OpenAI used the first week of June to ship two announcements that pull the assistant market in different directions at once. The company confirmed that ChatGPT has crossed one billion monthly active users, per fresh Sensor Tower data, hitting the mark roughly three years after launch and ahead of Google Maps, TikTok, Instagram and YouTube on time to first billion. It paired that distribution milestone with the general rollout of Dreaming, an always on memory layer that updates the user profile in the background without anyone pinning a fact by hand.
A long term profile that you do not own
The headline number gets the press release, but the memory upgrade is the part engineering teams need to read carefully. Until this week, ChatGPT memory was a short list of facts that users had to curate manually, and most people ignored it. Dreaming flips that pattern. The assistant now sorts through previous conversations and saves information in the background, reconciles contradictions and builds a richer profile of how each person works, who they work with and what they are trying to ship. Plus and Pro subscribers got the new behavior this week, with free accounts following in the coming weeks.
That shift in where context lives is the story for buyers. A year ago, swapping a model behind a thin retrieval layer was a one sprint job. With Dreaming style memory and hosted Codex artifacts, both the user state and the generated assets sit on the vendor's side, not yours. Migrating later means either rebuilding that context inside another provider or accepting that some customer history simply does not move with you. Contracts written this quarter need explicit data export, retention and portability clauses on memory objects, not just chat transcripts. Privacy teams should also be asking whether end users can inspect or delete the inferred profile, and whether deletion propagates back into any fine tuning corpora the vendor maintains. Most vendors are not in the habit of publishing that detail by default, and very few will offer it without a direct contractual push from a buyer with a meaningful annual commit.
Codex is OpenAI's enterprise wedge
OpenAI is no longer treating Codex as a coding sidecar. The company says Codex now has five million weekly active users and is positioned as a tool for any information and knowledge based work, not just programming. New plugins shipped for business and enterprise tiers this spring, including a preview that lets customers build interactive hosted websites and small applications straight from prompts, with the artifact refreshed as fresh data lands. The Windows version of Codex computer use shipped at the end of May, joining the existing macOS build, and the Codex mobile app is now live on iOS and Android. That is a much wider attack surface for a single procurement line, and finance teams should price it accordingly. The same key that runs a coding agent can now spin up customer facing front ends and drive a desktop session, which means a quiet expansion of who inside the organization can spend OpenAI dollars without an approval step.
Microsoft and OpenAI are now competitors
The commercial picture got more interesting on June 3. Microsoft AI chief Mustafa Suleyman told staff that the partnership with OpenAI is over in everything but name, framing it with the line that the team has to prove it can do everything from the ground up. He pointed at Microsoft's MAI Thinking 1 model and the new Scout assistant, both shown at Build 2026, as the start of a fully independent stack. OpenAI now competes with its largest distribution partner for the same enterprise seats while also fighting Anthropic, Google and Meta on the model side. The friendly co-sell motion that defined the 2024 cycle is gone, and renewal conversations this fall will reflect that fact directly.
How we are sizing the procurement risk
We are telling clients to do three things before signing or renewing an OpenAI or Azure OpenAI contract this quarter. First, put a metering layer in front of every key with hard ceilings per workflow, not per user, because the new website builder and computer use features can multiply token spend by ten inside a single session. Uber publicly capped employee AI spend in May after burning through its annual budget in four months, and we expect a similar correction at any organization that lets Codex run agentic workflows on an open tap. Second, write memory export and deletion clauses into the master agreement, with a 30 day SLA on portability and an audited list of what Dreaming has inferred about each account. That clause should include a clear definition of what counts as a memory object versus a chat log, because vendors will try to draw that line in their favor.
Third, dual track the Microsoft Scout pilot. The practical move is a 90 day side by side test once Scout reaches general availability, scoped to one internal workflow that already runs on Copilot. Two products on the same desk, the same prompts, the same evals, the same cost dashboard. Budget roughly twenty five thousand dollars for the bake off, including a small in house evaluator, and pick the winner on three metrics: cost per successful task, latency at the 95th percentile, and number of escalations to a human reviewer. That is enough signal to brief a steering committee and to negotiate a real discount before the contract quietly auto renews at full list.
Two dates that decide the 2027 roadmap
Two forcing functions are close enough to plan against this summer. Microsoft has signalled that Scout reaches general availability inside the Microsoft 365 commercial channel later in the season, which is the moment to lock in per seat pricing before the published list rate becomes the negotiation floor. OpenAI's next DevDay, expected in October, is the obvious window for a Dreaming API and for a clear statement on whether memory objects can be exported under the enterprise tier. If neither lands by the end of September, assume the platform stickiness is intentional and route the 2027 migration budget into the plan now, while renewal talks still give buyers room to push back rather than after the next price sheet arrives.
