Blitz Bureau
NEW DELHI: Microsoft has reportedly begun cancelling most of its direct Claude Code licences, according to The Verge, instead moving engineers toward using GitHub Copilot CLI.
According to a Fortune report, that comes just six months after the firm first opened up access to Claude Code, encouraging thousands of its developers, project managers, designers, and other employees to experiment with coding. The tech became popular fast. Perhaps too popular.
The scale at which employees use it is now prompting the firm to reverse course on a tool its own engineers had come to rely on. Cancelling Claude Code licences won’t affect Microsoft’s Foundry deal, which includes investing up to $5 billion in Anthropic and giving Foundry customers access to Claude models, as well as Anthropic’s $30 billion commitment to purchase Azure compute capacity, according to The Verge.
Microsoft isn’t the only company scaling back its internal AI use. Uber’s CTO Praveen Neppalli Naga told The Information in April that the firm had already burnt through its entire 2026 AI coding tools budget in just four months. That comes after the company had actively incentivised adoption through internal leader boards ranking teams by AI tool usage.
Firms today are pushing employees to use as much AI as possible to squeeze out the technology’s productivity gains. But that pressure is leading to cracks, and those cracks may be irreparable.
The reports may throw cold water on the bets tech’s biggest firms have placed on the technology. While some cling to the promise of an AI “renaissance” or “revolution,” the cost of adoption is proving a stubborn bottleneck.
These developments also suggest that the economics of replacing or augmenting human labour with AI may be more complicated than some early forecasts originally implied. That echoes what Bryan Catanzaro, vice president of applied deep learning at Nvidia, recently said in an interview with Axios.
“For my team, the cost of compute is far beyond the costs of the employees,” he said.
Anthropic didn’t immediately respond to Fortune’s request for comment. Microsoft didn’t provide a comment.
An emerging AI paradox: cheaper tokens, bigger bills
Uber and Microsoft aren’t the only firms pushing employees to use as much AI as possible. Like at Uber, a Meta employee crafted a leader board, fittingly named “Claudeonomics,” after Anthropic’s AI model, to track which workers are using the most AI. Amazon is pushing its employees to “toxenmaxx,” or use as many AI tokens as possible (the basic building blocks of AI compute).
But with a token-based pricing system, the work gets more expensive with more use and better efficiency. Goldman Sachs recently forecasted that agentic AI could drive a 24-fold increase in token consumption by 2030 as consumers and enterprises adopt AI agents, reaching a staggering 120 quadrillion tokens per month. As businesses turn to AI agents to boost productivity, aggregate costs could rise sharply even if the price of each token falls.
But as consumption increases, the cost of individual AI tokens is expected to fall sharply. A recent report from research firm Gartner found that by 2030, inference on a one-trillion-parameter LLM — in simple terms, a highly sophisticated AI model — will cost AI firms nearly 90 per cent less than it did in 2025.
Even so, Gartner predicted that cheaper tokens won’t translate to cheaper enterprise AI because agentic models require far more tokens per task than standard models, increased consumption can outpace falling unit costs, and AI providers won’t fully pass through lower costs to consumers. In turn, inference costs are likely to push higher.
“Chief Product Officers (CPOs) should not confuse the deflation of commodity tokens with the democratisation of frontier reasoning,” Gartner senior director analyst Will Sommer warned in a statement.
That reality may complicate the grand plans some firms have for deploying AI agents. Nvidia CEO Jensen Huang recently said he thinks 100 AI agents will one day work alongside every employee at his company.
Huang is part of a broader wave of CEOs touting an agentic future in which digital workers operate across the enterprise. But if token consumption rises faster than unit costs fall, that future could come with a much heavier bill than executives expect.
Time to pause?
Anthropic says its Claude AI is increasingly helping create new AI systems, a trend the company believes could speed up AI development. In a blog post, the company said the trend could eventually lead to “recursive self-improvement”, where an AI system becomes capable of designing, building and training its own successor with little or no human involvement.
The company stressed that such a stage has not yet arrived and may never fully materialise, but argued that governments, regulators and society need to start preparing for it now. “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for,” the company wrote in the research paper, warning that AI development appears to be speeding up rather than slowing down.
According to Anthropic, both public benchmarks and its own internal data show that AI is already helping engineers and researchers work faster. While these advances could unlock major advances in areas such as healthcare, science and productivity, the company says it also raises questions about how humans will maintain oversight as AI systems become more capable.
How is AI improving faster than expected?
Anthropic outlined how the shift has unfolded over the past few years. Initially, engineers wrote code manually. Then chatbots began assisting with small coding tasks. That evolved into coding agents capable of writing and editing files independently.
Today’s AI agents can run code, perform tasks on their own and even delegate work to other agents. The next logical step, Anthropic argues, is AI systems that help build and train future AI models.
The company also points to public benchmarks as evidence that AI capabilities are advancing rapidly. Anthropic said the amount of work AI systems can reliably complete has been growing quickly, with task duration doubling roughly every four months, compared with an earlier trend of every seven months.
In March 2024, Claude Opus 3 could handle software engineering tasks that took humans around four minutes to complete. A year later, Claude Sonnet 3.7 could manage tasks lasting around 90 minutes. By 2026, Claude Opus 4.6 was reportedly capable of handling tasks requiring roughly 12 hours of human effort.
According to Anthropic, Claude now writes most of the code used to build the company’s AI systems. As a result, engineers are able to get much more work done, with the average engineer producing around eight times more code each day than in 2024.
Claude is not just getting better at coding. Anthropic says it is also improving at research tasks. In one internal project, Claude-powered agents solved nearly all of a key research challenge, while its success rate on difficult coding tasks rose to 76 per cent in May 2026, up significantly in just six months.
Looking ahead, Anthropic outlined three possible futures. Progress of improvement could slow down, AI could continue delivering major productivity gains while humans remain in control, or AI systems could eventually begin building their own successors. The company said the second scenario appears the most likely based on current evidence.
Despite these improvements, Anthropic says humans still have an edge when it comes to big-picture thinking, making strategic decisions and deciding which problems are worth solving.
However, at the same time, Anthropic warns that the challenge is no longer just building more capable AI systems. Instead, the bottleneck is increasingly becoming human oversight, review and validation.
If AI starts advancing faster than society can safely manage, the company says there should be a way for governments and leading AI firms to coordinate a temporary slowdown in the development of the most advanced AI models.
However, it cautions that any pause would need to be coordinated globally, as a unilateral slowdown by one company could simply allow less cautious competitors to move ahead.