Replit OpenAI Default Replaces Claude at the Front of Agent

Replit has made OpenAI its default model provider, reversing its earlier dependence on Anthropic even though Claude remains available. The Replit OpenAI default changes which models most customers encounter first inside Replit Agent. It also turns an infrastructure decision into a visible contest over cost, coding performance, and commercial alignment.

CEO Amjad Masad initially described the change more sharply at TechCrunch Disrupt. He said Replit was no longer working with Anthropic and argued that ChatGPT had become better for the company’s needs. His later clarification narrowed that claim: Replit still supports Claude, but OpenAI is now its preferred partner.

That distinction matters. Replit has not removed Anthropic from its platform or abandoned its model-agnostic position. It has changed the default, which can still redirect substantial usage because many customers accept the platform’s initial selection.

The switch also follows a deeper OpenAI partnership announced in August. OpenAI models were already powering Replit’s Free Mode, while Replit had begun publishing production observations involving newer models from both providers.

The resulting story is not a clean Anthropic eviction. It is a strategic reversal in which Replit places OpenAI at the center while preserving enough model choice to respond if performance, economics, or customer demand shifts again.

What the Replit OpenAI Default Actually Changes

OpenAI now occupies Replit Agent’s preferred position, but Anthropic remains inside the product and its operating stack.

Masad told TechCrunch that Replit selected OpenAI based on performance, economics, and the strength of the partnership. His comments followed an appearance at TechCrunch Disrupt on October 15, 2026.

The original statement sounded conclusive. Masad said Replit was no longer working with Anthropic and cited internal evaluations, cost, and the companies’ commercial relationship.

By the weekend, Replit’s position had become more precise. According to the updated default-model account, customers can continue choosing Claude and other supported models. OpenAI is the preferred partner powering Replit Agent, not its exclusive provider.

That clarification separates three decisions that are easy to confuse.

First, Replit controls the model presented as the default. That choice affects ordinary users who do not inspect or change model settings.

Second, Replit can route different tasks to different models. A coding agent is not necessarily tied to one foundation model throughout an entire project.

Third, customers can still make an explicit Replit model choice where the product allows it. Claude therefore remains relevant even when OpenAI receives the initial workload.

Replit’s public subprocessor list reinforces the distinction. Its August 2026 update lists both OpenAI and Anthropic as AI model providers. It also lists Google and OpenRouter, indicating a broader supplier portfolio rather than a single-provider architecture.

The practical effect still favors OpenAI. Defaults shape behavior because they reduce the effort required to begin a task. A customer opening Replit Agent is more likely to use the model placed directly in the standard path than one that requires a deliberate selection.

That can redirect a large share of inference spending without blocking competing models. Masad said Replit had recently been one of Anthropic’s five largest customers, although neither company disclosed the underlying expenditure or usage.

The default therefore acts as a purchasing decision at scale. It determines which provider receives routine traffic, which model gathers more production feedback, and which partner becomes most closely associated with Replit’s user experience.

Yet the remaining Claude option limits how far the announcement can be taken. Replit has not shown that every workload performs better with OpenAI. It has instead decided that OpenAI offers the best overall combination for its standard experience right now.

That final qualification is important. Model performance changes quickly, and Replit’s own engineering work assumes that the strongest choice can vary across tasks. The company is choosing a front-runner without permanently closing the race.

Cost and Distribution Made the Partnership Timely

The switch followed months of product integration that gave OpenAI both an economic argument and a direct distribution channel through Replit.

Replit and OpenAI publicly expanded their partnership in August 2026. Replit introduced Free Mode using GPT-5.6 Luna, a model positioned for fast and economical everyday work.

Under that design, users can plan, explore ideas, request feedback, and handle lighter tasks without drawing from the same usage allowance as more complex builds. Replit can then move demanding work into other Agent modes.

OpenAI described the Free Mode partnership as a result of improving model economics. Masad said OpenAI’s price reductions helped Replit offer that experience to millions of users.

Replit separately said Free Mode allowed customers to create up to 30 times more under its revised product design. That figure comes from Replit and should be read as a company claim, not an independently audited measure of completed software.

