Hard Tasks
GPT-5.6-sol-high
< 40% pass@4
Faster model
improvement.
Ready When You Are
Training data, runnable environments and rewards for faster model improvement
Build Your Next Frontier
Ready to Run
Runnable environment
and verified reward
Ready Now
18K tasks
off-the-shelf
Choose Your Climb
Single capabilities or pre-built packs
18K ready-made datasets
MCP and CUA
5,000each
70+ enterprise runtimes
- Marketing + retail
- Customer service
- IT service mgmt
- Human resources
- Finance
- ML operations
- Site reliability
- Personal assistant
CYBER
4,000
3,000
DEFENSIVE
AUDIT AND PATCH
1,000
OFFENSIVE
black-box exploitation
penetration-testing
DEEPSWE
3,000
Production repositories
Deterministic verification
Feature work, bug fixes, incident response, multi-PR changes, refactoring and integrations against full toolchains.
GDPVAL
1,000
Economically valuable work
- • Finance
- • Human resources
- • Marketing
- • Retail
- • Customer service
- • Personal assistant
How to start
Your capabilities.
Your volume.
Start with samples, evaluate the fit, then select the tasks you want for your training program.
- Choose your capabilities
Tell us what you want to improve. Select one capability or a mix across the catalog.
- Evaluate the samples
Review representative tasks, environments, and rewards against your own models.
- Set your volume and delivery
Agree on task counts and QA criteria. Delivery is now.
Can I buy a single capability?
Yes. Select one capability or a mix across the catalog. Choose the task families or domains and the volume you need.
What comes with the tasks?
Each package includes a task specification, runnable environment, and reward verifier. Depending on the capability, it also includes a reference solution or customer-specific QA. Inspect the complete package in your samples.
How do I evaluate fit before buying?
Start with five representative samples. Review the tasks, environments, and rewards against your own models, then select the tasks and volume for your training program.
Be a hill-climb hero.
Try samples against your own model
