A New Era of AI Cost Discipline
The current wave of AI adoption is entering a phase where companies must balance model performance with strict budget limits, driving demand for Anthropic alternatives such as DeepSeek and other cheap AI models that offer competitive quality at far lower inference costs. Startups that depend on continuous AI usage are discovering that premium offerings, while powerful, can turn inference into their largest expense. As usage grows from thousands to billions of calls, small pricing differences become existential issues for early-stage firms. This tension is reshaping AI cost comparison strategies: instead of defaulting to the most hyped model, teams are running structured benchmarks, weighing latency, quality, and price for their actual workloads. The result is a quiet but important shift in the AI market, where budget AI providers are gaining traction not by marketing, but by clearing a simple bar: being good enough while keeping the lights on.
Lindy’s DeepSeek Switch: Millions Saved, Performance Up
AI agent platform Lindy has become a flagship case in the DeepSeek vs Anthropic debate. Founder Flo Crivello announced that Lindy “switched 100% of traffic to DeepSeek V4, churning from Anthropic models” and reported millions of dollars in savings alongside better results on many core tasks. For Lindy, inference had been the number-one cost, exceeding payroll, so a 2–5x reduction was described as transformative. DeepSeek V4-Pro’s pricing of USD 3.48 (approx. RM16) per million output tokens underpins that appeal, especially when full benchmark runs cost USD 1,071 (approx. RM4,930) versus USD 4,811 (approx. RM22,130) for Claude Opus 4.7. Lindy’s team spent months evaluating GLM, Kimi K2.5, and other models before deciding V4 was “way way better” for their needs. Anthropic models remain in marginal internal use, but no longer sit on Lindy’s primary production path.
Anthropic’s Enterprise Strength Meets Startup Reality
Anthropic has emerged as a dominant choice for large enterprises, where reliability, safety tooling, and procurement relationships carry heavy weight. For these customers, higher prices can be justified by trust, integrations, and support. Smaller companies, however, experience the same price points very differently. When a startup’s product runs AI nonstop, per-token costs compound into massive recurring bills, and even well-funded teams begin to question whether frontier models are worth the burn. Lindy’s move highlights this divide: the company remains “a big fan of Anthropic” and still taps Claude for edge cases, yet found it unsustainable as a default. This pattern suggests Anthropic could remain strong with big enterprise contracts while ceding ground among cost-sensitive startups, where cheap AI models and open-weight systems are closing the quality gap fast enough to be compelling replacements in daily operations.
Microsoft’s Take: Anthropic Is ‘Extremely Expensive’
The pressure on premium AI pricing is not only coming from startups. Microsoft AI CEO Mustafa Suleyman described Anthropic as “extremely expensive” and said that “many people are urgently looking for alternatives.” He added that Microsoft itself “pays a lot of money to Anthropic” and that the company’s goal is to “reduce and ultimately eliminate that cost.” This public stance signals two things: first, that even deep-pocketed tech giants feel the strain of closed-source frontier model pricing; second, that major platforms may prefer to build in-house capabilities rather than depend indefinitely on external labs. Suleyman’s aim is for Microsoft AI to stand alongside Google DeepMind, OpenAI, and Anthropic as a top-tier lab. In that context, budget AI providers and open-weight models are not only startup tools; they are part of a broader strategic repositioning away from expensive dependencies.
Performance vs Price: How the AI Market Is Rebalancing
The emerging AI cost comparison story is not simply about “cheap vs premium” but about when a slightly weaker model is still more than good enough. DeepSeek itself admits it trails US frontier labs by 3–6 months, yet on agentic tasks, V4-Pro leads open-weight peers and performs strongly for production use cases like Lindy’s AI employees. For many applications, that level of performance at significantly lower cost beats chasing cutting-edge benchmarks. This dynamic is creating a layered market: enterprises with complex compliance needs may pay for top-tier closed models, while startups and mid-size firms adopt budget AI providers and open-source systems where trade-offs are acceptable. As more companies publish results from model migrations, the default assumption that expensive models are always best is eroding, replaced by a more pragmatic question: which model makes the product and the business viable at scale?
