Replit's CEO says the company has stopped working with Anthropic. Speaking onstage at TechCrunch Disrupt on Thursday (reported October 16, 2026), Amjad Masad said the decision was "based on both cost, partnership, and performance," according to TechCrunch's report. He added a pointed line on quality: "I think ChatGPT is better now."
The record, though, is messier than the headline. TechCrunch reports that an anonymous person familiar with the relationship said the Anthropic partnership is still in place, and that a September Replit blog post listed Claude Fable 5 and Fable 5.1 among the models running in the production Replit Agent. TechCrunch asked Replit to clarify and said it would update its story. At the time of writing we have not seen a Replit or Anthropic statement that resolves the conflict.
This post separates what Masad said, what other sources show, and what a builder should take from a customer-vendor split at the center of the vibe-coding market.
TL;DR: claimed vs verified
| Question | Answer |
|---|---|
| Who said it? | Replit co-founder and CEO Amjad Masad, onstage at TechCrunch Disrupt (TechCrunch report dated October 16, 2026) |
| What exactly was claimed? | Replit is no longer working with Anthropic; the reasons were cost, partnership and performance |
| Performance claim | On Replit's internal and a public benchmark, "Sol and Astra" (OpenAI) do better at building vibe-coded apps from scratch |
| Is there a written statement? | Not one we found. Replit's blog index shows no post announcing a break, and Anthropic has not commented in the sources we read |
| What conflicts with it? | An anonymous source says the partnership continues; a September Replit post listed Claude Fable 5 and 5.1 in production Agent (as reported by TechCrunch) |
| What is on record about OpenAI? | Replit's blog features a "Free Mode" post from August 18, 2026 headlined "With OpenAI, Replit now helps you accomplish even more, every day" |
| What is still unknown? | Whether Claude models were removed, when, whether it is a commercial or a technical change, and what "not working with" means in practice |
What Masad actually said
TechCrunch's account, by reporter Lucas Ropek, gives four distinct claims.
- Replit and Anthropic are no longer working together. The wording reported is Masad's, from a live interview, not a press release.
- The reasons are cost, partnership and performance. Masad did not break out which mattered most.
- Replit benchmarks say OpenAI is ahead for its core use case. Per TechCrunch, he said Replit uses an internal benchmark and a public one, and that "Sol and Astra perform better on building vibe-coded applications from scratch."
- The relationship is better with OpenAI "for now." His phrasing: "we just felt like they were a better partner for now."
"For now" matters. It reads as a statement about the current model generation rather than a permanent alignment, which fits how fast the leaderboard has moved this year. See our comparison of Gemini 4 Argon, Opus 5.5, Grok 4.7 and GPT-6 Astra for how close the frontier has been.
He also shared business numbers in the same session. Replit is not disclosing annual recurring revenue yet. Enterprise was about 2% of revenue last year and is about 15% this year, and more than half of revenue now comes from business customers, per TechCrunch. That context explains why cost and partnership terms would matter: an enterprise-heavy vendor negotiates volume, support and indemnity as much as it picks the best model on a leaderboard.
What the public record shows
We checked Replit's own blog index on October 16. Three things stand out.
An OpenAI tie-up is documented. An August 18, 2026 featured post titled "Replit Introduces Free Mode to Expand What is Possible with AI" carries the subhead "With OpenAI, Replit now helps you accomplish even more, every day." TechCrunch also notes Replit announced an OpenAI partnership in August. A week later, on August 26, Replit published "Intelligent Model Routing on Replit," saying it now picks the best model for every task automatically.
Replit describes itself as model-flexible. In a September 29 engineering post, "Free the models: Harness design at the frontier," Replit argues that routers are less capable than the models they choose between, and that its Agent lets the main model decide the tier and effort of its own subagents. The post says GPT-6 Astra hands routine implementation to cheaper subagents, and that on DeepSWE and Terminal-Bench the Replit Agent is Pareto-efficient against Astra on its own. That is a Replit claim about its own harness, not an independent result, but it shows OpenAI's flagship was already central to Replit's agent design before Masad's remark. Our write-up of GPT-6.1 Sol and Astra cost efficiency covers why those two tiers are priced and used differently.
Claude has also been a first-class citizen. On June 17, 2026 Replit announced it is available directly inside Claude, letting people design in Claude Design and send the result to Replit. That is an integration where Claude is the front door to Replit, which is a different thing from Replit using Claude as an engine. A split in one direction does not automatically end the other, and nothing we found says it has.
