AI model loyalty, enterprise edition: it's not you, it's your price per token
SummaryAt Rippling, Anthropic models went from doing 57% of engineers' coding work in May to 18%, while OpenAI's share rose from 16% to 52%, The Information reports. The reason its CEO gave: Rippling wanted to cut costs, and OpenAI was cheaper.
· MODEL Editors · 2 min read
Some links may be affiliate links. Disclosure
Developers argue all year about which AI model writes the best code. At one big software company, the deciding vote turned out to belong to finance.
What was reported. In a story published October 24, 2026, titled "Anthropic's Margin is OpenAI's Opportunity," The Information's Laura Bratton and Stephanie Palazzolo report that OpenAI has been winning AI coding work away from Anthropic with discounts and free credits, while Anthropic has played hardball in negotiations. The example that took off on X on Saturday:
- At Rippling, the HR software company, Anthropic models handled 57% of the coding work of its engineers in May, and OpenAI models 16%.
- Five months later, that has flipped: OpenAI 52%, Anthropic 18%.
- CEO Parker Conrad told the outlet Rippling wanted to cut costs, and OpenAI proved cheaper.
Coverage of the same article says OpenAI also gave Replit more than 30% off a contract worth over $100 million a year, after Anthropic declined to extend Replit's earlier discount without a bigger spending commitment. Replit's CEO had already said onstage on October 16 that OpenAI is now the default for Replit Agent, citing cost, partnership and performance; Claude models remain available to customers who pick them, TechCrunch reported.
The other side. None of this says one model is better. Rippling still uses Anthropic for almost a fifth of the work, and when routing between providers is this easy, the next price cut or the next better model can move the work right back.
Our take. For years, picking a coding model felt like picking a favorite editor. At company scale it is turning into a procurement line: if two models are close enough on your tasks, the per-token price decides. If you build on these models, keep your prompts, evals and tooling portable so you can switch, in either direction, without a rewrite.
What we did and didn't do: we read The Information's headline and the reporters' public post, TechCrunch's Replit report, and other outlets' summaries of the article on October 24, 2026. The full Information article is paywalled and we did not read it in full; the Rippling and Replit figures are as reported there and have not been published by either company. We have not tested either company's models for this story. Not sponsored; no affiliate links.
MODEL Editors
Sources
FREE · EVERY WEEK
The top AI tools this week, tested.
Every week MODEL's editors pick the AI tools worth your time, try them hands-on, and tell you what's good, what's not, and who it's for. Free, in your inbox.
Free. Unsubscribe anytime. Our picks are editorial and can't be bought.
The weekly three is editorial. A subscription does not buy a place in it.
Continue
Hot take
The LiteLLM hack started with a coding-agent config file. The leaked key was force-pushed over, never revoked.
LiteLLM's October 23 root-cause report traces the poisoned 1.107.0 release back to an opencode.json file committed to a public repo on May 27. It held the master key to an internal agent platform, and that platform held an organization owner's unscoped GitHub CLI token. The key sat live for 136 days before anyone used it.
· MODEL Editors
Tech
The internet can now send smells. It knows five, and they're an office party
Ohio State and Zhejiang Lab researchers built Scentronic: an AI e-nose that recognizes a smell, sends its name over the internet, and a wearable that sprays the matching scent on the other end. The catalog is coffee, beer, red wine, cake and orange.
· MODEL Editors
Security
LiteLLM 1.107.0 shipped a file called cook.pth. It stole credentials.
For about two hours on October 21, the litellm package on PyPI carried a credential stealer that runs every time Python starts, even when you run pip. If you installed 1.107.0 or pulled a floating Docker tag that morning, treat the machine as compromised and rotate everything it could reach.
· MODEL Editors