Replit Says It Has Ended Anthropic Partnership, Citing AI Costs and Performance

Replit CEO and co-founder Amjad Masad says the AI-powered software development platform has stopped working with Anthropic, pointing to model costs, partnership considerations and coding performance. Speaking at TechCrunch Disrupt in October 2026, Masad said OpenAI’s models performed better in Replit’s evaluations for building applications from scratch. The comments mark a potential shift in the AI company’s model strategy as competition intensifies among providers of tools that can generate and modify software from natural-language instructions.

However, the status of the relationship remains disputed. TechCrunch reported that a person familiar with the partnership said it was still in place, while Replit’s September materials continued to reference Anthropic’s Claude models in its production environment. Replit has also announced a partnership with OpenAI, but that announcement alone does not establish that all Anthropic models have been removed. The available evidence therefore points to a significant disagreement over the relationship’s status, rather than a fully documented and independently confirmed termination.

Key Takeaways

• CEO’s statement: Amjad Masad said Replit was no longer working with Anthropic.
• Reasons cited: Model costs, partnership considerations and performance in coding tasks.
• OpenAI models: Masad said OpenAI’s Sol and Astra performed better in Replit’s evaluations for building applications from scratch.
• Conflicting evidence: An unnamed source told TechCrunch that the Anthropic partnership remained in place.
• Recent product references: Replit’s September materials still named Claude models in its production environment.
• Why it matters: The dispute highlights the growing importance of model quality, cost and reliability for AI coding platforms.
• Status caveat: The precise scope and timing of any separation have not been publicly clarified in the reporting reviewed.

What Did Replit’s CEO Say About Anthropic?

Masad made the remarks during an appearance at TechCrunch Disrupt, where he discussed the changing market for AI-powered software development.

According to TechCrunch’s report, Masad said Replit was no longer working with Anthropic. He described the decision as involving cost, partnership considerations and performance, and argued that OpenAI’s models had become more effective for the company’s application-building use case.

The comments are significant because Replit had previously relied on Anthropic’s Claude models in its AI development platform. Claude helped power Replit Agent, a tool designed to turn user instructions into working applications.

Masad’s remarks suggest that Replit is reassessing which model providers best serve its product and customers. But the statement does not, on its own, establish the exact date when any technical migration took place, whether every Claude model was removed or whether any contractual relationship remains active.

Those details matter because AI platforms often use multiple models for different tasks. A company may change its primary provider for application generation while retaining other models for particular workflows, testing or specialised functions.

Until Replit clarifies the scope of the reported change, it would be premature to describe the move as a complete and fully documented exit from every Anthropic product or service.

Why Replit Is Prioritising Cost and Coding Performance

AI coding platforms have to balance model quality against the cost of operating their products. Every request can involve a combination of model inference, tool use, code execution, debugging and repeated attempts to complete a task.

A model that generates a useful first draft but repeatedly fails to complete an application may require more calls and more user intervention. A model that solves the same task with fewer attempts can reduce costs while improving the user experience.

This is especially important for platforms that promise to help people build software by describing what they want in ordinary language. Such systems must do more than produce code snippets: they may need to plan an application, create files, configure dependencies, execute code, diagnose errors and make corrections.

The economics of AI coding

Model pricing is only one component of the cost. The total expense of completing a task can depend on:

• The number of model calls required.
• The amount of input and output processed.
• How often the agent needs to retry or debug.
• The computing resources required to run the generated application.
• The amount of human intervention needed before the result works.

A relatively expensive model may still be economical if it completes tasks reliably in fewer steps. Conversely, a cheaper model may become costly if it produces errors that trigger repeated calls or require extensive manual correction.

This means Replit’s reported decision cannot be reduced to a simple comparison of published token prices. The relevant measure is the cost of delivering a satisfactory result to a user.

Masad’s comments indicate that Replit evaluates model performance using internal tests and a public benchmark. However, the report does not provide enough independently verified benchmark data to quantify the size of OpenAI’s alleged advantage or determine whether it applies across all types of coding work.

