Orosz says OpenAI wins enterprises by not locking them in
Codex runs other models and counts the spend against your OpenAI commit, which Orosz reads as the smarter enterprise play.
Gergely Orosz says OpenAI is pulling ahead of Anthropic on enterprise strategy, and his reasoning is that OpenAI is the one giving buyers an exit. Codex is open source, it lets you use other models, and an OpenAI Codex subscription can be used in your own harness. He says Anthropic does none of those things, calling it "closed everything", with closed source Claude Code and no option to run other models inside it.
To lock in mid-sized and above companies, you need to NOT lock them in...
The trigger is a change Philip Kiely described, that enterprise teams can now use open models like GLM-5.3 Flash and Kimi K3 natively in Codex and count that spend against their OpenAI commit.
Enterprise teams can now use open models like GLM-5.3 Flash and Kimi K3 natively in Codex and count spend against their OpenAI commit.
The argument
Orosz's read is that procurement instincts run the other way from vendor instincts. A sane developer, and especially a company, does not want every egg in one basket, because the model that works well today might not survive the next iteration. So the vendor that makes model swapping easy is the one that gets signed.
"To lock in mid-sized and above companies, you need to NOT lock them in..." is how he puts it.
He flags the own-harness point as a second differentiator and is careful about how much it matters, noting that developers can decide for themselves whether they prefer it or whether they care at all. Some do.
For working engineers
The mechanism worth noticing is the commit. Routing a Kimi or GLM call through Codex does not reduce what you owe OpenAI, it moves that usage inside the contract you already signed. Openness at the model layer, accounting at the billing layer. That is a different kind of stickiness from the one Claude Code enforces by refusing to run anything but Claude, and arguably a more durable one, because it survives your team deciding a rival model is better.
This is a strategy read, not a test result. Nobody in the sources has benchmarked GLM-5.3 Flash or Kimi K3 inside Codex, or reported what the agent harness does when it is pointed at a model it was not tuned against. Orosz is arguing about where enterprise money goes, not about which model writes better code.
The practical question for anyone with a commit to spend is narrower than the strategy debate. If the open model option is real, it changes what happens when the frontier model you standardised on gets worse, slower or more expensive for a release or two. Swapping becomes a config change rather than a renegotiation. Whether that holds up in practice is still unreported.