What the AI inference price collapse really means
The AI inference price collapse is the rapid, aggressive reduction in API token prices for large models, as providers compete to turn high-end reasoning and agent workloads into low-margin infrastructure rather than premium software, reshaping how startups design products, plan margins, and choose between models. For years, AI inference costs made serious reasoning systems hard to sustain: long context windows, repeated tool calls, and large outputs could turn the token bill into the business model itself. Now, price cuts are sharp enough to reset assumptions. Tencent Cloud’s decision to cut DeepSeek-V4 prices by up to 97.5%, while keeping model capability unchanged, turns a flagship reasoning model into a mass-market option. At the same time, rival offers such as MiMo V2.5 Pro show that capable reasoning models can be priced for scale, not scarcity.
DeepSeek and MiMo turn reasoning models into a low-cost utility
On the reasoning side of the market, DeepSeek pricing and MiMo V2.5 Pro mark a clear shift toward model pricing competition as a core strategy. Tencent Cloud has announced that its DeepSeek-V4 series will see a maximum discount of 97.5%, without any stated change in model quality. MiMo V2.5 Pro is moving into the same buyer conversation by offering an API at about USD 1 (approx. RM4.60) per million input tokens and USD 3 (approx. RM13.80) per million output tokens for prompts up to 256,000 tokens. This brings heavy reasoning, coding, and agentic workloads into reach for smaller teams. Applications that plan, read files, write code, and loop over their own work no longer need to feel like experimental loss leaders. For founders, lower AI inference costs mean more freedom to extend context, support longer sessions, and delay hard usage caps while they search for product–market fit.
Google’s full‑stack advantage: cheap tokens as a strategic weapon
As API token prices fall, the advantage shifts from model size to control of the stack that runs those models. Google’s pitch with Gemini 3.5 Flash focuses on AI operational costs: if companies mix Flash with other models, they can cut huge token bills without giving up too much capability. Sundar Pichai has said that monthly usage of Google’s AI products has reached 3.2 quadrillion tokens, highlighting how quickly token volumes, and therefore AI inference costs, are rising. Analysts estimate that Google pays around 50% less, and possibly as much as 75% less, for internal AI compute because it owns chips, data centers, cloud, and the models themselves. By contrast, most AI providers rent this stack and pass on margins from chip vendors and cloud platforms. That structure gives Google room to undercut on token price while still defending its own margins.
How a race to the bottom reshapes startup economics
For startups, collapsing AI inference costs are not a curiosity; they rewrite the spreadsheet. A year ago, many teams avoided long-context reasoning models because AI operational costs could eat all the unit economics. Now, MiMo, DeepSeek, and peers are willing to cut prices aggressively to win production workloads. That has several effects. First, more experiments become viable: teams can route across multiple models, test caching strategies, and iterate agents without fearing a runaway token bill. Second, smaller, cheaper models move from backup to default for many workloads, with larger models reserved for narrow, high-value calls. Third, middleware vendors face pressure: when base API token prices drop, a routing layer must prove its worth through better observability, fallback logic, and governance rather than a thin pass-through fee. In this race to the bottom, the winners will be those who can turn cheaper tokens into better products, not those with the flashiest model benchmarks.