Breaking

OpenAI Says Internal Model Solved 100+ Math Problems, Intensifying RSI Debate

September 23, 2026 at 01:01 EDT

Foundation Models
AI Agents
Research & Papers

OpenAI and Anthropic have put new numbers behind the once-speculative idea of recursive self-improvement, the feedback loop in which AI systems help build better AI systems. The disclosures do not establish that full RSI or artificial superintelligence has arrived, but they have compressed the public debate from decades to years.

September 2026 · Automated AI research

The RSI debate moves from theory to measured lab work

Frontier labs now report internal models solving major math problems and AI agents doing measurable shares of AI R&D. Full recursive self-improvement has not been shown, but the loop is getting easier to measure.

100+

long-standing math problems claimed resolved

10,000

concurrent agents on the Navier-Stokes run

88h

from launch to agent-produced resolution

reported code output per engineer per day

Agent labor is becoming a visible input

Human workday

1.0

Agent-workdays

3.1

One reported research organization now uses about three agent-workdays for each human workday.

Code output changed by a full order of magnitude

2024 baseline

Q2 2026 output

The proposed loop

Humans set goals

→

Agents run work

→

Results are checked

→

Next model improves

The unresolved question is whether the goal-setting and verification steps can also become reliable enough to close the loop.

Why timelines feel shorter

Large agent swarms, cheaper GPT-6 tiers, and rapid gains in internal research automation all point toward faster capability cycles.

What still blocks full RSI

Research taste, proof review, compute, long-horizon safety evaluation, and alignment remain active bottlenecks.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.