The rapid rise of capable open-source AI models is prompting enterprises to rethink their dependence on a handful of frontier models from companies such as OpenAI and Anthropic, with cost emerging as an equally important factor as performance in AI deployment decisions.
The latest trigger came last week when Chinese startup Moonshot AI launched its open-source model Kimi K3, which ranked close to leading proprietary models on several benchmarks, including coding, while trailing only marginally on overall performance. The model’s popularity was such that Moonshot suspended new subscriptions within days after demand overwhelmed its computing capacity.
While few enterprises see frontier models disappearing from their AI stack, companies are increasingly adopting a hybrid strategy that combines large proprietary models with open-source and smaller language models (SLM), reserving expensive frontier models only for complex reasoning tasks while shifting routine workloads to lower-cost alternatives.
“We’re seeing companies get more deliberate, though I wouldn’t frame it as a rush toward open-source or SLMs,” Harish Doddala, SVP and general manager at AI software delivery platform Harness, said. Earlier, enterprises largely debated which frontier model to standardise on. Today, discussions have shifted to identifying which workloads genuinely require frontier-level capabilities and where lower-cost models can deliver comparable outcomes.
The shift reflects growing scrutiny over AI spending as enterprises move from experimentation to production. Token consumption, once a secondary consideration, is increasingly becoming a key metric for technology budgets.
Infosys has already adopted what it describes as a cost-aware AI architecture. “We use the right models – large, medium and small – for the right task, optimise token usage and balance deterministic rules with predictive model calls. This allows AI to scale at optimal cost while delivering better ROI,” Mohammed Rafee Tarafdar, the company’s chief technology officer, told Fe.
Arun Balasubramanian, managing director, India and Saarc at Dynatrace, said enterprises are increasingly adopting tiered AI strategies that intelligently route workloads across models to optimise token consumption without compromising innovation.
Industry executives, however, caution against interpreting the trend as a retreat from frontier AI models. Instead, they see the market evolving towards multi-model architectures where different models serve different business requirements.
“As enterprises move from experimentation to execution, the question is becoming economic — how do we make AI scalable across thousands of workflows without costs becoming unpredictable,” Goutham Parcha, vice president, Pega Applications at Pegasystems, said. Smaller and open-source models are increasingly suited for repeatable enterprise tasks such as summarisation, classification, documentation and data extraction, while frontier models continue to be preferred for more sophisticated reasoning.
The transition is not without trade-offs. Running open-source models internally requires companies to manage infrastructure, model evaluation and security, potentially shifting costs from AI inference to operations.
“Open-source gives you the recipe for the model, but you have to manage everything yourself,” a co-founder of vibe-coding startup Rocket, said, adding that operational complexity could offset savings from lower token costs.
Even so, venture investors believe economics will continue to drive adoption. “The strongest open models today come from Chinese labs, so trust concerns remain. But the cost-quality trade-off is becoming too compelling to ignore,” Vardhan Dharnidharka, principal at Stellaris Venture Partners, said.
The trend is also creating opportunities for Indian IT services companies. HCLTech, which recently invested in sovereign AI startup Sarvam, sees growing demand for smaller, industry-specific AI models. Wipro has announced plans to build enterprise-grade small language models through its AI-Native Business & Platforms unit, while LTM has reported early traction for its specialised voice and language models across sectors such as aviation.