The important correction to the “companies are abandoning AI” framing is that they are not. Microsoft is still expanding access to Anthropic models through Microsoft 365 Copilot, while Uber and Amazon continue to use AI internally. What is changing is the procurement model: broad, lightly governed access to powerful agents is being replaced by metering, limits, and tool consolidation.
The details reported by The Verge, the Financial Times, and follow-up coverage from TechRadar, Tom’s Hardware, and India Today describe separate companies reaching that conclusion through different failures. The shared lesson for Windows administrators and engineering leaders is blunt: an AI coding rollout without budget controls is a production cost-management project waiting to happen.
Microsoft’s Claude Code change is a control decision, not a Claude exit
Microsoft reportedly told many internal Claude Code users to move to GitHub Copilot CLI by June 30, 2026. TechRadar and Windows Central, citing reporting from The Verge, said the change applied to most direct Claude Code licenses in the company’s Experiences and Devices organization. Microsoft has not publicly released the full internal memo, the precise number of licenses removed, or a cost comparison between Claude Code and Copilot CLI.
That missing disclosure matters. The reporting does not establish that Claude Code was technically inadequate, nor that GitHub Copilot CLI is inherently cheaper for every task. It establishes that Microsoft wants developers using a tool it can shape with GitHub around its repositories, security requirements, workflows, and internal billing arrangements.
The distinction is especially relevant because Microsoft has continued to put Anthropic technology in customer-facing products. Microsoft’s own Microsoft 365 Copilot release notes say Claude became available in Copilot Chat in June, and Claude models were already being integrated across selected Copilot experiences, including Copilot Studio and Microsoft 365 applications. In other words, Microsoft is separating model access from the developer tool that consumes the model.
For IT departments, that is a more realistic pattern than a single-vendor story. A company can permit Claude models in a governed Microsoft 365 Copilot environment while removing an independently procured command-line agent that has different permissions, usage behavior, telemetry, and invoices. The model may remain available; the spending and governance path changes.
Microsoft’s later internal token guidance reinforces the point. TechRadar reported in August that employees were told their divisions had AI token-budget targets and could inspect their individual AI spending. The reported guidance said many engineers were consuming hundreds to several thousand dollars in tokens per month, while Microsoft was emphasizing business value rather than raw token usage.
That is a notable shift from the early enterprise AI posture, where adoption itself was often treated as proof of progress. Token consumption is not a productivity metric. It is an input cost, and the highest usage figure can describe either a breakthrough workflow or an inefficient agent repeatedly reading a repository, generating code, and repairing its own mistakes.
Uber’s budget problem was created by scale and variable consumption
Uber provides the clearest warning about what happens when a coding assistant moves from pilot project to default engineering tool. India Today reported that Chief Technology Officer Praveen Neppalli Naga said Uber had used its full 2026 AI coding budget by April, four months into the year, after broad internal adoption of Claude Code and related tools.
The company has not publicly disclosed the amount of that annual budget, so claims that attempt to translate the overrun into a share of Uber’s total research-and-development spending go beyond the available record. The meaningful fact is the timing: Uber’s allocated AI coding budget was reportedly exhausted before the first third of the year had ended.
The reason lies in the difference between a conventional software seat and an agent. A Visual Studio, Microsoft 365, or endpoint-management license is generally predictable once the number of users is known. A metered AI agent can be used briefly for an autocomplete-style suggestion or continuously for planning a change, reading thousands of files, modifying code, invoking tools, running tests, receiving errors, and iterating.
The latter workflow is valuable precisely because it can work across many steps. It is also expensive because every step may use input, output, cached, or reasoning tokens, depending on the provider and plan. A developer running multiple agents against a large codebase is no longer consuming a simple per-user service. They are triggering a compute workload.
Reports said Uber subsequently imposed a monthly limit for individual agentic coding tools and provided internal visibility into consumption, with exceptions available. The reported $1,500 per-person cap has not been publicly confirmed by Uber, but the broader response—limits plus dashboards—is the part every organization should take seriously.
A cap is not merely a finance restriction. It forces a design decision about which work deserves an expensive autonomous run. Repository-wide refactoring, incident investigation, migration planning, and test-generation may clear that bar. Asking an agent to repeatedly solve small tasks that a developer can finish in minutes may not.
