How default placement moves spend between labs

When a coding platform sends most sessions to one lab’s models, that product decision moves usage, spend and revenue between vendors without anyone signing a new contract. Replit making OpenAI its default for Replit Agent is a clean example of the mechanism, and it is a different question from whether the Anthropic partnership has ended.

The headline changed; the URL and opening sentence still carry the old claim

Readers arriving from social shares or the original link land on a headline that has changed. TechCrunch published the report on 2026-10-16 and later updated it. The URL slug still reads “replit-is-no-longer-working-with-anthropic-ceo-says”, and the first sentence of the body still says Replit “is no longer working with Anthropic”. The displayed headline is now “Replit makes OpenAI its default over Anthropic, CEO says”. Two claims live in the same article: a product default the CEO has stated repeatedly, and an ended partnership he has not restated since.

What Masad said onstage, and what he said after

Speaking at TechCrunch Disrupt, CEO and co-founder Amjad Masad said: “I think ChatGPT is better now,” adding that Replit publishes a public benchmark and that “consistently, Sol and Astra perform better on building vibe-coded applications from scratch. It was a decision based on both cost, partnership, and performance. And for all these reasons, we just felt like they were a better partner for now.”

After Disrupt, he clarified that customers can still access Anthropic models on the platform while OpenAI’s models are now the default. Over the weekend he told TechCrunch: “Replit remains model agnostic… Our users can choose Anthropic’s Claude or other supported models… OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership.”

His own post on X on 2026-10-17 used slightly different wording: “OpenAI is now our preferred partner powering Agent based on performance, cost, and strength of our partnership.” That is the CEO’s phrasing of the clarification; the longer version above is TechCrunch’s rendering of it.

One naming note for anyone reading the product docs alongside this. Replit’s documentation lists models including GPT-6 Sol, GPT-6 Astra and Claude Fable 5.1. “ChatGPT is now the default” is the shorthand TechCrunch used for that clarification, not documentation language.

The claim that the partnership ended is disputed

TechCrunch reports that Masad’s onstage comments “were rejected by insiders familiar with the partnership”. One person, who asked not to be named, said the partnership was still in place. Anthropic did not comment in the piece, and the company has published no statement about Replit since.

Recent material from Replit and Anthropic cuts against the claim that the partnership ended. A “free-the-models” engineering post on 2026-09-29 names Claude Fable 5 and Fable 5.1 running in Replit Agent’s production environment. The same post says: “The code quality of Astra and Fable 5.1 also let us use our code-review subagent less, with no drop in our eval scores.” A Replit product employee, Samuel Spitz, posted about building with Opus 5.5 on 2026-09-25. Anthropic also still publishes an undated Replit case study describing the older Agent 4 generation.

None of that settles whether a contract is active. TechCrunch still carries the onstage sentence as recorded, and the open questions are whether a contract exists, whether it changed, and what either company means by partnership. The dispute is unresolved on the public record.

Replit’s documentation never flags an OpenAI default

Replit’s model selector documentation describes two behaviours. Auto, which Replit calls intelligent model routing, chooses an appropriate model for each task and can switch as the task changes. Manual selection lets builders pick a model. Power Mode and Max Mode are where Core and Pro builders can choose a model, while Free Mode always uses routing. Enterprise workspaces use a single Auto mode in place of Free Mode, and builders there can turn Auto off to select an authorised model. Account and Workspace policies must keep at least one model from OpenAI, Anthropic and Google enabled.

Nothing in those pages names an OpenAI default or says which model Auto prefers. The selector page also omits some models: the docs changelog for 2026-10-16 added Claude Haiku 5.5 to Power Mode, and the selector page as it currently reads does not list it. The page itself says the in-product list is the source of truth and varies by rollout and account settings.

The documentation neither confirms nor contradicts the switch. It does show Claude models remaining selectable, which makes a Claude-dependent team’s access question an account question rather than a contract question.

Replit’s benchmark is Replit’s own harness

The benchmark Replit publishes is its own harness, reported in the “free-the-models” post. Replit says it ran Agent in Max mode on DeepSWE and Terminal-Bench “exactly as it ships to users”, that its figures are “means of four repetitions”, that three GPU tasks are excluded, and that it compared against published mini-swe-agent baselines plus a sidekick architecture it built itself. These are vendor-claimed numbers from conditions the vendor chose. The public runs test repository changes on DeepSWE and shell tasks on Terminal-Bench, not the from-scratch app building he described onstage, and the charts compare Replit Agent with Astra’s published baseline and a sidekick, with no Sol-versus-Claude run.

Independent leaderboards do not support a clean OpenAI lead. On Datacurve’s DeepSWE v1.1 board, Claude Opus 5 and GPT-6 Astra both sit at 74 percent within their error ranges, and Gemini 3.8 Flash also reaches 74 percent at $2.36 per task against Astra’s $4.43. Artificial Analysis, which runs Terminal-Bench 4.0 itself, puts Claude Sonnet 5.5 at the top with 63.6 percent, ahead of Claude Opus 5.5 at 59.6 percent.

These are different harnesses, tasks and model versions from Replit’s, so a direct comparison with its measurement is not sound.

The revenue mix is the CEO’s own estimate

Masad said enterprise moved from “like 2% of revenue last year” to “like 15% of our revenue” this year, with more than 50 percent of revenue from B2B, and that Replit had been until recently a top-five customer of Anthropic’s. Replit has published no ARR figure of its own, absolute revenue, definition separating enterprise contracts from B2B, or audit. The company is private. His backstage expectation that enterprise and government will supply more than half of revenue in 2027 is an expectation, not a number.

TechCrunch’s statement that the default shift “could significantly reduce Replit’s spend with Anthropic” is the reporter’s inference about how defaults behave, and it is not a measured amount. The mechanism is real in principle: a default change can move spend between labs when customers keep the default. No published figure shows how many Replit customers do.

What changes for a team building on Replit

If Claude is in your workflow, the documented position is that Claude models remain selectable on Core and Pro in Power Mode and Max Mode, and Enterprise workspaces must keep an Anthropic option enabled. Public documents do not show which model Auto picks for a given task. No third party has measured it.

The practical step is to pin the model you depend on in the selector and check the list for your specific account, since rollout and account policy can change what is available. For teams comparing platforms, the default belongs in the switching-cost calculation: an automatic choice you cannot see is still steering your traffic.