A Price Shock That Defies the Compute Crunch
While many providers react to rising compute demand by lifting rates, DeepSeek has moved in the opposite direction. The company has made its DeepSeek V4 Pro promotion permanent, resetting its AI API pricing to just 6 Renminbi per 1,000,000 output tokens—only 25% of its original standard rate. Instead of a temporary discount, this becomes the new baseline, turning what was a marketing campaign into a structural cost reset for developers. In stark contrast, OpenAI’s GPT 5.5 is listed at 30 USD (approx. RM138) per 1,000,000 output tokens, with a premium tier at 180 USD (approx. RM828) per 1,000,000 output tokens, making DeepSeek V4 Pro dozens to hundreds of times cheaper on a per-token basis. By reframing inference costs as a bargain rather than a bottleneck, DeepSeek positions its model as a default choice for cost-sensitive builders.
MiMo V2.5 Pro Pushes Reasoning Models Into Commodity Territory
MiMo V2.5 Pro shows how quickly AI API pricing is compressing even for complex reasoning workloads. According to its API pricing page, MiMo V2.5 Pro is offered at about 1 USD (approx. RM4.60) per 1,000,000 input tokens and 3 USD (approx. RM13.80) per 1,000,000 output tokens for prompts up to 256,000 tokens, with higher long-context pricing above that. These levels put MiMo directly into the same buying conversation as DeepSeek V4 Pro, especially for applications that rely on long contexts, frequent tool calls and large outputs. Historically, such workloads—coding agents, research assistants, legal review tools—were often uneconomic at scale because the token bill dominated unit economics. MiMo’s pricing makes those use cases far more approachable for startups, signaling that capable reasoning and agentic models are being priced like infrastructure utilities instead of premium, scarcity-priced software add‑ons.
From Capability Premium to Inference Cost Warfare
The combination of DeepSeek V4 Pro and MiMo V2.5 Pro has accelerated a shift from capability-driven positioning to direct competition on inference costs. DeepSeek’s permanent cut to one quarter of its original rate, coupled with MiMo’s low per-million-token pricing, puts pressure on other model vendors that previously justified higher prices with marginal quality gains or brand strength. For startups, this alters the decision matrix: model selection increasingly starts from a cost ceiling, with teams asking whether incremental accuracy or features are worth several times the token price. At the same time, builders are learning that headline AI API pricing is just one piece of the equation. Cache strategies, context-length tiers, verbosity controls and tool-calling efficiency can multiply or shrink real spend. As more models cluster in a similar capability band, sustained price undercutting risks turning base models into near-commodity infrastructure.
How Startup AI Economics Are Being Quietly Rewritten
For founders, falling inference costs reshape product and fundraising assumptions. When a reasoning-capable model can be accessed at 6 Renminbi per 1,000,000 output tokens or a few USD per 1,000,000 tokens, API bills become a smaller share of overall deployment expenses. That frees teams to run longer sessions, accept richer context windows and delay aggressive throttling, all of which improve product-market fit testing. It also makes experimentation with multi-model routing more realistic, because the marginal cost of trying a second or third provider per request is lower. Most startups do not need to train their own foundation model; they need reliable access, predictable bills and enough quality to deliver value. As MiMo, DeepSeek and other labs compete on price, startups gain negotiating leverage and the ability to design business models around sustainable per-user margins rather than fear of runaway token invoices.
Beyond the Race to the Bottom: Differentiation and Consolidation Ahead
An extended race-to-the-bottom on AI API pricing will not leave the market unchanged. Middleware and routing platforms already face tougher questions about their value when base model prices fall rapidly. If an aggregator simply passes requests through with a markup, cheaper underlying models can compress its margins. To justify their place, these layers must deliver better routing, observability, governance and billing control—functions that save time or reduce risk, not just relay traffic. For model providers, persistent price pressure is likely to drive consolidation and force differentiation beyond raw per-token rates. Latency guarantees, uptime, data policies, tooling ecosystems and enterprise support will matter as much as headline dollars per million tokens. The winners among startups will mirror this logic: those that build flexible, multi-model architectures and deeply understand their real inference costs will be best positioned as each new price cut resets the competitive landscape.