Loading page…
Company intelligence / Live profile
Track AI21 Labs's role in the AI economy through Superpower Daily reporting on models, tools, and business. This live company page updates automatically as new launches, partnerships, product moves, and market signals are published.
Latest coverage
Key developments
Its LIFT-Bench research scores content alongside formatting. Training with that score avoided shrinking answers, but outside tests showed avoided regressions—not a general breakthrough.
The strongest quantified cost claim comes from Deel’s payroll evaluation: 72% accuracy at one-thirty-third the per-query cost of Claude Opus 5.5. The results are specific to partner tests.
A study of internal developer traffic found that only about 11% of messages asked for code. The company's routing comparison tests budget calibration, not proof of equal answer quality.
Company-reported experiments favor checking candidates for factual search, merging reports for deep research, and dividing coding work by what can be verified.
Kueue replaced requests in team chat, but AI21 still needed changes to share capacity fairly and ensure that free GPUs could actually fit incoming jobs.
AI21’s company-reported tests suggest an independently trained checking model can recover answers an ensemble already found but failed to choose. The results are promising, but they rest on 100-question samples, automated grading, and a verifier trained partly through closed-model distillation.
Your company watchlist