Microsoft launched a new AI coding model this week called MAI-Code-1-Flash. The launch itself is not the story. What matters is one line they added to the specification document: average token usage per task.
Tokens are the units of text an AI system reads and writes. Every token costs money. For most of AI's short commercial history, companies and vendors focused almost entirely on what a model could do, not what it cost to do it. That era is ending.
Microsoft's model completes the same coding tasks as Anthropic's comparable tool while using roughly a third of the tokens, and it scores higher on quality benchmarks. The cost gap is the actual product differentiator here, not the capability score.
This matters because real-world AI costs have surprised almost everyone who has deployed these tools at scale. Uber encouraged its engineers to use AI tools as much as possible and even ranked employee usage on internal leaderboards. The result: the company burned through its entire 2026 AI budget in four months. Its CTO went on the record in April saying, "I'm back to the drawing board because the budget I thought I would need is blown away already." Uber has since capped each employee at $1,500 per month per AI coding tool.
Salesforce tells a similar story. CEO Marc Benioff expects the company to spend around $300 million on AI usage fees this year, almost entirely for coding work. At the same time, Salesforce has frozen software engineering hiring. The money that would have gone to salaries is going to AI tokens instead, and the company is actively building systems to route simpler tasks to cheaper models to keep costs from spiraling.
These are not small or reckless companies. They are among the most sophisticated technology operators in the world. If they are struggling to manage AI costs, every other business deploying these tools faces the same problem, just at a smaller scale.
The independent benchmarking firm Artificial Analysis tracks this cost dimension across all major AI models. Their data shows that the top two AI models by capability, GPT-5.5 and Claude Opus 4.8, score within one point of each other on their composite intelligence ranking. But running the same benchmark suite costs around $3,357 on GPT-5.5 and around $4,685 on Claude Opus 4.8. Same output quality, 40% cost difference.
This is the number that should interest any business operator. If two tools produce the same result and one costs 40% more, the price gap is not a technical detail. It is a budget line.
The shift is also starting to reshape how companies think about their internal teams. Salesforce is hiring 1,000 to 2,000 salespeople to sell its AI products to clients, while holding the engineering headcount flat. The logic is that AI can absorb coding work, but humans are still needed to explain and sell the output. That is a meaningful structural change in how large companies are allocating labor, and it is being driven not by capability but by cost arithmetic.
For anyone running a business that uses AI tools, or is considering it, the question to ask vendors and internal teams has changed. It is no longer "what can this model do." It is "what does each completed task cost us, and how does that compare to the alternatives." The companies that will get the most value from AI in the next two years are the ones that manage this like any other operating cost: tracked, compared, and optimized.