The new economics of AI: trading prestige for price
The shift from premium AI providers to open-source AI models and cheaper Anthropic alternative models is a market-wide reaction to mounting inference costs, as startups discover they can meet or exceed performance requirements while achieving dramatic AI cost reduction strategies that change the unit economics of their products. Anthropic has become a favourite among large enterprises, but its pricing is pressing smaller players. Microsoft AI CEO Mustafa Suleyman has been blunt, saying “Anthropic is extremely expensive, and I think many people are urgently looking for alternatives.” For cost-conscious founders, every model call hits the bottom line, especially in agent platforms and SaaS tools where usage grows quickly. This is prompting a closer look at affordable AI solutions such as DeepSeek and other open weights, where the promise is straightforward: similar capabilities, far lower bills, and less dependence on a single frontier vendor.
Lindy’s DeepSeek switch: saving millions without losing performance
The clearest signal of this shift comes from Lindy, an AI agent platform whose product runs models continuously on behalf of users. Founder Flo Crivello says the company has “switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models,” and reports that the move “saves us millions of $ and we’re seeing an increase in performance on many core use cases.” In earlier posts, he called inference Lindy’s “#1 cost by a lot (more than payroll)” and said cutting it by 2–5x would transform the business. DeepSeek V4-Pro is priced at USD 3.48 (approx. RM16.00) per million output tokens, and running one benchmark costs USD 1,071 (approx. RM4,930) compared to USD 4,811 (approx. RM22,130) for Claude Opus 4.7, making DeepSeek vs Anthropic a stark comparison on price.
Anthropic’s premium pricing hits a wall with smaller companies
Anthropic’s models remain highly regarded, but their premium status is turning into a barrier for startups and smaller enterprises. Crivello’s experience shows how a popular frontier model can become an operational drag when inference dominates the cost structure. Even Microsoft, one of Anthropic’s major enterprise customers, wants out of the current spending pattern. Suleyman said Microsoft’s goal is “to reduce and ultimately eliminate” what it pays Anthropic, underscoring that large buyers also feel the pressure. For startups, the trade-off is sharper: they must weigh the marginal gains of slightly stronger proprietary performance against very real cash burn. As alternatives improve, many find that Anthropic alternative models deliver “good enough” or even better results for their specific workflows, while freeing budget for engineering, distribution, and product development instead of cloud and API bills.
Open-source AI models mature into enterprise-ready options
DeepSeek’s rise illustrates a broader pattern: open-source AI models and open-weights systems are catching up on real-world tasks that matter to businesses. DeepSeek V4-Pro, for example, scores 1554 on the GDPval-AA benchmark for agentic real-world tasks, making it a leading open-weights model for agent scenarios at launch. DeepSeek itself says it trails the US frontier by about three to six months, but for many production use cases that lag does not impact customer value. Lindy evaluated other models such as GLM and Kimi K2.5, calling GLM-5.1 “incredible” and saying Kimi K2.5 nearly became its default, before concluding V4 was “way way better” for their needs. This kind of systematic benchmarking is making open and alternative models credible replacements in enterprise stacks, not experimental tools on the side.
Democratizing AI access and the future of model choice
Behind these switches is a larger market shift toward affordable AI solutions that widen access beyond the best-funded enterprises. Suleyman notes that outside developer circles, “most people are just not using this in their everyday life,” highlighting a large untapped market. To reach those users, providers and startups must keep costs low enough that AI feels like background infrastructure, not a luxury feature. That reality favours architectures built around Anthropic alternative models, open weights, and multi-model routing, where expensive proprietary systems are reserved for edge cases. Lindy still uses Claude internally under subsidized plans and may escalate to Claude Opus in rare failure scenarios, but calls that usage “marginal.” The direction is clear: the centre of gravity is moving from single-vendor dependence toward a mix of open and closed models, with price-performance, not brand, as the deciding factor.