Still, the product logic is clear. Model cost is not a secondary concern for an AI coding platform. Every agent turn can trigger planning, code generation, tool calls, testing, error recovery, and follow-up reasoning.

A small difference in the cost of one response can become significant across a long software-building session. The effect grows again when the platform serves millions of people.

Replit also has an incentive to make that complexity invisible. Most users want a working application, not a procurement exercise involving several model vendors.

The company introduced Intelligent Model Routing in August to select models based on quality, speed, and cost as tasks developed. Replit said its testing delivered comparable output quality at 65 percent lower cost than the previous Max Mode system.

That routing result is another company-reported figure. Replit has not published enough independent evidence to treat it as a universal result across every project.

However, routing and defaults solve different problems. Routing selects a model or configuration during work. A default establishes the preferred commercial and technical starting point for the whole system.

OpenAI also gained a distribution advantage outside Replit’s own interface. Since January 2026, ChatGPT users have been able to invoke Replit to create, update, and deploy applications from a conversation.

That integration connects OpenAI’s consumer interface with Replit’s development and deployment environment. It gives both companies an incentive to make the handoff feel consistent.

For OpenAI, the relationship extends its models into an end-to-end software production service. For Replit, ChatGPT becomes another place where potential customers can begin building.

Anthropic still holds a strong position in coding. Claude models remain embedded in several developer products and have a reputation among programmers for code generation and reasoning.

But model quality alone does not decide a platform default. Replit must consider inference cost, capacity, product integration, commercial terms, and how quickly a provider responds when production systems fail.

Masad’s explanation combined all those factors. He did not present the move as the result of one benchmark score. He framed OpenAI as the better overall partner for Replit’s current priorities.

Those priorities are also changing as Replit pursues larger organizations. Masad said enterprise contracts represented about 2 percent of company revenue last year and approximately 15 percent this year. He also said business customers now account for more than half of revenue.

Replit expects enterprise and government customers to contribute more than half of its revenue in 2027. Those buyers tend to care about reliability, security controls, procurement stability, and support alongside raw model capability.

A preferred partnership can simplify that story. It can also create concentration risk if the chosen provider suffers outages, raises costs, or falls behind on important workloads.

Keeping Claude available helps Replit manage that risk. The company can promote OpenAI as the standard while retaining leverage and technical alternatives.

Replit and Anthropic Now Compete Through the Default

The primary contest is not exclusive access to Replit; it is control over the standard path through which Replit users build software.

Anthropic previously occupied a central role in Replit Agent. An earlier Claude deployment study described Replit choosing Claude on Google Cloud after evaluating several models.

That case study said Claude-powered Replit Agent supported more than 100,000 applications through Cloud Run. It presented the relationship as a combination of Anthropic’s model capabilities, Replit’s agent system, and Google Cloud infrastructure.

Replit’s newer materials show that this history did not disappear overnight. In late September, the company published production observations involving Anthropic’s Fable 5 and Fable 5.1 alongside OpenAI’s GPT-6 Astra.

Those observations examined how models delegated work to specialized subagents. A subagent is a smaller delegated process that handles a limited task, such as exploring code, reviewing changes, or testing an interface.

Replit reported that GPT-6 Astra delegated more frequently to general workers. It also said the newer Anthropic models displayed strong engineering quality and reduced the need for a separate code-review agent.

That comparison complicates any simple claim that Replit found Anthropic technically inadequate. Its engineers were still describing valuable behavior from Anthropic models shortly before Masad announced the new default.

The reversal is therefore commercial and architectural as much as technical. Replit is evaluating complete agent outcomes rather than asking which standalone chatbot produces the best isolated code sample.

A foundation model operates inside a harness, which is the surrounding system that manages prompts, tools, context, tests, retries, and delegated work. The same model can behave differently when that harness changes.

Replit’s production system can also compensate for model weaknesses. It can send exploration to one component, implementation to another, and testing to a browser automation process.

This makes the Replit Anthropic partnership less binary than the original onstage comment suggested. Anthropic can lose default status while its models remain useful for specific tasks or customers.

For Anthropic, the pressure comes from usage displacement. If most customers keep the OpenAI default, Claude receives fewer routine requests through one of the most visible AI software-building platforms.

