On 24 July, Anthropic released Claude Opus 5 at the same price as its predecessor, with performance that closes in on its flagship model for everyday work tasks. Three days earlier, Microsoft signed a multi-billion dollar deal to rent Mistral's compute capacity and distribute its models through Azure and Copilot Studio. A different kind of signal, around the same time: a major consulting firm launched a line of pre-configured AI agents aimed at mid-sized companies. Three announcements, one movement: the price of access to AI keeps falling, across the whole stack, at the same pace.
AI adoption in the enterprise is the organisational work that turns technical access to artificial intelligence tools into real professional use: scoped use cases, integrated company data, access governance and change management measured over time, a workload whose cost does not follow the price of the models down.
Three announcements, one movement
Taken separately, these look like ordinary product news. Put together, they trace a trend that has held for two years: the cost of accessing a capable language model keeps falling, year over year.
Compute first. Microsoft is building its own data centres at speed, and still chooses to rent someone else's on top. Compute capacity is turning into a resource that gets negotiated and re-let, including disconnected environments built for regulated industries.
Models next. Opus 5 holds the previous generation's price while closing the gap to the high end. What required an experimentation budget two years ago now fits inside a standard subscription.
Agents last. A consulting firm whose business model has long rested on billable hours now sells packaged agents to mid-market companies. That is the most telling signal of the three: the value of basic technical configuration is getting cheaper too.
Why doesn't falling AI prices solve adoption?
Because the price of a model and the cost of adoption are two different things. The first is negotiated in dollars per million tokens and drops every year. The second is measured in redefined roles, rebuilt processes and teams trained for professional use, and no price cut shortens that work.
Sourced figures back this up. S&P Global finds that 46% of generative AI pilots get abandoned before reaching production. MIT puts at 95% the share of pilots that never reach measurable return on investment. And according to Deloitte, only 21% of companies deploying AI agents have governance rated as mature. None of these numbers is about model pricing. All of them describe an organisational problem.
- Redefining roles: deciding what an agent handles alone, what a human validates, what stays off-limits.
- Integrating company data: connecting real systems, not running demos on public data.
- Training teams for professional use: moving from an isolated personal prompt to a scoped, traceable, repeatable use case.
- Installing governance: monitoring agents in production, not auditing them once a year.
None of these four workstreams gets smaller because a model costs half as much. They take the same amount of leadership, IT and business-team time, whether Claude Opus 5 costs 5 dollars or 10 dollars per million tokens.
The regulatory calendar isn't giving anyone extra time
While prices fall, the regulatory calendar keeps moving. On 20 July, France's data protection authority, the CNIL, published an exploratory note with CIANum on agentic AI, systems that act on a user's behalf. On 2 August, the transparency obligations under Article 50 of the EU AI Act and the European Commission's sanctioning powers become applicable for prohibited practices, with fines that can reach 35 million euros or 7% of global revenue.
The easy read is to treat price deflation and regulatory pressure as separate stories. They are not. A company deploying cheaper agents without governance in place accumulates the same risk, just with more agents in circulation. The cost of governing does not fall either, and Gartner expects roughly 40% of agentic AI projects to be cancelled by 2027, largely on cost and unclear value grounds rather than on the technology itself.
What should a company do now?
The natural reaction to cheaper technology is budgetary: cut the envelope, buy off the shelf, treat the AI question as settled once the tool is installed. That is the mistake taking shape. A cheap, packaged AI agent in production remains, without proper scoping, a system nobody is really supervising, and a cancelled project a year from now costs more than the model ever saved.
The better move runs the other way. Use the budget freed up by falling prices to fund what does not get cheaper: mapping use cases, governing agent access, training teams. That is an organisational project, not a procurement one, and it does not scale down just because the underlying model got a discount.
Before adding one more agent to the stack, a company benefits from knowing where it actually stands on these four fronts. Koneetiv's AI maturity assessment measures technical foundations and governance in a few minutes, the two dimensions that decide whether an agent deployment holds up once the launch price settles.