• Microsoft (MSFT) plans to end most internal Claude Code licenses by June 30, 2026, and cut spending on Anthropic's models by over a third, as it steers engineers toward GitHub Copilot CLI.
  • Meta (META) is restricting employee use of Claude and OpenAI's Codex for AI development work, citing concerns over "distillation," and pushing its own MetaCode agent.
  • The pullback reflects a broader shift from unrestricted AI access to governed procurement, cost controls, and data-security reviews, even as both companies maintain external partnerships with Anthropic.

The Pullback

Microsoft and Meta are taking steps to reduce employee reliance on Anthropic's Claude, particularly for coding and other high-token tasks, according to people familiar with the matter. The moves, reported by The Information, signal a strategic recalibration of how the tech giants manage third-party AI tools internally.

Microsoft has informed its Experiences & Devices organization that it will end most internal Claude Code licenses by June 30, 2026, directing engineers to GitHub Copilot CLI instead. The decision followed unexpectedly high token-based usage costs that exhausted budgets far sooner than projected, the people said. Microsoft also restricted employee use of a newer Claude model while legal teams assessed Anthropic's data-retention terms. Reuters reported concerns that prompts and outputs could be retained for 30 days, and longer when flagged, creating issues for confidential and customer information.

Meta, meanwhile, has issued internal guidelines limiting employee use of Claude and OpenAI's Codex for work involving AI-model development. The stated concern is "distillation"—employees potentially using external-model outputs in ways that could transfer rival capabilities into Meta's internal systems. Meta is encouraging engineers to use its own in-house coding agent, MetaCode.

"It's a matter of cost discipline and data governance," said one person familiar with the matter, who spoke on condition of anonymity. "These are not decisions made lightly, but the bills added up faster than anyone expected."

A Dual Shift

The changes represent a dual shift: from external models to internal tools, and from unconstrained access to budgets, permitted-use rules, and data controls. Importantly, Microsoft is not severing ties with Anthropic at the company level. The two announced a strategic partnership in November 2025 to make Claude models available through Microsoft Foundry and Azure, including continued availability across elements of the Copilot product family. The internal-use retrenchment and the external cloud partnership can coexist: Microsoft can sell customers model choice while limiting costly or risky employee consumption.

Anthropic declined to comment. Microsoft and Meta did not respond to requests for comment.

The financial asymmetry matters. Microsoft can justify huge AI infrastructure spending by selling Azure capacity and AI services to customers; Meta's AI investment primarily aims to improve ad targeting, feed ranking, engagement and future consumer AI products rather than generate equivalent near-term cloud revenue. That makes Meta especially motivated to control outside-model bills while developing its own AI stack.

The Economics of AI Consumption

Agentic coding tools can generate unusually volatile costs because pricing is tied to tokens, context length, tool calls and repeated model reasoning. A coding agent that autonomously searches a codebase, runs tests, revises code and retries workflows can consume far more compute than a conventional chatbot interaction. The reported Microsoft action is thus a warning that "per-seat" AI budgeting may be inadequate for usage-metered agents.

The broader market is moving toward model portfolio management, vendor consolidation, and proof-of-ROI pressure. Enterprises are choosing different models for different work—cheap models for routine tasks, stronger models for complex reasoning, and private or self-hosted models for sensitive work. Large platforms want developers using their native environments—Microsoft with GitHub Copilot and Azure; Meta with MetaCode and internal models—so they retain the product data, developer feedback loop and eventual software revenue.

"We have a constant balance with the banks, which really we consider our partners and not only our binary competitors," said Cecile Mayer-Levi, head of private debt activity at Tikehau Capital SCA (TKKHF), referring to a different context. "It's much more of a convergence between the two solutions." Her comment, while about private credit, echoes the dynamic in AI: competition and partnership often coexist.

At the macroeconomic level, this story sits inside the vast AI capital-expenditure cycle. Meta guided to 2026 capital expenditures of $130 billion–$145 billion, while Microsoft's calendar-2026 capital-spending expectation has been reported around $175 billion. These investments support construction, chips, power equipment, data-center supply chains and cloud services, but they also increase exposure to power constraints, higher depreciation, chip scarcity and investor concerns that returns may arrive slowly.

Market Implications

For Anthropic, a large internal-spend cut at Microsoft or a restriction at Meta can be symbolically important even if the dollar amount is modest relative to Anthropic's total business. These are influential AI buyers and developers; their behavior can encourage other enterprises to question premium token expenditures.

