Labelbox
Labelbox builds AI training environments for frontier labs, background autonomous agents, and a robotics data engine
About Labelbox
Labelbox builds environments for training AI for frontier labs and lets engineering and research teams hand backlogs of tasks to agents. Recursion's agents start on their own: triggered by a red build, an alert, a new checkpoint, or a scheduled task, they keep working in the background until the task is done — fixing flaky tests, keeping dependencies updated, re-running evaluations, and comparing dozens of experiments in batches. Terra is Labelbox's robotics data engine, collecting egocentric and teleoperated demonstration data, annotated and validated by the AI diversity engine, delivered from pretraining through evaluation. Horizon targets reasoning, tool use, and computer use, providing reinforcement learning training grounds and evaluations. See the Recursion pricing below for billing terms.
Best for
Best use cases
Run background agent jobs triggered by a red build, an alert, a new checkpoint, or a scheduled task
Fix flaky tests, keep dependencies updated, re-run evaluations, and compare experiments in batches
Collect egocentric and teleoperated robotic demonstration data and complete annotation and validation
Build reinforcement learning training grounds and evaluations for reasoning, tool use, and computer use
Compare before choosing
The official comparison page states (the models and machines comparison table on the Recursion pricing page): claude-haiku-5-5 belongs to Anthropic, starting at $0.15 per session hour, typically $0.36, up to $0.72 during busy hours.
Features
Terra Robotics Data Engine
Collects egocentric and teleoperated demonstration data, annotated and validated by the AI diversity engine, delivered from pretraining through evaluation.
Horizon Reinforcement Learning Training Grounds and Evaluations
RL training grounds and evaluations for areas of high AI economic value such as reasoning, tool use, and computer use.
Recursion Background Autonomous Agent
An agent is triggered by a red build, an alert, a new checkpoint, or a scheduled task, starts on its own, and works in the background until the task is complete.
Pre-Launch Cost Estimate
Recursion shows the estimated per-session-hour price for each agent before launch, so you can decide the scope of the run.
Pros
- Recursion shows the estimated per-session-hour price for each agent before launch, so you can decide which tasks to run beforehand
- Recursion storage is not charged, and transcripts and session files incur no storage fees
- Saving agents and environments in Recursion consumes no compute and still works when the balance is empty
- Terra annotates and validates robotic demonstration data with the AI diversity engine, delivering from pretraining through evaluation
Cons
- Recursion rate limit: 300 requests per 60 seconds per quota, all API keys share the same quota, console usage is counted separately, and exceeding the limit returns 429 rate_limit_exceeded with Retry-After
- Recursion transcripts and session files are not kept permanently and are cleaned up over time; download anything you need to retain long term
Pricing Plans
claude-haiku-5-5 (Anthropic)
- Model entry claude-haiku-5-5 from Anthropic in the official comparison table (Recursion pricing page)
- Billed per session hour, one price covers the sandbox and its model costs
- The estimated per-session-hour price for each agent is visible before launch
* Pricing information is for reference only. Please refer to the latest announcement on Labelbox's official website.