OpenAI's revenue is $20 billion short of what everyone assumed
Issue #055 · 4 min read
OpenAI's revenue is $20 billion short of what everyone assumed
Plus: fired safety researchers push back, and GitHub Copilot gets a sandbox you'll actually want
The big story
For months the number everyone repeated was $70 billion — OpenAI's reported annualized revenue, the figure that justified every valuation conversation in the industry. A new report says the real number is $20 billion lower than that. That's TechCrunch.
Twenty billion dollars is not a rounding error. It's the kind of gap that changes what a Series C term sheet looks like, what an enterprise buyer assumes about vendor stability, and what every "OpenAI-powered" startup should be modeling for 2027.
The timing is not great for OpenAI either way you look at it.
The same week, three fired AI safety researchers published an open letter disputing the company's account of why they were let go, warning that the firings are creating what they call a chilling effect inside the safety team — that's TechCrunch again, and OpenAI has since doubled down on the decision rather than walking it back, according to The Verge.
A revenue miss is a spreadsheet problem. A safety team in open revolt is a trust problem. Having both in the same week is worse than either alone.
If you've built pricing or a roadmap assumption on "OpenAI's growth curve," this is the week to stress-test it rather than take the old number on faith.
What shipped
GitHub's local sandboxing for Copilot is now generally available across the CLI, the desktop app, and VS Code sessions using Agent Host. The pitch is straightforward: your coding agent gets a secure local sandbox instead of free rein on your machine. If you've been nervous about letting Copilot's agent mode touch real files, this is the control you've been asking for — see the GitHub changelog.
Claude Haiku 5.5 landed in GitHub Copilot, positioned for the high-volume, low-stakes work — subagents, quick edits, terminal tasks — rather than the big reasoning calls. Worth testing if your team's Copilot bill is mostly going to small repetitive requests that don't need a flagship model, per the GitHub changelog.
Anthropic is now scanning open-source projects for security holes, for free. It's a straightforward goodwill play aimed at the ecosystem Anthropic's own models are trained on and deployed into, reported by The Verge. If your product depends on open-source packages — and whose doesn't — it's worth knowing this exists before you pay for a commercial scanner that does the same thing.
One worth flagging rather than celebrating: Scale AI quietly removed 89 tasks from its SWE-Bench Pro coding benchmark after an audit found they'd been gamed, then graded its own fix — a detail from an investigation at theinference.org. If you cite SWE-Bench scores in a model comparison deck, that number just got a footnote it didn't have last week.
What I'd actually do this week
Pull your OpenAI-dependent cost assumptions and re-check them against the $20 billion gap. If a partner's growth story was baked into your 2027 plan, ask your finance lead to flag it now rather than after a renewal surprise.
Turn on local sandboxing if you're running Copilot in agent mode. It's GA, it's free, and the downside of not using it is an agent with full access to your machine.
If you're citing SWE-Bench Pro anywhere — a vendor comparison, a hiring rubric, a board deck — go read what got cut and why before you cite it again.
Reply if you've hit the Copilot sandbox or the Anthropic scanner this week — I want to know what it actually caught.