DeepSeek recently cut its API prices by 75% in a single move. It's the latest in a two-year pattern: the cost of running AI models keeps falling, sharply and often. If you've been holding off on AI investment because "it's still expensive," that reasoning has an expiry date — but the conclusion most people draw from falling prices is only half right.
The half that's right: the cost of raw AI capability — the model itself, the API call, the tokens processed — has been in freefall for two years and shows no sign of stopping. Competition between major labs, along with more efficient model architectures, keeps pushing prices down. For a Malaysian SME, the AI tool that looked financially out of reach eighteen months ago is very likely affordable now, and will be cheaper again in another eighteen.
The half that's wrong
Falling model costs do not mean falling AI project costs. This is the trap. The token price is a small fraction of what it actually costs an organisation to get real value from AI — the rest is integration, data cleanup, process redesign, and the human time to figure out what to actually build. A cheaper engine doesn't make the car free.
Analysts have flagged this gap explicitly: as models get cheaper, the economics of using AI to do real production work — not demos — look worse than the headline pricing suggests, because the surrounding cost of making AI reliable at scale hasn't fallen at the same rate the token price has.
Before your next AI budget conversation, separate two numbers: what the AI tool itself costs per month, and what it will cost in your team's time to make it actually useful (setup, training, workflow redesign). The second number is usually the one that determines whether the project succeeds — and it's the one falling model prices don't touch.
What this actually changes for Malaysian SMEs
Where falling costs genuinely matter: experimentation is now cheap enough that there's little excuse not to try. A pilot that would have cost a meaningful sum in API fees two years ago now costs very little. That lowers the bar for testing an idea before committing budget to a full rollout — which is exactly how AI adoption should happen anyway, rather than a single large procurement decision made on a vendor's roadmap slide.
Where it doesn't help: vendor lock-in, data governance, and the change-management work of getting a team to actually use a tool well. None of that gets cheaper because the underlying model does. If anything, cheaper AI means more tools competing for your team's attention, which makes the discipline of choosing carefully — rather than chasing the newest, cheapest option — more valuable, not less.
What This Means for You
Falling AI costs are a genuine reason to run more small experiments, faster. They are not a reason to expect AI adoption itself to get proportionally easier or cheaper — that part still depends on the same things it always has: a clear problem, a specific process, and people who understand both the tool and the work it's meant to improve.
Want the next one of these in your inbox?
Subscribe on the Insights page for weekly, practical perspectives on AI in Malaysia — no hype, no spam.