That affects more than immediate revenue. Production traffic generates evidence about failures, preferred workflows, latency, and how nontechnical users describe software tasks.

OpenAI can gain more of that feedback while strengthening its position in agentic coding. Anthropic must then persuade Replit, competing platforms, or users that Claude produces enough additional value to justify an explicit selection.

For Replit, the pressure runs in the opposite direction. Once the company promotes OpenAI as the standard, customers will associate Agent’s output quality with that choice.

A failed build will not be interpreted as an abstract supplier problem. Users will see it as a Replit problem, regardless of whether the underlying failure began in the model, harness, toolchain, or application environment.

The Replit OpenAI default therefore shifts accountability as well as traffic. Replit is publicly endorsing one provider while retaining the ability to use others behind the interface.

That arrangement resembles cloud procurement more than a consumer chatbot preference. Platforms often choose a primary supplier for economics and integration while maintaining alternatives for resilience.

The difference is that foundation models influence visible product behavior. Changing the default can alter coding style, interface design, error patterns, tool selection, and the amount of work completed before human intervention.

Developers should not interpret the decision as an industry-wide verdict on Claude versus OpenAI. Replit’s evaluation concerns its own harness, workloads, contracts, and customer distribution.

Another coding platform could reach the opposite conclusion using a different agent loop or user base. An enterprise maintaining a large existing repository also has different needs from a founder generating a new application from a prompt.

The more defensible conclusion is narrower. OpenAI currently delivers the combination Replit wants for the default path, while Anthropic remains valuable enough to keep available.

The Benchmark Case Is Still Incomplete

Replit has explained the criteria behind its decision, but outsiders cannot independently reproduce the complete comparison.

Masad said Replit uses an internal benchmark covering measures such as performance. He also pointed to a public benchmark and said OpenAI’s Sol and Astra models consistently performed better when building vibe-coded applications from scratch.

Vibe coding describes software creation driven mainly through natural-language instructions, with an AI system producing and revising much of the code. It emphasizes completed outcomes over manual implementation.

That definition creates an evaluation challenge. Building an application from scratch involves more than generating correct functions. The agent must interpret an ambiguous request, choose architecture, create an interface, connect services, test behavior, and recover from errors.

A benchmark that captures those steps can be useful. It can also favor the harness and workflow for which it was designed.

Replit is both the evaluator and the platform selecting a commercial partner. That does not make its findings invalid, but it limits what the findings establish outside Replit.

The company has published some methodological detail about its agent evaluations. It has also released production statistics showing how different models delegate work inside its system.

Those disclosures offer more substance than a bare claim that one model feels better. They still do not provide a complete, independently audited comparison of OpenAI and Anthropic across cost, reliability, security, and application quality.

Cost is particularly hard to compare from public information. A platform of Replit’s scale may negotiate terms that differ from published API conditions.

Its total expense also includes failed generations, cached context, retries, tools, subagents, and support. A model with a higher unit cost can be cheaper overall if it finishes tasks with fewer attempts.

The reverse can also happen. A lower-cost model may increase total spending if it generates more errors, uses more tokens, or requires heavier review.

Partnership strength is even less measurable. It can include joint engineering, capacity commitments, response times, marketing, product distribution, and negotiated commercial terms.

Readers should therefore separate Replit’s decision from a universal coding-model ranking. The company has evidence sufficient for its own product choice, but it has not published enough for independent observers to reproduce every conclusion.

The timing raises another uncertainty. Replit’s September engineering post described strong production behavior from models associated with both companies. OpenAI became the preferred partner only weeks later.

That suggests the decision involved more than a sudden technical victory. The August Free Mode launch, ChatGPT distribution, model economics, and enterprise strategy likely reinforced the benchmark results.

There is also a discrepancy around the status of the Anthropic relationship. People familiar with the partnership reportedly disputed Masad’s initial statement that Replit was no longer working with Anthropic.

Replit’s later clarification aligns more closely with its public product evidence. Claude remains selectable, Anthropic remains listed as a subprocessor, and the company still describes itself as model-agnostic.

The safest interpretation is that Anthropic lost preferred status, not platform access. Anything stronger would exceed the currently verified evidence.

