Subscription economics: SemiAnalysis prices the plans, and Microsoft and Meta cut Claude spend
Sources: SemiAnalysis, "Anthropic Subscriptions Offer 5x+ More Value Than OpenAI" (10-05) · The Information, "Microsoft Slashes Internal Claude Spending by a Third" (10-05) · The Decoder on Meta and Microsoft (10-05) · AI Weekly Espresso on Beijing "transfer stations" (10-05) · raw: SemiAnalysis RSS, The Information, The Decoder
TL;DR. SemiAnalysis measured what each AI subscription is worth in API-priced tokens by running experiments that isolate one token type (input, cache write, cache read, output) and watching each plan's usage meter move. Key numbers: at Anthropic, subscriptions are about 10% of revenue but can take over 40% of inference compute, lowering blended revenue per MW by about $36M. At the mid-tier "daily driver" models (Opus 5.5 vs GPT-6.1 Sol), Anthropic's $200 plan gives about 5x the API-equivalent value of OpenAI's. Frontier tiers (Fable 5.1 vs GPT-6 Astra) are similar, but Fable is capped at 50% of a plan. OpenAI halved the $200 plan's token allowance (old buyers keep it until Oct 29) and added a $500 tier that gives only 21% more Astra than the old $200 plan, with "Ultrafast" (up to 8x speed, 8x burn) as the headline. They also caught a provider silently A/B testing ~20% lower limits on one account. The same day, The Information reported Microsoft cut projected internal Claude spend (once on track for $1B+) by more than a third, and The Decoder reported Microsoft's cloud division cut per-employee monthly Claude budgets from $100,000 to $10,000 while Meta halved Claude Code users to 30,000. In Beijing, gray-market "transfer stations" resell Claude at 70-90% off through overseas accounts, sometimes relabeling Qwen output as Claude.
Why it matters
- The credit-cost ratio is a hidden price list. Plans sell a meter, not tokens, and the meter's per-token-type rates differ from API ratios, so value depends on the (plan, model, workload) tuple. Agentic workloads are cache-read heavy, which is where plan value diverges most.
- Subsidy is a compute allocation decision. 40% of compute for 10% of revenue is the Anthropic number that matters for MW economics on the compute economics page.
- Enterprise customers are now rationing. Microsoft and Meta cutting internal use (and pushing their own tools) is the demand-side version of the same pressure. It also makes a routing layer (send most calls to cheaper in-house models) a budget item, see LLM routing.
Relations
- Confirms the 10-05 digest's "cost unit is changing" thread: cache reads dominate agent bills (Kepler, 10-04: 97.37% cache reads), and plan meters price cache reads very differently from APIs.
- GPT-6's 100x price spread (Astra $10/$50 vs Luna $0.10/$0.50 per M tokens) is the API-side counterpart.