OpenAI quietly admits the model menu was too long

OpenAI shipped three things on 28 May: a tuning update to GPT-5.5 Instant, a workspace agent control panel for ChatGPT Enterprise, and app templates for Snowflake, Databricks, and GitHub Enterprise. It also announced that o3 and GPT-4.5 will be retired from ChatGPT, folded into the 5.x family. Anthropic shipped enterprise connector permissions the same day. The interesting story is not any one of these; it is what they look like stacked.

What was actually released. The GPT-5.5 Instant update is a tuning pass, not an architecture change. The release notes describe "more natural responses," better in-chat code and writing block formatting, and, the phrase worth circling, "fewer bullet-heavy outputs."¹ The workspace controls cover model selection per workspace, admin permissions, app access, and response capability configuration. The app templates are pre-built agent configurations for the three platforms where enterprise data actually lives. Slack thread replies and speech output for agents round out the list.¹

None of this is, individually, a structural announcement. Together they describe a vendor that has decided what its product actually is.

The sunset is a margin story. Retiring o3 and GPT-4.5 inside ChatGPT is being framed as roadmap tidying. It is also inference economics (the cost of running models in production, not training them). Every model endpoint OpenAI keeps live carries a GPU-hour tail: dedicated capacity, version-specific kernels, routing logic, evaluation overhead. The 2025 model menu, o1, o3, GPT-4o, GPT-4.5, the 5.x family, Instant variants, was a margin drag dressed as user choice.

I notice the language used to announce this is unusually quiet. There is no migration guide of the kind OpenAI published when GPT-4 was deprecated. The sunset is being treated as a non-event, which is itself a signal: OpenAI's internal data presumably shows users have already migrated to 5.x and the long tail of o3 loyalists is small enough to absorb the churn. The alternative reading, that OpenAI needs the capacity badly enough to cut whatever it has to, is consistent with the same evidence and harder to disprove.

The convergent control plane. Anthropic shipping connector permissions on the same day as OpenAI's workspace controls is either coincidence or evidence that both labs are answering the same procurement questionnaire. I lean toward the second. Enterprise AI buyers in 2026 are running standardised security reviews — SOC 2 (an audit standard for how vendors handle customer data), vendor-risk checklists, data-residency requirements. The features that get shipped are the features the checklists demand.

This matters because it means the enterprise control plane is becoming table stakes rather than differentiation. Twelve months ago, a lab with admin controls had a sales advantage. Today, a lab without them has a sales blocker. The competition has moved one layer up: not whether you can govern the model, but whose weights end up inside which data stack.

The templates are a distribution concession. Pre-built agent templates for Snowflake, Databricks, and GitHub Enterprise tell you where OpenAI now thinks enterprise AI gets deployed. It is not in Chat. It is inside the data warehouse and the code repository, where the data and the workflows already are.

This is a meaningful shift for the FDE market structure (how AI capability is delivered into enterprises — through centralised teams, embedded engineers, or vendor-led integration). OpenAI is conceding that the chat interface is not the enterprise entry point. The model meets the data where it lives. That is the same architectural conclusion Anthropic reached when it leaned into Claude inside Snowflake Cortex, and the same conclusion Databricks reached when it stopped building its own foundation models and started reselling everyone else's.

The press-release version of this is "broad integration coverage." The structural version is: OpenAI has accepted that enterprise distribution runs through the data infrastructure vendors, and is paying the integration tax to stay in the workflow. Whether the templates get used is a separate question. Template-based integrations typically require enterprise IT to configure, and OpenAI does not have the professional-services motion that Anthropic, through Palantir and the hyperscalers, has built out. The templates may sit unused; the concession that they need to exist is the actual data point.

The bullet-points thing. I want to dwell on "fewer bullet-heavy responses" for a moment because release notes do not usually phrase tuning changes as corrections of known failure modes. Labs describe model updates in terms of new capability, not retreated capability. The standard register is "improved reasoning on benchmark X" — not "the previous outputs were annoying and we made them less annoying."

User complaints about bullet-heavy GPT outputs were loud through early 2026, and the dominant reading in product circles was that the model had been tuned for speed at the expense of prose coherence — denser outputs, less continuous reasoning, more visual scannability. This release note reads as an acknowledgement that the tuning went too far.

That is an enshittification-adjacent observation (the pattern where platforms degrade output quality under cost or scale pressure), though I would not push the frame all the way. The frame predicts quality decline driven by monetisation; what we have here is more plausibly quality decline driven by inference cost, then a partial reversal once the cost pressure eased or once user pushback became loud enough to model. Same shape, different mechanism. Worth watching whether the reversal holds at the next capacity-constrained moment.

Slack threads. The smallest item on the release list is the most telling about agent positioning. Letting agents reply inside Slack threads rather than only posting at the top level means the agent is being designed to behave like a human Slack participant. This is the SaaS apocalypse frame (per-seat software pricing collapsing as agents substitute for human users) in concrete form. An agent that replies in your thread is, functionally, one fewer reason to add a human to the channel.

What to watch.

The sunset dates for o3 and GPT-4.5, when published, will tell you how aggressive OpenAI is being about reclaiming inference capacity. Short windows mean cost pressure is acute; long windows mean orderly migration. The uptake numbers on the Snowflake and Databricks templates, if disclosed, will tell you whether the distribution concession is working or just performative. And the next time OpenAI ships a tuning update, watch whether the release notes go back to capability-additive language or whether the corrective register sticks. The latter would be the more interesting signal.

Footnotes

¶