Bengaluru: The availability of powerful, low-cost AI models from Chinese companies is healthy for the AI ecosystem and gives startups more choice rather than posing a threat, OpenAI’s global head of startups Marc Manara said.Startups should use different AI models for different tasks instead of depending on a single provider, he told TOI in an interview in Bengaluru on Thursday.His comments come as Chinese companies such as Moonshot AI, which recently launched Kimi K3, intensify competition by offering capable AI models at sharply lower prices.“We’re big supporters of open-weight models. We have released our own open-weight models,” Manara said.Open-weight models allow developers to download and adapt the parameters that determine how an AI system behaves. Unlike closed models, which are typically accessed only through a provider application or API, they can be customised for specific use cases.A single startup may use as many as five AI models within the same product, Manara said. It could deploy a frontier reasoning model for complex tasks, a smaller model for routine workloads and a fine-tuned open-weight model for other functions.“That’s honestly how companies should be built today,” he said.Manara also rejected suggestions that competition from Chinese and other open-weight models was the primary driver behind OpenAI’s recent price cuts.“I don’t think that’s really the motivating factor for price cuts,” he said, adding that lower prices enable startups to build more AI-powered products without costs becoming prohibitive.OpenAI recently reduced the price of GPT-5.6 Luna, its smallest new model, by 80%. Manara said cheaper models make it easier for startups to support free-tier users and other cost-sensitive workloads.Manara also cautioned companies against treating higher AI usage as a measure of productivity.AI services are generally priced by “tokens” — the small units of text an AI model processes or generates. Some organisations have begun ranking employees based on the number of tokens they consume.Such leaderboards create “perverse incentives”, Manara said, encouraging employees to use more AI simply to climb the rankings even when it adds little business value. He described the practice as “token maxxing”.Instead, companies should focus on “value maxxing” — measuring whether AI improves output, speeds up work and delivers tangible business results. They should also avoid using their most powerful and expensive model for every task, he added.India is among OpenAI’s top three startup markets globally, Manara said. The company currently has about five people working with startups in the country across business and technical roles.OpenAI is also looking to address demand for local inference in India, although Manara did not provide a timeline. Inference is the computing process through which a trained AI model processes a request and generates a response. Local inference would allow more of that processing to take place within India.The push comes amid calls for India to develop its own AI models and infrastructure. Sarvam AI co-founder Pratyush Kumar said at the company’s Epoch 2026 conference that India should use the best global AI models while continuing to build indigenous capabilities.Meanwhile, OpenAI’s Codex has crossed 10 million weekly active users globally. Its adoption in India has grown about 27-fold since the beginning of the year, Manara said.
