Microsoft and Meta Just Slashed Their Claude Budgets. Here’s the Checklist for Engineers.

As Anthropic moves from API partner to enterprise rival, two of its largest customers are cutting access. What it means for teams mid-migration.

Microsoft’s cloud division just cut per-employee Claude spend from $100,000 a month to $10,000. Meta halved its Claude Code footprint from 60,000 seats to 30,000. Both are leaning harder on their own tools — Copilot, Meta’s internal AI — while Anthropic walks upmarket to sell the same enterprise platform its customers were building on top of.

Read that sequence slowly. Two of Anthropic’s largest customers are now throttling it, because the API provider turned into a rival. I’ve watched this exact movie with databases, with cloud, with search APIs. The ending is always the same: the people who didn’t plan for the vendor’s incentives to change get surprised, and the people who did shrug and reroute. Here’s the order I’d actually do things in.

1. Find out whose budget you’re living inside

If you’re an engineer at Microsoft or Meta, your Claude access isn’t yours — it’s a line item in someone’s cost center, and that line just moved. A $10,000/month ceiling sounds generous until you learn your team’s agentic test harness burns that in a long afternoon. Before you optimize anything, find the cap, find who owns it, and find out whether it’s per-seat, per-team, or pooled. You cannot plan against a number you haven’t seen.

2. Instrument spend per workflow, not per month

The monthly invoice tells you nothing useful. You need cost attributed to the workflow: which agent, which repo, which pipeline. When a cut lands, you want to kill the three jobs eating 70% of the budget, not blind-ration everyone. Log token usage with tags from day one.

1
2
3
4
5
6
7
8
9
10

resp = client.messages.create(
    model="claude-...",           # pin a version, see item 4
    max_tokens=1024,
    metadata={"user_id": team_id}, # tag at the API boundary
    messages=msgs,
)
# emit to your own meter, keyed by workflow + model + env
meter.record(workflow="pr-review", model=resp.model,
             in_tok=resp.usage.input_tokens,
             out_tok=resp.usage.output_tokens)

3. Put a model-agnostic shim between your code and the vendor

If your prompts call anthropic.messages.create directly in 400 places, you don’t have an integration, you have a hostage situation. Route everything through one internal interface that takes a prompt and a model name and returns text. Swapping providers should be a config change and a prompt tweak, not a sprint. This is the single highest-leverage thing on the list, and it’s boring, which is why people skip it.

4. Pin versions and keep eval sets you own

When you’re forced to move — to Copilot, to an internal model, to another frontier vendor — the question is “does the replacement actually hold up on my work?” You can’t answer that without a golden set of real tasks and graded outputs. Pin the model string you’re on today so behavior doesn’t drift underneath you, and run every candidate through the same evals. Vendor benchmarks are marketing; your eval set is evidence.

5. Treat “internal alternative” as a real option, not a demotion

Both companies are pushing their own tools partly for strategy and partly for cost. Some of your workflows genuinely don’t need a frontier model — a cheaper internal one is fine for lint-level review, templating, classification. Tier your work. Save the expensive calls for the tasks that actually need the reasoning.

6. Negotiate on the assumption that incentives change again

None of this means Claude is worse — the models are excellent and that was never the point. The point is that your vendor’s roadmap and your employer’s roadmap just stopped pointing the same direction, and that can happen with any provider overnight. Multi-vendor isn’t a purity test. It’s leverage. Have a second provider wired up and warm, even if it carries 10% of traffic, so your fallback is a dial you turn, not a project you start.

If you only do one thing this week: build the shim in item 3. Everything else is easier once swapping a model is a config change instead of a migration. The budget already moved. Make sure your code can.