If your teams build on Replit, more of their Agent traffic is about to run on OpenAI models instead of Anthropic's, and nothing in Replit's standard DPA obliges it to tell you first. OpenAI was already on Replit's subprocessor list before the switch, and the DPA's objection right only attaches to a change of subprocessor. The controls you do have are in the admin console: restrict which models your Workspaces can reach, then re-test what your builders shipped on Claude before you trust the new default.
On stage at TechCrunch Disrupt this week, Replit CEO Amjad Masad said the company was no longer working with Anthropic. "It was a decision based on both cost, partnership, and performance," he said, according to TechCrunch. Over the weekend he softened it in a statement: "Replit remains model agnostic," users "can choose Anthropic's Claude or other supported models," but "OpenAI is now our preferred partner for powering Replit Agent." TechCrunch's reading is the one that matters for buyers. Claude is still selectable, OpenAI is now the default, and most customers never touch the default.
What Replit Actually Changed
Replit changed the default model behind Replit Agent, and it kept Claude in the catalog. Masad told TechCrunch that Replit had recently been a top-five Anthropic customer and that its benchmarks show OpenAI's models ahead: "consistently, Sol and Astra perform better on building vibe-coded applications from scratch." He also said, plainly, "I think ChatGPT is better now."
The account is disputed. TechCrunch reports that a person familiar with the partnership said it was still in place, and that a late-September Replit blog post still named Claude Fable 5 and Fable 5.1 as running in Replit Agent's production environment. The report carries no on-record response from Anthropic or OpenAI. So treat the business story (is the partnership over?) as open. The product story is less ambiguous: Replit's preferred model is now OpenAI's, by the CEO's own statement.
The current model selector documentation shows both families side by side. Power Mode lists GPT-6 Sol, GPT-6 Luna Fast, Claude Sonnet 5 and Claude Sonnet 4.6. Max Mode lists GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5, Opus 4.8, Opus 5, Opus 5 Fast and Kimi K3. Free Mode builders "cannot use the Model selector or choose a model manually," and a builder's choice "applies to your next message."
The customers this hits are growing fast. Masad said enterprise contracts were "like 2% of revenue last year" and about 15% this year, and that "more than 50% of our revenue is B2B," per the same report. Backstage he said he expects enterprise and government to pass half of revenue in 2027.
Why Your DPA Doesn't Trigger
The switch triggers no contractual notice because no subprocessor was added. Replit's subprocessor list, marked "Last updated: August 2026," already names Anthropic, OpenAI and Google for AI models, alongside Fireworks AI and Baseten for AI infrastructure and OpenRouter for "AI model selection for Integrations." Your data could already go to OpenAI under the paper you signed. Moving the default changes how much of it goes there.
Replit's data processing agreement, dated April 1, 2026, gives Replit "a general authorization" to engage subprocessors. It sets no advance-notice period. It lets you "object to a change of a Subprocessor based on reasonable grounds" relating to data protection law, and if that cannot be resolved you "may terminate the affected portion of the Services," which the DPA calls your "sole and exclusive remedy." A shift in routing volume between two vendors who were both already listed is hard to argue is "a change of a Subprocessor."
That leaves the job with your own vendor-risk process. If your privacy team cleared Replit on the basis that prompts and code went mostly to Anthropic, that assessment no longer describes the data flow. The same gap showed up when Atlassian's OpenAI deal moved Rovo routing without giving admins a provider choice, and it belongs on the list of security-review questions that change the answer: which model provider actually receives the data, and who decides.
What Changes for Retention and Training
At the API level, the two providers' published defaults look alike, so the retention question is really about what Replit negotiated. OpenAI's API data documentation says abuse-monitoring logs are "retained for up to 30 days" by default, that API data "is not used to train or improve OpenAI models" unless the customer opts in, and that zero data retention is "subject to prior approval by OpenAI." Anthropic's commercial retention page, dated July 1, 2026, says it deletes API inputs and outputs "within 30 days of receipt or generation," with exceptions that include a zero data retention agreement.
Those are the providers' terms with Replit, not with you. Replit's DPA says nothing about AI model providers, retention at the model vendor, or training. Whether Replit holds ZDR with OpenAI on the same footing it held with Anthropic is something only Replit can answer, and you should get the answer in writing. Residency is a second difference: OpenAI offers in-region processing only in the United States, Europe and the UAE, per the same documentation. ZDR has gone wrong before on this beat; our piece on why Fable 5's ZDR was never zero covers how a provider's safety exceptions can override a customer's retention setting.
