Five shifts OpenAI just telegraphed

Five shifts worth probing

Tibo, who leads product for ChatGPT and Codex at OpenAI, sat down with podcaster Matthew Berman on 24 August and laid out where agentic AI goes next. The 44-minute interview, published on Berman’s Forward Future channel and summarised by BidClub, carries five signals worth taking seriously.

The growth claim first. Codex hit 20 million users on a curve Berman describes as turning vertical once it was folded into ChatGPT, putting a coding agent in front of an existing user base of product managers, designers and sales teams. Asked why, Tibo pointed to distribution rather than rivalry with Anthropic. I don’t tend to look at the competition that much, he said.

Laptops hit their limit. Tibo’s tweet that Codex will seem primitive in two to three months carried a reasoning worth taking seriously. Today’s laptops were designed around human throughput — typing speed, attention, the handful of apps a person can have open. A model does not share those limits and may eventually handle 100 applications opened at the same time perfectly fine. The next generation of agents will need more than the resources on your desk.

Ultra-fast reshapes the work. Berman set up the third signal: today’s speeds push him to kick off ten to fifteen parallel agents, which is pretty significant cognitive overhead. Ultra-fast, running ten to fourteen times faster, might let him run three or four instead, or stay in flow on one. Tibo agreed, noting that the speedup feels strongest when the agent is generating code or copy and weakest when it is making many tool calls — network and orchestration overhead become the new bottleneck.

Voice is now production-ready. The new voice supports tool use, Tibo said, and it has changed his own morning. In the morning, I just sit there with my phone and dictate a couple of things for ChatGPT, and then it just goes and does it — it has access to all my tools. Every time the interface leans into something more natural, humans choose the path of least resistance.

20MCodex users reached, on a growth curve the host describes as turning from flat to vertical after the merge into ChatGPT.

One interface, everyone. In Tibo’s framing, you and your mom will use the same thing. Labels like software engineer or designer are, in his words, just human concepts we have invented to deal with abstractions. The interface adapts to you; you do not adapt to it.

Efficiency compounds month on month. Outside ultra-fast, speeds are about 60% faster than three months ago. Luna, optimised by Soul, dropped 80% in price; Tibo said Luna, which sat at the frontier six months ago, is now cheap enough that some uses are effectively free. He expects the next model to be more efficient again, and committed to passing the gains on rather than just pocket that interesting gain.

The bigger picture is where all five signals point. The next wave is not a single new model — it is the seam between model, voice, agent and interface dissolving. Agents run faster, speak instead of type, and stop asking which interface you are.

How to try it this afternoon

Five experiments, twenty minutes total:

• Voice plus tool use. Dictate one task with a connected tool — calendar, inbox, file — and notice whether the cognitive mode feels different from typing. (Our piece on Google’s free iPhone dictation app covers an offline alternative if data residency matters to you.)
• High-speed tier. Switch on whatever ultra or pro mode your provider offers and run a generation-heavy task and a tool-heavy task back to back. Time both — the gap tells you which workloads benefit.
• One-interface test. Open your general assistant and your coding assistant in the same window. Ask each for the other’s job and notice where the seam is.
• Pricing audit. Compare what you spent on AI last quarter with what the same workloads would cost on today’s cheaper frontier models. If your budget has not been revisited in six months, you are probably overpaying — see our guide to usage-based pricing.
• Workflow reshape. Run one task the old way (parallel, multi-agent) and one the new way (one in flow, high speed). The one you prefer is the workflow to invest in next quarter.

For a UK team, the practical question is not which tool to pick — it is whether your current way of working is set up to ride the wave, or to be left scrambling when it lands. Probe these five patterns this afternoon, before they become table stakes.