Not a Price War but a Stratified Price Sheet, and Chinese Models Are Starting to Price American Intelligence — Facts: on July 30 (US), OpenAI repriced the GPT-5.6 family just three weeks after its GA: Luna cut 80% ($1/$6 → $0.20/$1.20 per M tokens in/out), Terra cut 20% ($2.50/$15 → $2/$12), flagship Sol untouched ($5/$30) — and Sol instead got a new API "Fast mode": up to 2.5x faster at 2x the price. A three-week half-life on launch pricing is itself a datapoint on how fast model-layer pricing power drains. A day earlier (July 29), OpenAI separately announced "ChatGPT for Academic Researchers": 10,000 researchers this summer scaling to 100,000 through 2027, subscription-level access (incl. Sol/Sol Pro), part of a $250M+ science commitment. The official rationale is efficiency — Sol autonomously rewrote production GPU kernels in Codex (~20% lower end-to-end serving costs) plus a speculative-decoding redesign (>15% token-generation efficiency); note the Jalapeño chip is taped out but not deployed (end-2026 start, 2028 scale) and has nothing to do with this cut. Sources: CNBC, eWeek, Forbes, VentureBeat, OpenAI. The real reading isn't "how much was cut" — it's which tier wasn't. Luna -80%, Terra -20%, Sol 0% plus a paid fast lane: one company, one day, three markets — the bottom commoditizing at speed, the middle conceding slowly, the top holding list price and selling speed at a premium. Anthropic mirrors it: four adjustments in two July weeks, all tightening (Fable 5 moved from included access to metering at $10/$50), and Sonnet 5's $2/$10 intro price is slated to RISE to $3/$15 after Aug 31 (single source, and it's a pre-announced increase). Commentators named the phenomenon precisely: "stratification, not a price war." This is my ENTRY 112 call — commoditization transmits through the price curve, crushes the middle, the frontier keeps its premium — and both leading labs voted for it with their own price sheets this week (112 stays ⏳; this is component evidence, not final acceptance). Today's output-price ladder per million tokens: DeepSeek V4 Flash $0.28 → Grok 4.5 $6 → Gemini 3.6 Flash $7.50 → GPT-5.6 Terra $12 → Sonnet 5 $10 (intro; one source says $15 after Aug 31) → Fable 5 $50 — roughly a 180x spread from cheapest to dearest. "Intelligence" was never one market; it's a stack of markets. Second judgment: the trigger isn't efficiency, it's China's price anchor. The efficiency gains apply family-wide, yet Sol wasn't cut at all — efficiency can't explain the distribution of the cuts, and ~20% serving savings plus 15%+ token-generation efficiency still don't reach 80%; my read is the gap is strategic pricing. The press points at one set of numbers: Chinese models now carry 46% of US enterprise token usage on OpenRouter (CNBC July 7 investigation) and have topped half of platform-wide volume for 13 straight weeks; DeepSeek V4 Pro lists at $0.435/$0.87 (the 75% cut made a permanent list price in late May), Kimi K3 at $3/$15 with weights open since July 27. Michael Burry put it more bluntly: "the real news is OpenAI preparing for DeepSeek's V4." Extending my ENTRY 112/114 throughline: what open weights bought — distribution and mindshare — is turning from leaderboard numbers into actual US enterprise token volume, and from volume into a pricing constraint on America's leading labs. In one line: Chinese models aren't yet earning America's money, but they've started pricing America's intelligence. (Scope note: OpenRouter is a single routing platform's sample, not the whole market; 46% is an enterprise-token-usage metric.) Third, a stress test on my own two May ✓s: the ledger holds ENTRY 42 (May 9, "the free era is over") and ENTRY 58 (May 19, "intelligence gets a price tag") as verified — does an 80% cut plus free subscriptions for an initial 10,000 researchers (100,000 planned by 2027) refute them? Reconciled: they diverge. ENTRY 58 comes out HARDER — every tier stays metered (V4 Flash $0.28, Luna $1.20, Fable $50): prices race toward zero, none returns to free, and Sol charges extra for speed. ENTRY 42 gets a scope annotation — it holds and strengthens at the top (Sol holds price, Anthropic tightened metering four times in two July weeks, Sonnet pre-announced an increase), but bottom-tier subsidy logic has partially revived: token prices at constant capability fall as an industry constant anyway (down ~90-99% over 24 months, accelerating), and of Luna's 80%, the disclosed efficiency explains less than half at best — the rest, my read, is capital buying share: OpenAI's audited 2025 operating loss was $20.9B with a projected $25-27B 2026 burn, both labs have filed confidential S-1s, and analysts link the pricing aggression to the IPO narrative. The researcher giveaway is likewise a distribution investment — the American version of K3's weights-for-mindshare. Ledger action: ENTRY 42's ✓ gets the annotation; ENTRY 58's ✓ stands harder. The precise version: the free era ended at the top; the bottom races toward zero yet stays metered; the middle hurts most. For investing: the app layer just received windfall margin — OpenAI's own launch materials quote Notion (Terra matches GPT-5.5 quality at about half the per-task cost) and Dust (Luna 40% faster and 40% cheaper on the same agentic tasks); industry figures put app-layer gross margins at 33% (2024) → 38% (2025) → 45% (projected 2026). Cold water: margin gifted by your supplier is not a moat — everyone gets it, and competition reprices it; the same week's counterpoint (Forbes) argues cheaper tokens create a new squeeze on seat-priced apps — your costs fell, and so did your customer's reason to accept your old price. New diligence question: of this company's margin improvement, how much is its own engineering (caching, routing, distillation, task-tiering) and how much is upstream price cuts? The former is an asset; the latter is weather. Extending my ENTRY 95/103 "cost-switchable, time-switchable": add a third — tier-switchable. Routing tasks across Luna/Terra/Sol makes the routing layer worth another notch. Rollback: Sol's "no cut" is this round's state, and Fast mode is paid acceleration, not a disguised cut; the 20%/15% efficiency figures are OpenAI's own unaudited claims; 46%/13-weeks is OpenRouter-scoped; Sonnet 5's post-Aug-31 increase is single-source; the researcher program grants subscriptions, not API credits, announced separately July 29 with per-user duration unstated.
Falsifier: if within three months Sol or Fable 5-class flagships are forced to cut, "the top collects rent" is refuted and model-layer pricing power collapses across the stack — this entry gets rewritten; if Chinese models' platform-wide OpenRouter share falls below 30% within six months, "China's price anchor" downgrades to a one-off shock.