Users should also ask how Replit model choice appears in practice. A visible selector does not necessarily reveal every model invoked during a long agent session.

Replit’s architecture allows a core model to delegate to other processes. Intelligent routing can also change the system’s behavior as a task evolves.

Enterprise buyers need clearer answers about logging, data handling, retention, residency, and which providers process project context. Replit’s subprocessor disclosure identifies vendors, but procurement teams may need workload-level detail.

Development teams also need records that connect model changes with resulting code and decisions. A searchable engineering knowledge base can help preserve evaluation notes, incident findings, and architectural context across these shifts.

Security remains part of the evaluation rather than a feature that any foundation model solves alone. Replit’s agent habitat includes sandboxing, networking controls, identity, authorization, orchestration, and error recovery.

Masad has argued that agents need that surrounding environment to work safely. The point supports Replit’s platform strategy, but it also means a model comparison cannot carry the entire safety claim.

OpenAI’s default position should be judged through observable outcomes: completed applications, defect rates, recovery behavior, security incidents, latency, and total cost per successful task.

Until Replit publishes comparable data across those dimensions, the announcement remains a well-informed platform decision rather than a final technical verdict.

Three Signals Will Test Replit’s OpenAI Bet

The next phase will show whether the default improves completed work, survives enterprise scrutiny, and remains open to real customer choice.

The first signal is Replit’s next set of public agent evaluations. The most useful release would compare OpenAI and Anthropic models on the same tasks, inside the same harness, with total costs and failure rates included.

A comparison limited to final benchmark scores would leave the main economic claim unresolved. Readers need to know how many turns, retries, tool calls, and delegated tasks each successful build required.

If OpenAI maintains better completion rates at lower total cost, Replit’s decision will look technically grounded. If the advantage appears only under narrow application-generation tests, the broader preference will look more dependent on commercial alignment.

The second signal is customer behavior around Claude. Replit has said users can still choose Anthropic models, but it has not disclosed how frequently they do so after the default change.

A meaningful share of customers manually selecting Claude would show that model identity still matters. It could also push Replit to preserve visible controls and prevent routing from becoming an opaque substitute for choice.

Low switching rates would be harder to interpret. They might show satisfaction with OpenAI, but they could simply confirm that users rarely change defaults.

Replit should therefore pair selection rates with outcome measures. Retention, task completion, support requests, and repeat usage would reveal more than the model-selector count alone.

The third signal is Anthropic’s response across coding platforms. Anthropic can compete through model improvements, commercial terms, deeper integrations, or stronger evidence for complex repository work.

A new Claude release that wins Replit’s own evaluations would test the meaning of “model agnostic.” Replit would then face a choice between maintaining its preferred partnership and changing the default again.

If Replit switches readily when evidence changes, its neutrality claim becomes more credible. If OpenAI remains the default despite a clear disadvantage, commercial alignment will appear more influential than benchmark leadership.

Enterprise adoption will put additional pressure on that balance. Masad expects enterprise and government business to represent a majority of revenue in 2027.

Those customers often demand predictable suppliers and extensive controls. Some also prefer multiple approved model providers to reduce dependence on one vendor.

Replit can respond by keeping OpenAI as the standard while documenting where alternatives are available. It can also provide administrators with controls over approved providers, routing, data exposure, and audit records.

The company’s larger strategic test concerns trust. Replit wants users to avoid thinking about model selection, yet it also wants to describe itself as model-agnostic.

Those goals can coexist only if automatic decisions remain inspectable. Users do not need to choose a model for every turn, but they should understand which providers can process their work and why the platform changed course.

The Replit OpenAI default is best understood as a current preference, not a permanent settlement. It reflects OpenAI’s combination of price performance, distribution, and results inside Replit’s agent system.

Claude’s continued availability keeps competitive pressure inside the product. It also gives Replit an alternative if OpenAI’s economics or performance changes.

For developers and enterprise buyers, the immediate action is straightforward. Evaluate completed workflows rather than provider reputations, record which model handled critical work, and watch Replit’s next benchmark release closely.

The decisive evidence will not be another executive quote. It will be whether Replit publishes comparable outcomes, preserves meaningful model choice, and changes its routing when the underlying data changes.