Why the conflict in the reporting matters
Three readings fit all the evidence at once:
- Narrow reading. Replit stopped using Anthropic as a primary or default model for its Agent, while some Claude models still exist in a picker or in legacy paths. The September post listing Fable models could be accurate for September and stale by October.
- Commercial reading. The enterprise or API contract between the two companies ended or lapsed, while individual users can still choose Claude through other routes.
- Overstatement reading. Masad was speaking loosely onstage and the actual change is smaller, which is why a source close to the relationship says it continues.
We cannot tell these apart from public sources. That is why this is a claimed-versus-verified post and not a confirmed split. We use the same standard in other cases, such as our look at the Anthropic Scion and Meta Mononoke attribution dispute and the report on EU AI Office probes of OpenAI and Anthropic: attribute the claim, list what is documented, and say what is missing.
The performance claim: what it can and cannot tell you
"ChatGPT is better now" is a strong line, but read it for scope.
- It is one task. Masad specified building vibe-coded applications from scratch, the core Replit workload. That is different from editing a large existing codebase, debugging, or writing secure code.
- It is the customer's benchmark. Replit describes an internal benchmark and one public one, and has published its approach to evaluating Replit Agent at scale in a June 23 post. We have not seen the new numbers, so the size of the gap is unknown.
- Other tests disagree on other axes. On security, for example, our summary of the Last Secure Code benchmark for vibe-coding agents found that even strong models produce insecure apps most of the time, so "which model builds apps best" depends on how you score it.
- Cost is part of the claim. Masad listed cost first. Prices and capacity limits have been shifting at Anthropic; we covered the pause of Claude startup credits over capacity earlier this month. We have no evidence that this is linked to Replit's decision, and it should not be read as such.
Vibe coding as a wave of generated apps, illustrating the Replit model choice
Why a vibe-coding platform's model choice matters
Replit sits where non-engineers turn prompts into running software, the pattern we described in software for one and personal apps built with AI coding agents. Whichever model powers the agent determines how often an app works on the first click, how much each build costs the platform, and how much margin is left for a free tier. Replit's "Free Mode" framing makes inference cost a front-line business issue.
Consumer spending data points the same way. a16z's list of the top consumer AI products, which we summarized in our top 100 consumer AI spending post, shows that a small share of users pays, so platforms with large free tiers are sensitive to the per-token price of whichever lab they rely on.
For the labs, losing or reducing a flagship coding-platform customer is a signal in a market where coding is the biggest revenue line. For OpenAI, it is a reference win. Our 28-day log of OpenAI Astra updates shows how aggressively the company has been pushing the coding story.
What developers should take from it
- Treat the underlying model as a dependency that can change. If you build on Replit, Cursor-style tools or any agent platform, assume the model behind the default can change within weeks, sometimes without a big announcement.
- Test behavior, not brand. Keep a small set of prompts and acceptance checks that represent your app type, and rerun them after a platform change.
- Check the picker. Open your own workspace's model settings and changelog today. Whatever the headline, your account is the ground truth for what you are being billed for.
- Do not infer a ranking. One company's reasons are cost, partnership and performance. Only the third is about the model, and it is self-reported.
- Watch for confirmation. The things that would settle this are a Replit blog or changelog entry, an Anthropic statement, or a TechCrunch update. We will add an update section here if any of them appears.
How to check this yourself
- Read the original TechCrunch report and look for an update note at the top.
- Scan the Replit blog for a post on model changes, and review the product changelog and model picker in your account.
- Compare against Replit's September 29 harness post to see which models it names.
- Look for any Anthropic customer or partner page that still lists Replit.
Bottom line
Verified: Masad said, onstage, that Replit is no longer working with Anthropic, and gave cost, partnership and performance as reasons. Verified: Replit and OpenAI announced a partnership in August, and Replit's engineering blog features GPT-6 Astra. Not verified: that all Claude usage in Replit has ended, that the benchmark gap is large, or that Anthropic agrees with the description. Until Replit or Anthropic publishes something, read this as a strong CEO claim with a conflicting source, not a closed case.
Related reading
- Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
- GPT-6.1 Sol and Astra: cost efficiency
- OpenAI Astra: 28 days of updates
- Anthropic pauses Claude startups credits over capacity
- Last Secure Code benchmark for vibe-coding agents
- Software for one: personal apps built with AI coding agents
- a16z top 100 consumer AI products and who pays
Details are accurate as of October 16, 2026 and may change as Replit and Anthropic respond.