Why OpenAI’s Models Matter to Replit

Replit announced a partnership with OpenAI in August 2026. The relationship gives Replit another major model provider to consider as it develops its software-building tools.

During his TechCrunch Disrupt appearance, Masad said OpenAI’s Sol and Astra models consistently performed better at building applications from scratch in Replit’s evaluations. This is a claim made by Replit’s CEO about the company’s testing; it should not be treated as a universal, independently established ranking of AI coding models.

Performance can vary substantially according to the task. A model that performs well when generating a new web application may not necessarily be the best choice for debugging a large existing codebase, explaining unfamiliar code or handling a specialised programming language.

The result can also depend on how the benchmark is designed, which tools the model can access, how success is measured and how much time or computing power is available.

For Replit, the practical question is whether the models it chooses can help customers complete projects more quickly and reliably at a sustainable cost. A close integration with a model provider can also influence product features, technical support and the pace at which new capabilities reach users.

The OpenAI relationship is therefore an important part of Replit’s model strategy, but it does not by itself prove that OpenAI is superior across every programming task or that Anthropic’s models are no longer useful.

The Contradiction: Is the Anthropic Partnership Actually Over?

The most important qualification in this story is that the public reporting contains conflicting accounts.

TechCrunch reported Masad’s statement that Replit was no longer working with Anthropic. The same report also cited an unnamed person familiar with the relationship who said the partnership remained in place.

There is additional context. Replit materials published in September referenced Claude Fable 5 and Claude Fable 5.1 as models operating in the production environment of Replit Agent. That reference raises questions about whether a complete transition had occurred by the time of Masad’s comments.

It does not necessarily prove that the partnership remained unchanged. Product documentation may refer to an earlier configuration, and a company can change its commercial relationship before every technical reference is updated. Equally, a CEO’s statement may describe a strategic decision whose implementation is still underway.

The available reporting does not resolve these possibilities. TechCrunch said it had contacted Replit for clarification and would update its report if it received further information.

Question	What the reporting establishes
Did Masad say Replit had stopped working with Anthropic?	Yes, according to TechCrunch’s report of his remarks
Did he cite cost and performance?	Yes, alongside partnership considerations
Had Replit announced an OpenAI partnership?	Yes, in August 2026
Did conflicting information emerge?	Yes; an unnamed source said the Anthropic partnership remained in place
Were Claude models still referenced in September materials?	Yes, according to TechCrunch
Is a complete termination independently confirmed?	Not by the available reporting reviewed

The distinction between a public statement and a verified operational change is important in technology reporting. Companies can have commercial contracts, product integrations and model deployments that change at different speeds. Without further clarification, the most accurate conclusion is that Replit’s CEO has described a separation, but the precise status and scope remain unresolved.

What This Means for Anthropic and Claude

If Replit has substantially reduced or ended its use of Anthropic models, the development would illustrate how quickly AI model providers can lose a customer when a platform’s performance or cost requirements change.

Anthropic’s Claude family has been used for code generation and agentic software development. Replit was an important example of a platform embedding Claude into a product that allows users to build applications without managing every stage of the conventional development process.

Losing or reducing a relationship of this kind could affect the distribution of a model provider’s technology, especially when the customer serves a large developer audience. It could also signal that AI coding platforms are becoming more willing to change model providers as competing systems improve.

But the commercial significance should not be overstated. The available reporting does not establish the financial value of Replit’s Anthropic relationship, the amount of revenue Anthropic may lose or whether other parts of the relationship remain active.

Nor does one customer’s reported decision establish that Claude has become broadly uncompetitive. Model performance varies by task, and many software teams use different models for planning, coding, testing and review.

For Anthropic, the episode underscores the need to compete on a combination of model quality, cost, reliability and the quality of its relationships with customers.

What It Means for AI Coding Platforms

The dispute reflects a broader shift in the software development market. AI coding platforms increasingly compete on the complete experience of building an application rather than on access to a particular language model alone.

Customers expect these tools to understand instructions, create functional interfaces, connect databases, identify bugs and help deploy working software. A model provider is one component of that system, alongside the platform’s tools, execution environment, security controls and user interface.

