The development

What happened

OpenAI's 18 November report says coding agents are completing more work inside its AI research organization, but performance varies sharply by task and humans remain central. Its estimates for tasks requiring two to four human hours rose from 23% success without intervention in March to 46% in October; four-to-eight-hour tasks rose from 13% to 35%. These are OpenAI's internal measurements, not an independent productivity trial. The company cautions that researchers may have changed which tasks they delegate as agents improved.

Key facts

  • For 1 August–18 October, OpenAI estimates success without human intervention at 79% for monitoring training runs, versus 25% for writing inference and serving code and 29% for building evaluation, reinforcement-learning and agent environments.

  • October median daily agent use was valued at $2,400 at API list prices, with $23,000 at the 90th percentile. That is not cash spending: activity, model prices, behavior and harness changes all affect the measure.

  • A late-August/early-September survey of 101 staff asked what they would trade to keep frontier coding agents for three months. Separate median answers were 40% of team headcount or 30% of compute: hypothetical preferences, not cuts or a representative causal study.

  • Ten researcher interviews described faster implementation, debugging and data review, alongside poor research taste, low-quality code or data, and invalid analysis assumptions. Researchers reported needing close oversight to avoid wasted experiments.

  • Logged experiments per active experimenter averaged 1.68 times the 2025 baseline in the four weeks ending 18 October. OpenAI says more experiments or changed lines of code need not mean faster research progress; its internal capability index also uses a different benchmark mix.

  • Interviews with key contributors attributed the named posttraining and safety-monitoring ideas for GPT-6 Astra and GPT-6.2 Astra to people. Some GPT-6.2 pretraining ideas developed through exchanges with Codex, not fully autonomous model invention.

The context

Why it matters

TideReach assessment: the useful signal is where AI assistance changes research work, not a single headline multiplier. Stronger monitoring and implementation can accelerate parts of a pipeline while weak experiment design or evaluation remains a bottleneck. Task-selection changes, self-reported value and vendor-defined capability metrics prevent these results from establishing an overall causal productivity gain or completed recursive self-improvement. This is a research update, not a new public model release.