Anthropic's Enterprise Moat Under Pressure: Microsoft Cuts Claude Spend by a Third, Meta's Claude Code Users Halve

Per The Information: Microsoft cut internal Claude spending by over a third; Meta's Claude Code users halved as it pushed internal AI tools (Muse Code with 6,000+ internal users, MetaCode for internal-only). Anthropic's two biggest corporate customers are pulling back as the company ships Sonnet 5.5 with the same $2/$10 list pricing.

Anthropic’s two largest corporate customers are pulling back from Claude at the same moment the company is shipping a new flagship model. Per The Information’s reporting (Oct 5, 2026): Microsoft cut its internal Claude spending by over a third, and Meta’s Claude Code users halved as the company pushed internal AI tools. This is the first concrete evidence that Anthropic’s enterprise moat — the multi-billion-dollar “everyone wants Claude” narrative from 2025 — is being renegotiated in real time.

The numbers and the moves:

  • Microsoft cut internal Claude spend by over a third through 2026, per the report.
  • Meta’s Claude Code users halved as the company expanded Muse Code (external testing) and MetaCode (internal-only) — both positioned as direct Claude Code replacements powered by Meta’s own models.
  • Meta’s Muse Code recently passed 6,000 internal users during the testing phase, per Investing.com’s coverage of the report.
  • The pullback coincides with Anthropic shipping Claude Sonnet 5.5 with the same $2/$10 list pricing as Sonnet 5, faster inference, and a self-reported SWE-bench Pro 81.3% that ties or beats the leading frontier coding models.

What the Microsoft number means

Microsoft’s “over a third” reduction in Claude spending is the kind of number that doesn’t get noticed quarter-to-quarter but does change the trajectory. Microsoft was on track (per their executives earlier this year) to spend “at least” a meaningful amount annually on Claude. Cutting internal Claude spend by 33%+ in a single year, while Microsoft is also OpenAI’s biggest backer, suggests Microsoft is hedging — running Claude where it’s still differentiated, replacing it elsewhere with OpenAI and its own Maia silicon.

For Anthropic, the Microsoft pullback is doubly painful because Microsoft was their largest referenceable enterprise customer. Losing a third of that spend is roughly a 5-7% hit to the run-rate forecast, before factoring in what the next year’s renewals look like.

What the Meta number means

Meta’s “Claude Code users halved” is the cleaner signal. Meta had internal Claude adoption, and the company built Muse Code (a Claude Code competitor powered by Meta’s own models) explicitly to replace it. Muse Code entered external testing in August and has now reached 6,000 internal users, with MetaCode (internal-only) running alongside.

The pattern is what you’d expect: when a frontier model lab has its own coding agent ambition (Meta does — Llama + agentic tooling), and the third-party agent is the standard internal tool, building a replacement is the obvious play. The 6,000-user number is the one to watch: it’s the order-of-magnitude signal that Meta is actually migrating off Claude Code, not just saying they will.

Why now — the Sonnet 5.5 timing

The Sonnet 5.5 launch (same $2/$10 list as Sonnet 5) is the cost-of-switching argument Anthropic wants to make. The pricing is deliberately competitive — it matches OpenAI’s GPT-6.1 Sol list price, the same number the developer community has been benchmark-recalibrating around since DevDay.

Sonnet 5.5’s self-reported numbers are also positioned as the answer to the “why pay for Claude” question:

  • SWE-bench Pro 81.3% — competitive with or ahead of the leading coding-agent benchmarks.
  • OSWorld 2.1 strict 43.5% (vs partial 80.1%) — the strict-mode number is the honest one; partial-mode inflated scores have been a category-wide problem.
  • Terminal-Bench 70.6%, CursorBench 55.5%, FrontierCode Xhigh 52.1% — broader agentic-coding coverage.

These are self-reported. Independent verification matters. But the price tier ($2/$10) is the more meaningful data point for enterprise buyers who are doing cost-of-switching math today.

What it means for builders

Three takeaways:

  1. Frontier model vendors are now competing for the same enterprise budgets. Microsoft is cutting Claude while pushing OpenAI + Maia; Meta is building Muse Code to replace Claude Code. The 2025 “everyone buys Claude” pattern is fragmenting. If you’re an enterprise buyer, this is your leverage window — multi-model routing is no longer optional, it’s the default.

  2. The $2/$10 price tier is now the reference line. OpenAI’s GPT-6.1 Sol, Anthropic’s Sonnet 5.5, and likely Google’s Gemini 4 Argon (when it leaves Fairwind) all target $2 input / $10 output per million tokens. If your agent workload is cost-bottlenecked, this is the price floor to benchmark against — and you should plan for it dropping further as open-weight models (DeepSeek V4.1 Flash, etc.) close the gap.

  3. “Self-reported benchmarks” are the new paper-launched. Sonnet 5.5’s SWE-bench Pro 81.3% is Anthropic’s number, on Anthropic’s harness. Independent arena runs (lmarena, the-chatbot-arena) will surface in the next 1-2 weeks. Don’t migrate production workload on a self-reported number; wait for the independent verification.

What we still don’t know: whether the Microsoft and Meta pullbacks are isolated (driven by their own model ambitions) or the leading edge of a broader enterprise trend. Q4 2026 Anthropic revenue and the next 90 days of customer-rotation reporting will tell us. If two more Fortune 100s follow Microsoft’s pattern, the enterprise moat story is over.

Sources