However, it does not automatically mean that Claude is losing on capability. In fact, reports suggest employees adopted Claude Code enthusiastically. The tension is that high-quality models can be both productive and expensive at scale. The competition is increasingly about the full equation: business value = model quality + workflow fit - inference cost - security/compliance risk.

Policy and regulatory context also looms. The EU AI Act's General-Purpose AI Code of Practice is intended to help providers comply with obligations involving transparency, copyright and safety. Big model providers and customers must increasingly account for documentation, risk controls and responsible deployment. Privacy and data handling are central: Microsoft's reported concern about Claude retention terms illustrates why corporate legal teams focus on where prompts, outputs, source code and customer data are stored, who may access them and for how long.

Antitrust and platform power may also come into play. Policymakers may examine whether dominant cloud and software platforms use internal policies, bundling or default settings to disadvantage third-party models. Yet there is a legitimate counterargument: companies have valid reasons to control costs, secure proprietary data and standardize development tools.

Stakeholders and Debate

The main debate is not simply "Claude versus Copilot" or "build versus buy." It is whether companies should maximize employee access to the strongest available AI model, or standardize around proprietary platforms that offer better control of cost, data and strategic feedback. Critics can see enforced internal alternatives as lock-in; supporters see them as basic financial and security hygiene.

Engineers and knowledge workers may lose access to a preferred tool, face quotas or be required to use a lower-cost internal alternative. That can reduce flexibility, but it can also improve consistency, security and support. Corporate customers may gain access to Claude through Azure even as Microsoft employees face stricter internal limits. This highlights a distinction between customer-choice offerings and internal cost governance.

Anthropic faces revenue and influence risk if major technology companies lower internal consumption or build substitutes, while still benefiting from broader cloud-distribution partnerships. Microsoft and Meta shareholders could welcome spending discipline, but will want evidence that internally developed tools match or exceed external alternatives without harming developer productivity. Consumers and advertisers may see little immediate effect, but over time, cost control around external coding models could affect the speed and quality of AI features, ad systems, privacy protections and platform moderation tools.

Smaller enterprises and startups may see validation of their concerns about runaway AI bills, but also better pricing, model routing products and cost-management tools.

Background and Outlook

The situation emerged from several converging trends: frontier models such as Claude became highly valued for coding, reasoning and long-context work; companies deployed them quickly across engineering organizations, often before accurately forecasting agentic-workflow usage; consumption-based token bills exposed the true cost of heavy use; competitive pressure made it unattractive for Meta and Microsoft to rely deeply on rival AI products while trying to build competing systems; and security, intellectual-property and data-retention concerns made unrestricted external-model access harder to justify.

There are precedents in enterprise software: large technology companies commonly move employees from outside tools to internally controlled systems once a strategic capability matures. What is new is the speed of AI-model progress and the fact that usage itself—prompts, outputs, coding interactions and feedback—may help shape future product performance.

Near-term, more corporate AI policies will set token caps, team budgets, access tiers and approved use cases. Anthropic and other model vendors will face stronger pressure to offer predictable pricing, volume discounts, private deployment and clearer data-handling guarantees. Microsoft's GitHub Copilot and Meta's MetaCode will receive more internal usage, generating valuable feedback and potentially accelerating their product development. Engineering teams may test productivity more rigorously as they compare external frontier models against in-house alternatives.

The likely end state is multi-model, governed AI, not a single winning assistant. Organizations will route work among proprietary, open-weight and third-party models according to performance, cost, latency, data sensitivity and regulatory requirements. Microsoft's continued Claude availability through Azure supports that interpretation: the firm wants customers to have choice while it internally controls what workloads are subsidized and what data leaves its boundaries.

For Meta, the long-run test is whether proprietary tools can materially improve the company's core advertising and consumer-product economics enough to justify its extraordinary AI infrastructure spending. For Microsoft, the test is whether Azure and Copilot monetization can sustain investment while keeping internal AI usage efficient. Both companies' recent results show substantial revenue strength, but the rising AI-capex burden means investors will increasingly demand measurable returns rather than broad AI ambition alone.

*Correction: A previous version of this article misstated the date by which Microsoft plans to end most internal Claude Code licenses. It is June 30, 2026, not June 30, 2025. Additionally, the article has been updated to clarify that Microsoft's partnership with Anthropic remains in place for external customers.