Replit names OpenAI its preferred Agent partner while Claude remains available

Replit CEO Amjad Masad says OpenAI is now the preferred partner for Replit Agent, and that its models are the default. The choice follows Replit's August OpenAI partnership, but it does not remove Anthropic's Claude from the choices Masad described to TechCrunch after his October 15 Disrupt appearance. Replit's model selector documentation currently lists models from both companies.

That distinction matters to a builder who depends on a particular model, and to a team that sets which providers its staff may use. BIG CHANGE reviewed the interview report and Replit's public product material. We did not inspect an account's live selector or reproduce Replit's model comparisons.

The big change

• What changed: Replit's CEO has named OpenAI the preferred partner behind Agent while continuing to describe Replit as model agnostic. The model a user gets automatically and the models a paying user may select are separate product decisions.
• Why it matters: A default can steer ordinary use without closing off alternatives. Replit's published controls let Core and Pro builders select an available primary model in Power or Max Mode, while enterprise administrators can limit the models offered in their workspaces.
• What to watch: The available model list and Replit's routing choices will show how this preference reaches users. Neither the CEO's statement nor a model menu measures how often a provider is actually used.

What Masad said, and what he clarified

Speaking at TechCrunch Disrupt on October 15, Masad said Replit had favored OpenAI because of performance, cost and the partnership. He cited an internal benchmark and said a public benchmark also favored OpenAI's Sol and Astra models for building applications from scratch. TechCrunch later obtained a narrower clarification: Replit users can still choose Claude or other supported models, and OpenAI is the preferred partner for Replit Agent. The updated report also says Masad identified OpenAI as the default.

The distinction corrects the impression left by an earlier version of the report, which said Replit was no longer working with Anthropic. TechCrunch says a person familiar with the arrangement disputed that account. The current evidence establishes Masad's preference and his statement about user choice; it does not establish that Anthropic access ended or quantify a change in either company's business with Replit.

Replit had already made OpenAI part of its product. Its August 18 Free Mode announcement named OpenAI's GPT-5.6 Luna as the model powering that mode. The August 26 routing announcement described automatic model selection for tasks and manual selection for Core and Pro subscribers. Replit's September 29 engineering account discussed both Claude and GPT models in its Agent production and evaluations. Masad's October comments therefore concern a preferred provider within a product that already used more than one model family.

What builders can choose

Replit's current Model selector documentation says Free Mode uses automatic routing and has no manual model selector. Core and Pro users can choose an available primary model in Power or Max Mode. The documented lists include OpenAI and Anthropic models, with different lists for the two modes. Enterprise administrators can restrict a workspace's approved model set, according to Replit's routing announcement.

Those controls explain why a preferred partner does not mean one model serves every Agent task. Replit says its automatic routing may select a different model as a task changes. A manually selected primary model also does not document every specialist model used by the underlying agent: Replit's engineering account describes a core loop that can delegate work to subagents. The current menu remains the authority for a particular user's access because the documentation says availability varies with rollout, organization settings and authorization.

Masad attributed the partner decision partly to benchmark results, but the TechCrunch interview does not provide the test tasks, versions and settings needed to assess that comparison. Replit's September engineering post contains evaluations of an Agent configuration using GPT-6 Astra; those results do not establish the claimed across-provider preference for a user's present task. We found no verified public artifact for the specific benchmark Masad referenced and did not run a comparison ourselves.

For teams choosing Replit Agent, the practical question is whether their required model appears in the selector for their plan and workspace, then how the resulting Agent behaves on their own work. The default may change what an untouched account uses; it does not by itself show that customers switched models, spent less, or gained better results.

Sources & further reading

• Replit makes OpenAI its default over Anthropic, CEO says. Kirsten Korosec, Lucas Ropek and Rebecca Bellan, TechCrunch, updated October 17, 2026 after an original October 16 publication. Original reporting from Masad's Disrupt appearance and subsequent clarification. It establishes his statements and the report's correction, not independent product performance or usage.
• Model selector. Replit documentation, undated page read October 18, 2026. Lists automatic routing, plan and mode requirements, manual model options, and rollout and organization limits. BIG CHANGE did not inspect a live account.
• Intelligent Model Routing on Replit. Replit, published August 26 and updated August 27, 2026. Describes automatic routing, Core and Pro manual selection, and enterprise approved-model controls. Its cost and quality results are company claims, not independent measurements used here.
• Replit Introduces Free Mode to Expand What is Possible with AI. Replit, published August 18 and updated August 19, 2026. Establishes the earlier OpenAI-powered Free Mode partnership and its stated mode structure.
• Free the models: Harness design at the frontier. Replit engineering, September 29, 2026. Describes Agent delegation and reports Claude and GPT models in production/evaluations. Its own benchmark configurations do not verify the later broad partner comparison.