How Much Control Enterprise Admins Have
Enterprise admins can narrow the pool of models the Agent may use, but Replit's documentation does not let them remove OpenAI. The enterprise model controls page says: "At least one model from OpenAI, Anthropic, and Google must remain enabled." Each provider can be set to Enable all, Enable selected or Disabled, and open-source providers start disabled. The providers reference lists six: OpenAI, Anthropic, Google, DeepSeek, Kimi and Z.ai.
Routing is the default. Replit's August 28 changelog made Intelligent Model Routing the default Agent mode for Enterprise Workspaces, with Auto on. In Auto, routing picks a mode and an authorized model for each task, weighing quality and cost, from the models your policy authorizes. Builders can switch Auto off and pick a model themselves. The documentation describes no admin setting that forces Auto off for everyone, and you "cannot set a different policy for a user group in the same Workspace."
Put those together and the strongest pin you can apply is this: in each Workspace, Enable selected with exactly one OpenAI model, one Google model, and the Claude model your team validated. Routing can still send a task to the OpenAI or Google model you left on. Replit has not published how "preferred partner" changes routing weights inside Auto, so do not assume it leaves them alone. One more line from the docs is worth reading twice: "These controls are not spending caps."
Why Masad's Benchmark Is Not Your Eval
Masad's claim is a vendor claim about a vendor benchmark, and it measures the wrong task for most enterprise use. He described Replit's internal benchmark plus a public one, on "building vibe-coded applications from scratch." Your builders are more likely extending an app that already exists, with your auth, your schemas and your internal APIs. A model that wins at greenfield scaffolding can still regress on a large change to code it did not write.
Small eval sets also mislead. When we looked at agent benchmark confidence intervals, only 3 of 36 model gaps held up once sample size was accounted for. And prompts tuned against one model family stop paying off on the next: instructions your teams wrote into Replit projects while Claude was the default may behave differently under GPT-6 Sol.
The steel-man for Replit is reasonable. Cost per task matters at its scale, OpenAI's models may genuinely be better at its core workload, and Claude stays available to anyone who asks for it. None of that tells you how your own apps behave on the new default. That is a test you run, and the model router buyer's guide explains why a router's judgment is the part you have to verify yourself.
What to Do Before Your Builders Notice
This Week:
- Have your Replit account admin open Settings, then Advanced, then Agent modes and model providers, and set every provider to Enable selected. Keep the one Claude model you validated, plus the minimum one OpenAI and one Google model the policy forces you to keep. Use "Copy to all" so no Workspace is left on Enable all.
- Email your Replit account team three questions and ask for written answers: does Replit hold zero data retention with OpenAI for Enterprise traffic, which region processes that traffic, and how did the default change alter Auto mode's routing weights.
- Ask your privacy lead to update the vendor record for Replit. OpenAI is now the preferred provider behind the Agent, whatever the original assessment assumed.
This Month:
- Pick ten apps your teams built in Replit Agent and replay their last few real change requests on the validated Claude model and on GPT-6 Sol. Score each run on whether the change merged and passed your tests.
- Write the result into your Replit usage guidance: which model builders should select, and when to switch Auto off.
Before Renewal:
- Ask for a contract term that requires notice when Replit changes the default model or the routing preference for your Workspaces, separate from the subprocessor clause. Our review of AI vendor exit clauses found standard terms give 0 to 90 days to get out, so ask for this clause alongside the exit terms.
- Ask for the right to turn Auto off for an entire Workspace, so a pin cannot be bypassed one builder at a time.
The Bigger Picture
Every AI application you buy sits on a model its vendor can swap, and the subprocessor list only tells you who could receive your data. Who actually receives it is decided by a default setting. Harvey swapped models when agent costs sank its margin. OpenAI cut Cursor's model access on its own timetable. Replit moved on cost, partnership and performance, by its CEO's account. Cloud buyers went through the same thing with hosting regions, and the fix then was to name the region in the contract. The equivalent now is naming the model.
Masad called Replit "model agnostic." Your admin console currently says at least three providers must stay enabled, so set the pool yourself this week.
Continue Reading
- Atlassian's OpenAI Deal Sends Jira Data Where Admins Can't Pick
- Harvey Swapped Models After Agents Sank Its Margin to -50%
- OpenAI Cuts Cursor Off Nov 12. BYOK Voids Your ZDR.
- AI Vendor Security Review: 6 Questions That Change the Answer
- Fable 5 Needs Retention On. ZDR Was Never Zero.
- Model Router Buyer's Guide: Buy Failover, Not Judgment