As models improve, platform companies may find it worthwhile to switch providers, negotiate new commercial terms or route different tasks to different models. This can create pressure on model vendors to demonstrate measurable value rather than relying only on brand recognition or benchmark headlines.

It also raises questions about portability and vendor dependence. Platforms that build their products around one provider may face engineering work when they change models. Supporting multiple providers can reduce dependence on a single supplier, but it adds complexity in testing, performance optimisation and cost management.

The best technical choice may therefore differ from one platform to another. Replit’s needs are shaped by its own product architecture, customer base, workload patterns and pricing model.

Why the News Matters for Developers and Businesses

For individual developers, the reported change is not evidence that they must stop using Claude or switch to OpenAI. The decision concerns Replit’s provider relationship and the platform’s evaluation of its own product needs.

Developers should continue to choose tools based on their specific tasks, including the programming languages they use, the complexity of their projects, their budget and how reliably a system handles debugging and deployment.

For businesses buying AI development tools, the story highlights the importance of testing complete workflows. A model that performs impressively on a public benchmark may behave differently when given a company’s real codebase, security requirements and deployment constraints.

Teams should compare not only code quality but also completion rates, error frequency, total task cost, data-handling policies and the ability to recover from failures. They should also understand whether a product can change its underlying model without notice and what that means for reliability and contractual commitments.

The lesson is not that one model provider has permanently won. It is that the economics of AI software development are changing quickly, and platform companies are under pressure to keep evaluating their choices.

The Bigger Picture

Replit’s public comments put the economics of AI coding at the centre of the competition between model providers. Cost, coding performance and the quality of a commercial relationship can all influence which models a platform chooses to integrate. OpenAI’s growing role in Replit’s strategy illustrates how model providers compete not only for end users but also for distribution through other software products.

However, the conflicting accounts mean the precise status of Replit’s Anthropic relationship remains an open question. The episode is best understood as a reported strategic shift whose implementation has not been fully clarified, rather than definitive evidence that Anthropic has been removed from all of Replit’s systems. Further statements or updated product documentation would help establish what has actually changed.

Looking Ahead

The next development to watch is whether Replit clarifies the scope of Masad’s statement, including whether Claude models have been removed from production, whether the commercial partnership has ended and whether any limited integrations remain. Updated product documentation, an official company statement or a response from Anthropic could resolve the discrepancy. Until then, claims about a complete termination should remain qualified.

For the wider AI industry, the story reinforces a practical test for model providers: they must deliver enough quality and reliability to justify their costs within real products. As coding agents become more capable, companies will continue comparing models on the full cost and success rate of completing tasks. Those comparisons will shape partnerships, product design and competition across the AI software development market.

Frequently Asked Questions

1. Has Replit officially ended its partnership with Anthropic?

Replit CEO Amjad Masad said the company was no longer working with Anthropic. However, TechCrunch also reported that an unnamed person familiar with the relationship said the partnership remained in place. The exact status has not been conclusively clarified in the reporting reviewed.

2. Why did Replit say it stopped working with Anthropic?

Masad cited costs, partnership considerations and model performance. He said OpenAI’s Sol and Astra performed better in Replit’s evaluations for building applications from scratch.

3. Does this mean Replit now uses only OpenAI models?

That has not been established. Replit announced an OpenAI partnership in August 2026, but the available reporting does not confirm that every Anthropic model has been removed or that OpenAI is the exclusive provider.

4. What is Replit Agent?

Replit Agent is an AI-powered tool that helps users create applications from natural-language instructions. It can assist with writing code, creating files, running applications and correcting errors within Replit’s development environment.

5. Does this prove OpenAI’s coding models are better than Claude?

No. Masad described the results of Replit’s internal and public evaluations for a specific use case: building applications from scratch. Model performance varies by task, benchmark design and tool configuration, so the statement does not establish a universal ranking.

6. What should developers take away from the dispute?

Developers should evaluate AI coding tools based on their own workflows, including code quality, reliability, total cost, debugging ability and deployment needs.

Editorial note, 17 October 2026: An incomplete sentence at the end of this article was removed. No missing wording has been reconstructed.