The frontier is now competing on cost per task, and that is what turns AI's threat to incumbent software into a live one.

The releases this week share one property that matters more than any benchmark: they are cheaper to run. Microsoft says its in-house MAI models cut costs up to 89% versus OpenAI models for comparable workloads; Anthropic's Opus 5 delivers close to Fable 5 performance at half the price; MiniMax open-sourced a coding model it prices at 8% of Claude Sonnet; DeepSeek is competing explicitly on deployment cost rather than raw capability. The frontier is no longer defined solely by what a model can do, but by how little it costs to do it at scale.

That shift is what makes the adoption evidence this week consequential rather than anecdotal. Curative rebuilt its own CRM with AI-assisted coding in roughly two months and cancelled a $600k-a-year Salesforce contract. One company is not a trend, but Wall Street analysts increasing sell recommendations on Salesforce and Adobe on AI-related fears suggests the market is pricing the same logic: when building custom software becomes cheap, the premium on licensed platforms narrows. The vendors are not passive here — Atlassian is embedding agents directly into Jira, ServiceNow is being rewarded for its AI narrative, and Salesforce itself just won a $1.6 billion Veterans Affairs contract to deploy AI agents. The contest is over who captures the value that cheaper models create.

Our reading is that the decisive variable for the next 18 months is not model capability but the cost curve underneath it, and the parties positioned to control that curve — the ones bundling agents into existing workflows, or supplying the inference hardware — are better placed than either pure-play labs or the incumbents whose pricing assumed software was expensive to replace.

What we track and how it's verified