Amazon’s $1.8 million overrun exposes the missing guardrail
The Financial Times reported that an internal Amazon project using Claude Sonnet accumulated about $1.8 million in costs while attempting to match author information to product listings. Tom’s Hardware and TechRadar independently relayed the Financial Times reporting, describing the figure as an 860% overrun that went undetected for roughly five months.
Because the underlying documentation has not been published, the project’s exact token volume, model configuration, automation design, and approval chain remain unknown. What Amazon did tell the Financial Times, according to the report, was that the examples reflected experimentation and learning rather than normal operating practice.
That response is reasonable as far as it goes, but it does not reduce the operational significance. A basic data-matching task is exactly the type of workload that can appear safe to automate: repetitive, measurable, and apparently bounded. The Amazon account shows how it can become open-ended when agents encounter edge cases and are allowed to continue generating work without a hard spend stop.
The reported case also included other overruns, including an extra $541,000 associated with a financial-auditing tool. The irony is instructive. Traditional cost controls are designed around known consumption units: seats, virtual machines, storage, API calls, or transactions. Agentic systems complicate that model by changing their own work volume in response to what they find.
A budget alert that arrives after the month closes is therefore inadequate. So is an executive dashboard that reports aggregate AI usage but cannot identify which project, repository, workflow, model, and user drove the bill. The question is not whether an agent is “enabled.” It is whether an organization can stop it when its unit economics no longer make sense.
GitHub Copilot CLI has the same metering problem
Moving from Claude Code to GitHub Copilot CLI does not make usage-based cost disappear. GitHub’s documentation explicitly says Copilot CLI can draw from a user’s AI-credit entitlement or, in organization-owned repositories, bill usage directly to the organization when the relevant policy is enabled.
That makes Microsoft’s internal migration more significant than a vendor preference. GitHub Copilot CLI gives Microsoft a route to centralize the policy and billing surface inside a product it owns, but it still requires controls. An organization that treats the migration as a way to restore unlimited usage will reproduce the same problem under a different invoice.
The research evidence is also more qualified than AI marketing implies. A July preprint studying Microsoft’s early 2026 rollout of Claude Code and GitHub Copilot CLI found adoption associated with roughly 24% more merged pull requests during the observation period. The authors cautioned that pull requests are a proxy for output, not a measure of business value, code quality, maintenance burden, security, or customer impact.
That limitation should shape how enterprises calculate return on investment. More pull requests may represent useful throughput. They may also represent smaller changes, agent-generated churn, or additional review work transferred to other engineers. Cost per accepted change is useful; cost per customer outcome is better.
The practical policy shift is already clear
The companies in these reports are converging on a mature operating model for expensive agents. It has little to do with punishing developers for using AI and everything to do with preventing unbounded workloads from acquiring their own budget.
A workable enterprise rollout needs all of the following:
- Spending limits should apply per user, project, organization, and automated workflow, with an actual stop condition rather than a notification after the money has been spent.
- Teams should track cost by model, repository, task type, agent run, and successful outcome so that a high bill can be investigated before a quarterly review.
- Autonomous agents should receive narrow credentials and explicit execution limits, particularly when they can run tests, invoke cloud services, query data stores, or modify repositories.
- Exception approvals should be tied to a defined engineering objective, such as a migration or incident, rather than granted permanently to the heaviest users.
- Organizations should measure review time, defects, rollbacks, and delivery outcomes alongside pull requests and token volume.
Microsoft’s June 30 internal deadline has passed, but the operational work begins after the switch. Whether engineers use GitHub Copilot CLI, Claude Code, Cursor, or another agent, the durable change is that AI coding is now entering the same governance category as cloud compute: valuable, scalable, and capable of producing a surprising bill if nobody owns the meter.
