Skip to content

AI & Cloud · · 7 min read

The Price of Intelligence Keeps Halving. Your AI Budget Should Not Be About Tokens.

Three weeks after launching Gemini 3.6 Flash, Google released 3.7 Flash — a more capable model for coding and agents — at an introductory price of half what its predecessor cost per million tokens. In the same month Anthropic moved computer use, browser use, the Skills API and the Files API from preview to general availability. Put those two facts together and a pattern most AI budgets still ignore becomes clear: the intelligence is getting cheaper fast, and the parts that were never cheap are the parts that remain.

What actually happened in August

Model pricing did not drift down this month; it stepped down. Google describes 3.7 Flash as its most intelligent workhorse model yet for coding and agents, and launched it at half the per-token cost of the model it replaced — one that was itself only weeks old. Open-weight labs kept pace with their own large releases, and aggregators counted more new models in August than in any previous month.

Meanwhile the tooling around models matured. The capabilities that let a model operate a computer, drive a browser, read files and follow packaged skills are no longer experimental previews with caveats — they are supported platform features you can build a product on. The Gemini app also passed one billion monthly users, which is a reminder that your customers already use these systems every day and will expect your software to meet that standard.

Where the money in an AI project really goes

When we scope an AI feature for a Malaysian business, token spend is rarely in the top three costs — and it is falling while the others are not. The real cost lives in connecting the model to your data cleanly, defining what it may read and write, building the review queue where a person checks its output, logging every decision for audit, and handling the cases where it is wrong or unavailable.

That work does not get cheaper when a lab halves its prices. It is ordinary software engineering: permissions, validation, integrations, interfaces. If a proposal spends most of its pages on which model to use and one paragraph on how its answers will be checked, the estimate is upside down.

Design so the model is replaceable

If a better, cheaper model arrives every few weeks, the worst architectural decision you can make is to weld your product to one of them. We put the model behind a thin interface of our own — a single place where prompts, structured outputs and fallbacks are defined — so that swapping providers is a configuration change with an evaluation run, not a rewrite.

The evaluation set is the asset here. A few hundred real examples from your business, with the answers your team considers correct, lets you test a new model in an afternoon and take the price cut with confidence. Without it, every new release is a guess.

What to put in the budget instead

Plan for a modest, metered model spend with headroom, and expect that line to shrink. Spend the saved budget where it compounds: on the data pipeline that gives the model good context, on the interface where people review and correct its work, and on the test set that lets you upgrade safely.

For an SME this is good news. The capability that was out of reach two years ago is now a commodity input. The competitive edge is no longer access to a strong model — everyone has that — but how well the system around it fits the way your business actually works.

Key takeaways

  • Model prices stepped down again in August; token spend is now a minor and shrinking line in an AI project.
  • The durable costs are integration, permissions, review workflows and audit — plain software engineering.
  • Keep the model behind your own interface and maintain a real evaluation set so each new release is a cheap upgrade, not a rewrite.

Choose how to get started

Build with us

A system engineered around how you work.

Scoped in a proposal, designed as a prototype, delivered in milestones you can review.

Start a project

See the work

Systems already running in production.

Marketplaces, operations portals and mobile apps, with the case studies behind them.

View our work
WhatsApp