What Falling Inference Costs Really Mean
Falling AI inference costs describe the rapid decline in the price of running large models through AI APIs, where providers bill developers based on tokens processed instead of flat software licenses, which changes how teams budget, test, and scale AI features because usage-based costs are now low enough that experiments, long sessions, and complex agent workflows can be run without making the token bill the entire business model. This shift is turning AI API pricing into a race to the bottom, especially for reasoning-heavy models that once felt too expensive for everyday use. Instead of treating advanced models as premium luxuries, providers are starting to price them like core infrastructure. For developers, this removes a major financial barrier: the ability to prototype, iterate, and keep agents running longer without fearing runaway invoices, even as models gain longer context windows and more tools.
DeepSeek’s Permanent Price Cut and Its Signal
DeepSeek V4 Pro has turned a temporary promotion into a permanent reset, cutting its standard rate to 25% of the original price and charging 6 Renminbi for every 1,000,000 output tokens. Consumers still use the platform for free, while enterprise developers see their development and inference costs drop sharply. This move puts DeepSeek pricing far below some headline models: the same source notes GPT 5.5 at USD 30 (approx. RM138) per 1,000,000 output tokens, with a premium tier at USD 180 (approx. RM828) for 1,000,000 output tokens. The gap is so wide that inference costs falling is no longer a future trend; it is here. Capable models are being priced aggressively to win developer mindshare. For startups, using a powerful model no longer means betting the runway on token spend; it becomes a line item that can be optimized like any other cloud cost.
MiMo and the New Competition in Reasoning Models
MiMo V2.5 Pro shows how quickly AI model competition is shifting from accuracy bragging rights to AI API pricing. According to Xiaomi’s MiMo API pricing page, MiMo V2.5 Pro is listed at about USD 1 (approx. RM5) per million input tokens and USD 3 (approx. RM14) per million output tokens for prompts up to 256,000 tokens. These terms place MiMo directly in the same buying conversation as DeepSeek V4 Pro, especially for long-context, tool-heavy agents. Reasoning-heavy workflows—like code-writing agents that read files, plan, write, check, and loop—used to explode budgets. Now they start to look like manageable infrastructure costs. Startups gain a wider menu of tradeoffs: they can choose between slightly different capabilities, latencies, or tooling while keeping inference costs under control, and they can run more A/B tests, multi-model routing, and complex flows without turning every prompt into a board-level expense.
Google’s Full-Stack Response to the Price War
As AI API pricing compresses, big platforms are shifting the battlefield to infrastructure and total cost of ownership. Google’s Gemini 3.5 Flash aims to compete on cost and speed, not only frontier capabilities. Sundar Pichai highlighted that monthly usage of Google’s AI products reached 3.2 quadrillion tokens and argued that companies mixing Flash with other frontier models could save more than USD 1 billion (approx. RM4,600,000,000) a year. Analysts quoted by Business Insider estimate Google pays around 50% less—and possibly up to 75% less—for its own AI compute by owning chips, data centers, cloud, and models end to end. This full-stack control lets it undercut rivals whose inference costs depend on external cloud margins. The message for enterprises and startups is clear: pick models not only on raw IQ but on long-run economics, latency, uptime, and how well they fit into existing cloud contracts.
New Business Models and How Builders Should Respond
When inference costs fall this far, the economics of what you can build changes. Products that used to require strict usage caps—AI research assistants, coding copilots, legal review tools, data-cleaning agents—can now offer longer sessions, richer context, and more aggressive automation without every extra token destroying margin. Some teams can offer core AI features free or bundle them into existing plans, using higher tiers only for edge cases like ultra-long contexts or heavy tool use. Middleware platforms also face a new test: routing layers must offer better observability, fallback logic, and governance to justify their margins when base model prices are compressed. For startups, the playbook is to treat AI models as interchangeable infrastructure, track token economics early, run multi-model evaluations, and design products around sustained agent usage rather than single-shot prompts. In a world of cheap tokens, iteration speed and product fit matter more than perfect models.