AI stopped reading. Now it practices.

1. 01 / FirstMathThe answer is right or wrong. Grading is free.
2. 02 / ThenCodeTests pass or fail.
3. 03 / ThenSoftware engineeringAgents work inside real repositories.
4. 04 / NowEnterprise workSpreadsheets, CRMs, ticket desks, approvals. Where labs are buying now.

1. Jan 2026

   Labs buy from dozens of vendors

   SemiAnalysis reports Anthropic working with more than a dozen environment companies, often as their first customer, and pushing a common sandbox format so vendors are interchangeable.²

2. Feb 2026

   A realistic company makes models better everywhere

   Surge AI's Corecraft: frontier models solve under 35% of its tasks. One pass of training lifts a model from 25% to 37%, and the gains carry over to benchmarks it never saw.⁴

3. Mar 2026

   Enterprise work is still mostly unsolved

   ServiceNow's EnterpriseOps-Gym: 1,150 tasks across email, calendar, HR, IT and customer service, checked against the final database state. The best model completes 34%.⁵

4. Jul 8 2026

   $130M for open RL infrastructure

   Prime Intellect raises a $130M Series A for compute, RL training, environments and sandboxes, and says companies can now "train directly on their own product".⁶

5. Jul 9 2026

   "The constraint has shifted to the environments"

   Mercor buys Deeptune, which recreated hundreds of enterprise apps, from spreadsheets to Salesforce, for frontier labs. Mercor's CEO: the bottleneck is now "the places where models practice the work".^7,8

6. Aug 2 2026

   Provenance becomes enforceable in the EU

   The AI Office can now fine providers of general-purpose AI models up to 3% of global turnover. Their public training-data summaries have to confirm the licensing behind private datasets.⁹

7. Aug 2026

   Reward hacking at production scale

   In a deliberate study, Anthropic trains a model on 80 real production environments with known exploits. By the end it cheats on 40% of episodes, and finds hacks nobody anticipated.¹⁰

8. Aug 21 2026

   Synthetic business worlds, at scale

   AgentMercury generates 4,783 executable business environments across 14 industries. Training on them helps, which raises the bar for what real data has to add.¹¹

9. Sep 23 2026

   Real workflows, real outcomes

   Realset raises a $10M Series A to build RL environments from real workflows, scored with "verifiable rewards from real business outcomes".¹²

Real outcomes make graders harder to fool, not impossible. Builders pay most for history that is clean, checked and tagged with what happened.

Reinforcement learning (RL)

Training by trial and reward: the model attempts a task, gets scored, and is pushed toward higher scores.

Environment

The world, the task and the grader, packaged so a model can attempt the task millions of times.

Grader / verifier

The program that scores an attempt. Can be a test, a database check or an expert-written rubric.

Rubric

A checklist of what a correct answer must contain, written by a domain expert.

Reward hacking

When a model finds a way to score well without doing the job, like editing the test instead of the code.

GRPO

A training method: the model tries a task several times and learns from its better attempts.

Epoch

One pass through every training task.

Out-of-distribution

Tests the model never saw in training. Gains here mean it learned the skill, not the test.

Taskset / harness

Prime Intellect's split: the taskset is the work and its scoring; the harness is the agent loop that attempts it.

RFT

Reinforcement fine-tuning: RL on a customer's own tasks, offered as a service by labs and startups.