Why Anthropic’s Pricing Is Triggering an Enterprise Rethink
Enterprise AI model migration refers to companies switching from one large language model provider to another, or to open-source AI models, in order to improve the balance between performance, cost per token, and operational control as AI usage grows from experiments into large-scale, business-critical workloads. That shift is now squarely focused on Anthropic pricing costs. Microsoft AI CEO Mustafa Suleyman has called Anthropic “extremely expensive” and said “many people are urgently looking for alternatives,” signalling that even major buyers see the bill as unsustainable. His comment that Microsoft’s goal is to “reduce and ultimately eliminate” what it pays Anthropic highlights a wider pattern: once AI moves from pilot projects to constant, high-volume use, enterprise AI expenses can balloon. As customers start benchmarking price and performance more rigorously, Anthropic’s premium positioning is turning from selling point to structural risk.
DeepSeek vs Anthropic: When Price Gaps Become Business Model Problems
The clearest proof of a changing market is the DeepSeek vs Anthropic story. Flo Crivello, CEO of AI agent platform Lindy, said his company has “switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models” and that the move “saves us millions of $ and we’re seeing an increase in performance on many core use cases.” For a product that runs models continuously, inference is the largest cost line, larger than payroll. 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. RM4,920), compared with USD 4,811 (approx. RM22,100) for Claude Opus 4.7. When requests reach billions per month, this kind of 4x cost difference turns into transformative savings and undercuts the logic of paying for the highest-priced frontier option.
Smaller Companies Turn to Open Source AI Models
While large enterprises negotiated premium contracts with Anthropic, smaller companies have been more willing to experiment with open source AI models and upstart providers. Lindy’s switch to DeepSeek V4-Pro came after months of systematic evaluation of several open or open-weights options, including Kimi K2.5 and GLM-5.1, which its CEO described as “incredible.” For these firms, the priority is not brand prestige but a workable cost-performance envelope for agentic workloads, such as AI employees handling real tasks for users. The pattern is emerging: start with a leading closed model for speed and reliability, then migrate once cheaper AI model alternatives reach acceptable quality. The engineering lift can be large — Lindy described the migration as “100x more work than we thought” — but ongoing inference savings and performance improvements make the pain worthwhile for many product teams with tight budgets.
Anthropic’s Premium Strategy and Its Growing Vulnerability
Anthropic has strong traction with enterprises and is widely seen as one of the three most important AI labs, alongside OpenAI and Google DeepMind. Its models command premium pricing based on safety, reliability, and high benchmark scores. But that premium position creates vulnerability as soon as credible alternatives narrow the performance gap while undercutting prices. Mustafa Suleyman’s public goal for Microsoft to “become one of the top four labs in the world” makes clear that big buyers do not want permanent dependency on expensive third-party models. At the same time, DeepSeek claims to trail the US frontier by only 3–6 months on capability, which is short enough for many real-world agentic use cases. Together, these forces pressure Anthropic to prove that its higher enterprise AI expenses are justified by meaningfully better outcomes, not just brand and early-mover advantage.
The New Cost-Performance Playbook for Enterprise AI
The emerging playbook for enterprises is to treat AI model selection as a cost-performance optimization, not a one-way platform bet. In practice, that means benchmarking multiple AI model alternatives, including open source AI models, measuring token-level costs against success on real tasks, and planning for multi-model architectures that can route traffic dynamically. Lindy’s experience shows that once a cheaper model like DeepSeek meets or exceeds performance on core use cases, financial gravity takes over. For hyperscalers like Microsoft, even deeper dynamics apply: if a partner’s models are “extremely expensive,” it becomes attractive to build competing models internally and “reduce and ultimately eliminate” that external spend. As AI workloads scale, the winners are likely to be providers that balance strong performance with sustainable pricing and give customers room to swap models without betting the company on any single vendor.
