The New Economics of Picking an AI Model
The shift from expensive frontier AI models to cheap AI model alternatives describes how cost-sensitive startups evaluate, compare, and replace premium systems with lower-priced open or semi-open options when the cost-to-performance ratio no longer justifies the premium. For many young companies, inference spend has quietly become the largest bill, overtaking payroll and cloud hosting. That pressure is forcing founders to ask whether paying for top-tier names still makes sense when newer, cheaper tools deliver comparable results. This is where AI cost comparison has turned from a procurement detail into a core strategic choice. Instead of chasing the absolute “best” benchmark score, teams now weigh every extra point of quality against its price in tokens and infrastructure work, and they are beginning to walk away from expensive contracts.
Lindy’s DeepSeek Switch: Millions Saved, Performance Gained
Nothing captures the mood better than Lindy, an AI agent platform whose product depends on continuous model calls. Founder and CEO Flo Crivello said his company has shifted 100% of production traffic away from Anthropic and over to DeepSeek V4, citing “millions of dollars” in savings and better performance on many core tasks. For a startup where inference was “#1 cost by a lot (more than payroll),” that is a turning point rather than a minor optimization. DeepSeek V4-Pro is priced at USD 3.48 (approx. RM16.00) per million output tokens, and running the Artificial Analysis Intelligence Index benchmark costs USD 1,071 (approx. RM4,930) versus USD 4,811 (approx. RM22,150) for Claude Opus 4.7. In this DeepSeek vs Anthropic comparison, the cheaper model is more than four times less expensive on that benchmark while delivering good enough or better results for Lindy’s agentic workloads.
Microsoft’s AI Chief Calls Anthropic “Extremely Expensive”
Cost concerns are not limited to startups. Microsoft AI CEO Mustafa Suleyman has openly called Anthropic “extremely expensive” and said many customers “are urgently looking for alternatives.” He added that Microsoft itself pays “a lot of money” to Anthropic and wants to “reduce and ultimately eliminate that cost.” For one of Anthropic’s most important enterprise buyers to talk this way signals a broader reckoning with how much premium AI access should cost. Suleyman’s comments sit alongside his ambition for Microsoft to become one of the “top four labs in the world,” building its own frontier models instead of relying on partners. The subtext is clear: if even the largest platforms feel squeezed, smaller firms chasing an affordable AI API will keep experimenting with open source AI models and lower-priced closed systems that give them more control over both spend and roadmap.
Startups Chase Value While Enterprises Pay for Relationships
A divide is emerging between how enterprises and startups think about AI purchasing. Enterprises still flock to Anthropic and other premium labs, often valuing deep account support, risk frameworks, and long-term contracts over raw cost. According to Flo Crivello, Anthropic will likely “be fine because of enterprise relationships,” even as his own company has marginalized it in production. Startups, by contrast, are benchmarking cheap AI model alternatives relentlessly. They test DeepSeek, GLM, and Kimi K2.5; they measure quality on real tasks and plug in whichever option offers the best cost-to-performance ratio at scale. For many of them, an affordable AI API is not a nice-to-have but a survival requirement. That is why open source AI models and newer contenders are gaining traction: they let founders control their margins instead of donating them to model providers.
Cost-to-Performance Ratio Becomes the Dominant Metric
Behind every switch to a cheaper model sits months of engineering and measurement. Lindy’s team spent a long time benchmarking alternatives, nearly selected Kimi K2.5, and still uses Anthropic internally when a subsidized plan makes sense. Crivello described the migration to DeepSeek V4 as “100x more work than we thought,” underlining that no founder takes this decision lightly. But once the numbers show a 2–5x reduction in inference cost with equal or better performance, the business case is overwhelming. DeepSeek acknowledges it trails the “US frontier” by three to six months, yet for many production agentic use cases that gap no longer matters. In this environment, the cost-to-performance ratio now outweighs brand prestige: teams prioritize total token spend, reliability on specific workflows, and the freedom to change providers again if the economics shift.
