Loading...

OpenAI's own researchers are getting seriously hooked on their internal AI coding agents, valuing these tools over a significant chunk of their human colleagues or compute power. This deep dive into their internal operations offers a fascinating, if sometimes messy, glimpse into the future of AI-assisted development.
The Decoder, citing OpenAI's own internal report, reveals a massive surge in the use of internal tools like Codex and Dots. The theoretical token costs alone are eye-watering: a median researcher would have consumed $2,400 worth of tokens per day in October, a 62% jump from the previous month. For the top 10% of users, that figure doubled to a staggering $23,000 daily. These agents are effectively running 5.7 workdays per researcher per calendar day, and even more in pre- and post-training teams, with over two-thirds of tracked researchers running at least four agents concurrently at peak usage.
OpenAI calculated that the theoretical API equivalent of this agent runtime was twice as high as researchers' regular salary costs in September, though it's worth noting this excludes equity packages, which make up a huge chunk of compensation at OpenAI, according to The Decoder. Despite this intense usage, internal agent activity only accounted for 2% of the company's total research compute.
The success rates for autonomous agents working without human intervention on two- to four-hour tasks increased from 23% in March to 46% in October, while four- to eight-hour tasks saw a climb from 13% to 35%, The Decoder reports, citing OpenAI. However, an Epoch AI taxonomy breakdown shows these systems primarily handle simpler tasks. For instance, they monitor training runs 79% of the time and answer technical questions independently 72% of the time. But when it comes to more complex engineering – like building evaluation and reinforcement-learning environments or writing code for inference and serving systems – success rates drop significantly to 29% and 25% respectively.
In an internal, non-representative survey of 101 employees, researchers indicated they'd rather give up 40% of their team's staff or 30% of their compute capacity than go three months without these coding agents, according to The Decoder. A separate quick poll of 28 employees found 27 preferred the latest internal model over a mid-career human engineer.
Qualitative interviews with ten researchers, however, highlighted the downsides. These models often lack "taste" for good research ideas, code, and data quality. They get sidetracked, make faulty assumptions, and produce "AI slop" – subpar code changes and unusable data artifacts. Close human oversight is crucial to avoid wasting compute cycles, The Decoder states. Interestingly, the key conceptual ideas for OpenAI's upcoming GPT-6 Astra and GPT-6.2 Astra came exclusively from humans, with only individual researchers gaining important insights through dialogue with Codex in pretraining for GPT-6.2 Astra.
Alright, small business owners and team leads, this isn't just navel-gazing from a tech giant. This is a peek behind the curtain at where AI development is headed, and what you can expect from future AI tools for your own teams. The headline fact — that OpenAI's own researchers value these AI agents more than nearly half their human colleagues — should grab your attention. It shows that even with all the "AI slop" and wasted compute, the productivity gains are that significant.
What this tells me is that AI coding assistants, and by extension, AI agents for other knowledge work, are becoming indispensable, even if they're still messy teenagers. You're not going to get a fully autonomous AI engineer anytime soon, but you are going to get tools that dramatically amplify your existing team. The "AI slop" issue is real, and it reinforces what we already know: AI tools are fantastic for routine, well-defined tasks (like monitoring or answering technical questions), but they utterly fall apart on complex, nuanced, or creative problems. This means your human experts aren't going anywhere; they'll just be spending less time on grunt work and more time on the strategic, "taste"-driven decisions that AI can't make. For small teams, this is crucial: it means you can stretch your talent further, but you absolutely must have human oversight to catch the inevitable AI screw-ups and guide the tools effectively. Don't expect magic; expect powerful, but often clumsy, co-pilots.
Open-source Hermes Agent is coming to Microsoft Teams, allowing workers to package successful AI workflows into team-wide reusable skills.

Reson8 open-sourced Rumble-2-Flash, a 0.5B parameter ASR model for 40+ languages that outperforms models up to 60x larger, with Apache 2.0 licensing.

Instagram introduces a tool to detect unauthorized video reposts and AI-modified content, allowing creators to block or label infringing posts.

Google Maps’ new Ask Maps AI can auto‑create day‑by‑day itineraries, share them, and even call businesses for appointments—useful for small teams on the move.

Comments