Microsoft caps AI tokens: your default model is the lever

The company that sells you Copilot just told its own engineers to use fewer tokens. Microsoft EVP Jay Parikh emailed staff that every division now runs against an AI token budget target, and that the default model in internal GitHub Copilot is switching to OpenAI's cheaper GPT-5.6 Sol. The quotable line — "tokenmaxxing is not what we are optimizing for" — is funny. The mechanism underneath it is the part worth copying.

What actually happened

404 Media first reported the memo, and CNBC confirmed the model switch. The details as reported:

• As of July 2026, each Microsoft division operates under an AI token budget target, with an internal dashboard so individuals can see their own consumption.
• GPT-5.6 Sol becomes the default model in internal GitHub Copilot, chosen because it costs less per token than what staff were reaching for.
• Internal guidance acknowledges many engineers currently run hundreds of dollars to a few thousand dollars a month in tokens.
• Parikh's framing: he wants people "focused on maximizing outcomes that move the needle," not on usage volume.

This is the same shape we covered when Uber and Tesla capped employee AI spend — except Microsoft reached for a different dial first.

Why the default model matters for your business

Changing the default beats policing the cap. A per-seat cap is a fight you have every month with your best engineer. Changing which model answers by default is a config change nobody has to argue about, and it moves the bill immediately. Caps are a ceiling; defaults are the floor everything else builds on.

Cheap by default, expensive on request. The right architecture is tiered: a low-cost model handles the boring 80% — renames, test scaffolding, boilerplate, doc lookups — and the expensive model gets called explicitly when a task actually warrants it. If your team can't switch models without changing tools, you don't have a cost lever, you have a subscription.

Know what a task costs, not what a seat costs. Microsoft's own engineers were running four-figure months. If a company with that much leverage on OpenAI pricing has to meter internally, a five-person shop paying list price should assume its worst month is coming. Track spend per repo or per workflow, not per person — that tells you which automations pay for themselves and which ones you're subsidizing out of habit.

Sources: 404 Media, CNBC, The Register.