Codex isn't a developer tool anymore. That's exactly why it matters for your firm.

On 2 June 2026, OpenAI announced six new plugins for its coding agent Codex. None of them were for developers. They were built for equity researchers, bankers, salespeople and designers, with corporate finance, private equity, legal and marketing strategy named as next on the roadmap. A tool that started as a way to help engineers write code is now being explicitly repackaged for the same operational roles that already sit inside a law firm, an accountancy practice, an insurance brokerage or an architecture studio. That's worth your attention regardless of whether anyone at your firm has ever opened a terminal.

What actually launched, and when

Two announcements sit behind this. On 2 June, OpenAI's "Codex for every role" update shipped the six role-specific plugins, a feature called Sites that builds and deploys a hosted internal app from a prompt, and an expanded Annotations capability that now works across documents, spreadsheets and slides rather than just code.

On 9 July, OpenAI launched GPT-5.6, its newest model family, across ChatGPT, Codex and the API. It comes in three tiers: Sol, the flagship; Terra, a mid-cost option built for everyday work; and Luna, the fastest and cheapest. The naming itself is a signal of intent. OpenAI has said the generation number tracks the underlying model while Sol, Terra and Luna are durable capability tiers meant to keep advancing on their own cadence across future releases, rather than being replaced wholesale each time. Alongside the model, OpenAI shipped a new ChatGPT Work agent and a desktop app that merges Codex directly into ChatGPT, continuing the same direction as the June announcement: fewer separate tools, more of it living inside the interface staff already use.

The adoption numbers, and why the mix matters more than the headline

Codex passed 5 million weekly active users around the June announcement, roughly six to eight times its usage since the desktop app launched in February. The more useful figure sits underneath that one: non-developers now make up around 20% of Codex's user base, and that segment is growing about three times faster than the developer segment. The centre of gravity on the platform driving this entire trend is already shifting toward the same operational roles most professional services firms employ, not away from them.

A few other figures worth knowing: OpenAI reports around 2 million business customers, 62 connected apps and 110 reusable "skills." Skills are the detail that matters most for a firm like yours: a way to encode a recurring workflow once and invoke it repeatedly on demand, rather than re-explaining the task from scratch every time. That's precisely the shape of most professional services back-office work, client intake, engagement letters, policy renewal checklists, month-end reconciliation steps.

Worth a caveat: these figures come from OpenAI's own reporting, not an independent audit. Treat the direction as real and the exact multiples as broadly indicative rather than precise.

Why "coding tool" was always the wrong frame

The plugins OpenAI shipped in June target equity research, banking, sales and design today, with legal, corporate finance, private equity and marketing strategy explicitly named as next. Legal appearing on that list by name is the direct signal for the audience this newsletter is written for.

The underlying mechanism was never really about code. It's an agent that reads and reasons over documents and data, then produces an editable output, a document, a spreadsheet, a slide deck, a small internal app. Software development happened to be the first market where that was easy to prove out and sell. The same mechanism applies to anything that turns raw information into a structured work product, which is a fair description of most professional services output.

What GPT-5.6 Sol adds on top

Sol introduces an "ultra" mode that coordinates four agents in parallel by default on demanding tasks, trading higher token spend for faster, stronger results. That's relevant to anyone weighing whether a large document review, a due diligence pack, or a complex drafting job can now be handed over as a single request rather than a chain of manual steps.

OpenAI also says Sol's cyber safeguards block roughly ten times more potentially harmful activity than earlier models, and has been explicit that this creates more friction for entirely legitimate requests as a trade-off, hence a built-in option to retry a blocked prompt on a lower-capability model. Worth knowing if a request gets refused unexpectedly: it may be an overcautious filter rather than an actual problem with what you asked for.

One more detail worth flagging simply because it's a pattern worth watching across the industry, not because it's unique to this launch: OpenAI ran GPT-5.6 Sol through an initial preview with a small group of government-vetted partners before wider release, at the US government's request, and has said publicly it doesn't want that kind of access process to become the standard going forward. Every frontier lab is currently operating inside some version of this dynamic. It's context worth having when you're evaluating any of them, not a reason to favour or avoid a particular one.

What this means for your firm

The planning question isn't "should our developers get a coding assistant." Most professional services firms don't employ developers in any numbers that matters here. The actual question is: which of our repeatable, document-heavy workflows already resemble the plugin categories OpenAI just shipped for other industries. For most firms in law, accountancy, insurance and architecture, the honest answer is several.

A few practical steps worth taking now rather than waiting for this to arrive informally through staff already experimenting on their own accounts:

1. Map your own repeatable tasks the way OpenAI mapped its plugin roles. Client intake, engagement letters, standard correspondence, renewal checklists, reconciliation steps. Check whether a general-purpose agent tool already covers most of that before commissioning something bespoke.
2. Treat a working skill as the unit of rollout, not a licence. One workflow that's actually been built, tested and handed to a team is worth more than a tool sitting unused on ninety desktops because nobody defined what it's for.
3. Fund training and governance alongside the tool, not after it. The same reporting that shows non-developer adoption accelerating this fast also shows policy and oversight consistently lagging behind actual usage inside the organisations driving the growth.
4. Don't dismiss anything described as a "coding agent." The fastest-growing user segment on the platform behind this entire trend already looks like your own operational staff, not your engineers, assuming you have any.

The takeaway

This isn't a distant trend to monitor. It's the market actively repackaging a tool built for one profession into something explicitly aimed at the operational roles most professional services firms already staff, and it's moving faster than most firms are adjusting policy to keep up. The advantage sits with whoever plans this deliberately now, rather than discovering six months from now that half the admin team has already been doing it quietly on personal accounts.

Sources: OpenAI's official GPT-5.6 and "Codex for every role" announcements and system card; launch coverage from 9to5Mac, TestingCatalog and Latent Space's AI News; and adoption-figures analysis from Digital Applied, BuildFastWithAI and Gradually.ai, all citing OpenAI's own reported usage telemetry.