Anthropic could reach human-level AI research before OpenAI, says Vals AI CEO
Key points
Key takeaways from a The Information - TITV interview with Vals AI Co-Founder & CEO Rayan Krishnan (September ̃2026):
Frontier scorecard. Vals AI's benchmark ranks Anthropic's Sonnet 5.5 second overall, behind its own flagship Opus 5.5 but ahead of OpenAI's Astra at fourth, a strong result fora model positioned as a cheaper, mid-tier option.
Cost nuance. Sonnet 5.5 is cheaper per token but token-hungry, so in expectation it costs roughly twice as much as its predecessor Sonnet 5 for the same task.
No test leakage. Vals AI builds its own benchmarks across finance, law, healthcare and coding, scoring real professional tasks rather than quiz-style questions, keeps test sets private with zero data retention,and refuses to sell training data to labs, so scores cannot be gamed.
OpenAI's pause. Krishnan calls OpenAI's decision to hold back its latest Astra modela natural conclusion after earlier safety incidents. OpenAI has outlined offline alignment evaluations, audits of RL environments, controllability work and external auditors, but he says a well-defined threshold for when scaling can resume is still missing.
RSI timeline. Vals AI's recursive self-improvement research measures models' ability to build their own successors via five proxy tasks against human baselines, projecting Anthropic to reach human-level researchers by August ̃2027 if recent progress continues, ahead of OpenAI around the end of 2027.
After the threshold. Once models can do much of the work of building their successors, model development becomes hard to predict and requires fully automated evaluations. Krishnan expects chip and energy constraints to blunt any fast-takeoff scenario, urging more investment in understanding models before release.
Government role. Krishnan welcomes more public involvement in AI evaluation, including US agencies and the UK's AI Safety Institute,and anticipates evaluators facing accreditation or licensing as AI-safety legislation advances.
Read more: The Information