References
- Primary source: Telangana Tribune
Published: August 8, 2026 at 7:47 AM UTC
Loading…
www.telanganatribune.com - Related coverage: techradar.com
Amazon quietly burned $1.8 million on an AI coding project that went horribly wrong | TechRadar
One failed Claude deployment exposed cracks in Amazon's AI cost oversight systemwww.techradar.com - Related coverage: techradar.com
'Tokenmaxxing is not what we are optimizing for': Microsoft tells engineer to calm down on AI usage | TechRadar
Microsoft forced to introduce new limits on AI token usagewww.techradar.com - Related coverage: indiatoday.in
Claude ate up Uber’s full year AI budget in 4 months? Here is what really happened - India Today
Uber said its AI coding budget was exhausted by April 2026 after engineers rapidly adopted Claude Code. The surge highlighted how usage-based pricing and widespread internal uptake can outpace clear returns.www.indiatoday.in
- Related coverage: seo.software
Uber’s AI Coding Budget Blowout Is a Warning for Teams Scaling Claude Code
Uber’s AI coding budget blowout shows what happens when coding agents scale without cost controls. Here’s how teams should govern spend and ROI.seo.software - Related coverage: linkedin.com
🚨 Uber burned its entire 2026 AI budget in 4 months. 84% of engineers using Claude Code. Monthly cost: $500–$2,000 per engineer. Uber's COO: "The link between AI spend and customer value is not… | Maciej Kraus | 19 comments
🚨 Uber burned its entire 2026 AI budget in 4 months. 84% of engineers using Claude Code. Monthly cost: $500–$2,000 per engineer. Uber's COO: "The link between AI spend and customer value is not there yet." Their fix: cap every employee at $1,500/month in token spend. Translation - they capped...www.linkedin.com
- Related coverage: linkedin.com
After giving 5,000 engineers Claude Code access in Dec 2025, Uber exhausted its entire annual AI budget by April 2026. Between orgs forcing a culture of "token maxing", and frontier models like… | Brittany Hobbs
After giving 5,000 engineers Claude Code access in Dec 2025, Uber exhausted its entire annual AI budget by April 2026. Between orgs forcing a culture of "token maxing", and frontier models like Anthropic and OpenAI defaulting users to the highest cost models, the token economics are going off... - Related coverage: groundtruth.day
Uber reportedly burned through its whole 2026 AI coding budget in four months — Ground Truth
The clearest enterprise cost figure yet for AI coding agents: Uber's CTO is reported to have said the company exhausted its Claude Code budget in a third of the year.
groundtruth.day
- Related coverage: windowsforum.com
Microsoft Cancels Most Claude Code Licenses, Pushes Copilot CLI
Microsoft is reportedly canceling most internal Claude Code licenses in its Experiences and Devices organization by June 30, 2026, after giving thousands of...windowsforum.com - Related coverage: byteiota.com
Loading…
byteiota.com - Related coverage: tomshardware.com
- Related coverage: windowscentral.com
Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI — a move likely driven by financial motives | Windows Central
Claude Code was popular among Microsoft engineers, but the company now wants them to shift to GitHub Copilot CLI.www.windowscentral.com - Related coverage: windowscentral.com
Microsoft CEO Satya Nadella says AI tokenmaxxing is costly: "I'm a tokenmaxxer too, it's addictive." | Windows Central
The executive wants staffers to rethink how they use frontier AI models to solve problems.www.windowscentral.com - Related coverage: tomshardware.com
Github Copilot customers report up to 100-fold price hikes — AI sticker shock bites as Microsoft switches to usage-based pricing | Tom's Hardware
The AI investment chickens have come home to roost.www.tomshardware.com - Related coverage: learn.microsoft.com
Release Notes for Microsoft 365 Copilot | Microsoft Learn
Lists the features that reach General Availability in each release of Microsoft 365 Copilot.learn.microsoft.com - Related coverage: learn.microsoft.com
Copilot in Microsoft 365 apps with Anthropic models | Microsoft Learn
Learn about the new toggle for Microsoft 365 apps for Anthropic in Europe.learn.microsoft.com - Related coverage: microsoft.com
Powering Frontier Transformation with Copilot and agents | Microsoft 365 Blog
Wave 3 of Microsoft 365 Copilot introduces Copilot Cowork, multi‑model intelligence, and enterprise‑ready AI—built to get real work done.www.microsoft.com - Related coverage: techcommunity.microsoft.com
Loading…
techcommunity.microsoft.com - Related coverage: news.microsoft.com
Introducing the Frontier Suite - Source EMEA
news.microsoft.com
- Related coverage: news.microsoft.com
Microsoft kündigt neue agentische KI-Funktionen an - Source EMEA
news.microsoft.com
- Related coverage: microsoft.com
Expanding model choice in Microsoft 365 Copilot | Microsoft 365 Blog
We’re excited to announce that we’re expanding the models that power Microsoft 365 Copilot with the addition of Anthropic models.www.microsoft.com - Related coverage: support.microsoft.com
Usare Claude con strumento ricerche in Microsoft 365 Copilot | Microsoft Support
Usare Claude con strumento ricerche in Microsoft 365 Copilotsupport.microsoft.com - Related coverage: techcommunity.microsoft.com
- Related coverage: adoption.microsoft.com