so no agent outspends the ceiling you set.
Pricing
Our promise is inference robots can trust: fast, reliable, and never above the limit you set.
Token pricesUSD per million tokens, async lanes
| Input and output prices in US dollars per million tokens | Model | Input | Output | | -------------------------------------------------------- | ------- | ------- | ------ | | GLM-5.3 | $_0.40 | $_1.80 | | | DeepSeek V4 | $_0.09 | $_0.18 | |
Control rateactions per second, per robot 50Hz
Markupon provider token costs 0%
Versus real-time frontier lanesestimated list prices, same request
An example agent request costs 8.3× less on glm-5.3:async than on a real-time frontier lane at an estimated $3 / $15 per million tokens.
Output priceUSD per million tokens, by lane $0.18
Frontier, estimatedAbstract lanes
Cost per request
Example request$0.00148.3× cheaper, estimated
Cached input$0.0880% off input
Real-time frontier (estimate)Abstract
GLM-5.3 DeepSeek V4 Flash Kimi K3
$0.00$0.50$1.00$1.501003001k3k10k30k100k Output tokens (log scale) The price list
“Every request reserves its maximum cost against your hard monthly limit before it runs and settles to measured usage afterwards.”
Abstract terms · Spend limits and billing
Abstract runs in our cloud or yours,
wherever your robots are.We run the inference stack ourselves, which means we can deploy it wherever your data and your robots live.
Talk to us about self-hosting 1. ### In our cloud 001 2. ### In your VPC 002 3. ### On your own GPUs 003
Hard spend limits.
Every request reserves its maximum cost against your monthly limit before it runs, then settles to measured usage. API keys are shown once and stored only as digests.
Writing
1. 01Inference engineering harness 2. 02Data annotation harness 3. 03Robot control harness Read the blog
Careers
We’re a small team building towards abundant labor, backed by HF0, Northside and Susa Ventures, with angels from Anthropic, OpenAI and Dyna Robotics, and part of NVIDIA Inception. If that is what you want to spend the next few years on, we would like to hear from you.