A business owner watches a broadcast on a tablet propped against a carafe in a kitchen where dinner is waiting

· 6 min read

OpenAI DevDay 2026: what non-developers should watch, and what they can skip

OpenAI holds its DevDay tonight in San Francisco. This article appears before the keynote, so it reports none of what will be said. It offers something else: a way for a business owner having software built to interpret it, and a list of what to skip.

OpenAI holds DevDay 2026 on September 29 at Fort Mason, San Francisco, with a Sam Altman keynote announced for 10 a.m. Pacific time, or 7 p.m. in Paris, streamed live and freely accessible according to devday.openai.com. This article is published beforehand: it says nothing about the content, but explains what to take from it.

7 p.m. in Paristhe keynote time, 10 a.m. Pacific (devday.openai.com)
$650registration fee to attend in person
8 citiesannounced DevDay Exchange editions, including Paris

A developers' conference addresses developers

DevDay's format is designed for people integrating models into their own programs: new programming interfaces, agent tools, and usage-based pricing. A business owner having a management tool built will use none of those components directly. They will suffer or benefit from them through their chosen tool, months later, sometimes without ever knowing. That distance makes it worth watching the keynote for two or three specific details among the demonstrations instead of following it as entertainment.

The first line to note: prices

A price announced on stage eventually reaches your bill, after a delay. In July, OpenAI publicly released the three GPT-5.6 versions, Sol, Terra, and Luna, at per-million-token prices reported by Engadget on July 8, 2026: $5 input and $30 output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. A fivefold difference between the most and least expensive within the same engine family. If new pricing is announced tonight, the useful question is which tier your tool uses by default and whether you can change it without waiting for an update. The full reasoning on agent workflow costs is in token prices fall, bills rise.

The second line: what is disappearing

Retirement announcements flash past on a slide in three seconds and affect your product for months. A recent example illustrates this: on August 29, 2026, Latent Space reported the end of Cursor's direct access to OpenAI models, with a cutoff scheduled for November 12, 2026, for reasons unrelated to Cursor users. We described this mechanism in your tool can lose its engine. Tonight, keep an eye on models or interfaces announced as deprecated and the accompanying date. A retired model means a product to retest, sometimes partly rewrite.

What to skip

Stage demonstrations show the best case, rehearsed, on a prepared machine. They say nothing about what the same model produces for your business, with your rules and data. Model score comparisons are no better at predicting delivered software quality: the measure that matters to you concerns a finished product, not an isolated task, which is why we publish our own defeats. Enterprise partnerships concern organisations of a different size from yours, and adoption numbers announced on stage cannot be verified.

Why this keynote still matters to you

Codex is one of the four assistants Maestro can control, alongside Claude Code, Cursor, and Mistral Vibe, and that choice can change mid-conversation. An announcement making an assistant better or cheaper therefore becomes a setting rather than a migration. Apply that calculation to any tool you consider: if an evening's news forces a complete rebuild, the dependency was too tight from the beginning. The full range of options appears at /comparatif. A second reason to follow the evening is the calendar: OpenAI announces DevDay Exchange events in eight cities, including Paris, and local events often translate announcements into demonstrations closer to real use than the main stage. Attending in San Francisco costs $650 in registration; the livestream costs nothing, and the written account published the following day takes ten minutes to read, still the best return on your time.

In practice, tomorrow morning

Two actions suffice. Note the announced prices and compare them with what your tool consumes today, using your token counter rather than a back-of-the-envelope estimate. Then find the deprecation list in the post published after the keynote, and put the nearest date in your calendar alongside the name of whoever must handle it. If neither piece of information changes anything for you, the evening was entertainment, and you missed nothing by having dinner.

Compare alternatives to Lovable: pricing, code and local workflows

Back to the journal

Take the baton.

Leave your email to try Maestro in the first waves.

The beta is open by invitation on macOS 13 and later. Leave your email for an upcoming wave of access. Windows is in development.

The beta is currently available on macOS 13 or later. Your answer helps us plan other versions.

Your email is only used to let you know when access opens. Nothing else, we promise.