ChatGPT is no longer OpenAI's most important product. Here's why.

OpenAI Explores Integrating Codex Features Into ChatGPT as Executives Weigh Customer Retention Strategies

The next phase of competition among leading artificial intelligence companies may center less on raw model performance and more on mechanisms that encourage continued use of specific platforms, according to one industry executive. For several years, OpenAI and Anthropic have released successive versions of their large language models with incremental gains in benchmark scores. Samuel Colvin, chief executive of the AI infrastructure startup Pydantic, said in an interview that the financial incentives for these companies are now changing.

"A year ago, what they cared about was revenue," Colvin stated. "Now, when one assumes they're both trying to IPO, their profit margin becomes really important." Training successive frontier models requires repeated capital outlays measured in billions of dollars, and rival laboratories have demonstrated the ability to replicate many performance characteristics within months. Colvin argued that this pattern makes it difficult to sustain high margins if competition remains limited to model quality alone.

Instead, he said, both OpenAI and Anthropic are directing resources toward products designed to increase the cost for customers to switch providers. "They are doing their very best to find ways of locking people in that are not related to model quality," Colvin said. "That's where I think Claude Code and Codex and all that work is coming from." These coding assistants allow developers to process large volumes of tokens on extended projects, producing higher usage levels than typical conversational interactions.

The resulting codebases can become extensive enough that subsequent maintenance, debugging, or modification requires continued access to the same tools that generated the original output. Colvin suggested this dynamic could create practical dependence even if alternative models achieve comparable technical results. Anthropic has introduced features under the Cowork label that extend its coding tools into broader workflow environments, while OpenAI has indicated plans to incorporate elements of its Codex system directly into the ChatGPT interface.

Enterprise customers have responded by developing internal systems that reduce reliance on any single provider. Walmart operates an internal coding assistant called Code Puppy that routes requests across models supplied by OpenAI, Anthropic, Google, and additional vendors. The system is intended to give the company greater oversight of its generated code and the ability to adjust usage based on pricing or performance changes across suppliers.

This pattern illustrates a divergence between the objectives of model developers seeking recurring high-margin usage and the preferences of large organizations that prioritize flexibility in model selection and control over their software assets. Colvin's comments reflect observations drawn from his company's work supplying validation and data-handling tools to teams that build on these platforms. The companies named have not issued public statements confirming the strategic emphasis described.

The discussion occurs against a backdrop in which usage-based pricing for token consumption remains the primary revenue mechanism for both OpenAI and Anthropic. Coding workloads that involve repeated generation, testing, and iteration can produce substantially larger token counts per project than one-off queries. Whether integration of coding capabilities into general-purpose chat interfaces will measurably alter customer switching costs remains subject to future adoption patterns and competitive responses from other providers.

Colvin noted that the shift in emphasis he described is still emerging and could be influenced by regulatory developments, changes in capital availability, or technical breakthroughs that alter the cost of training. No public financial filings from either company have yet detailed the contribution of coding-specific products to overall margins. The observations therefore represent one participant's assessment of directional trends rather than confirmed corporate strategy documents.

Industry participants continue to monitor how enterprise procurement teams balance the productivity gains reported from AI coding tools against the operational risks associated with concentrated vendor relationships. Walmart's approach of maintaining model portability provides one documented example of an alternative procurement model. Additional organizations have disclosed similar multi-model architectures in technical presentations and procurement guidelines issued over the past year.

The outcome of these competing priorities will depend on measurable factors including token pricing trends, the portability of generated code across different model families, and the pace at which new entrants offer comparable capabilities without the same integration requirements. Colvin's remarks supply one data point in an ongoing discussion among developers, investors, and enterprise technology leaders.