OpenAI Slashed Its Cheapest Model 80% in One Day—And Just Told You Where the AI Profit Pool Is Collapsing

The price grid below is the cleanest evidence that frontier and budget have become separate markets. OpenAI's flagship Sol sits at $5 input / $30 output per million tokens. Its cheapest model Luna now lists at $0.20 input / $1.20 output—15× cheaper on input, 25× cheaper on output. Terra, the middle tier, dropped 20% to $2/$12, putting it on roughly equal footing with Anthropic's Claude Sonnet 4.6 at $3/$15.

The competitive picture is asymmetric. DeepSeek V4-Pro made its 75% promo discount permanent on May 23, 2026, settling at $0.435 input / $0.87 output—now 3–19× cheaper than OpenAI GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8. China's budget tier is structurally below $1.50 combined per million tokens. OpenAI's Luna now sits roughly at parity with Qwen3.7-Plus ($1.28/M combined) and below Gemini 3.1 Flash-Lite—meaning OpenAI defended the small-model tier not with a marketing discount but by undercutting its own future revenue to remain in the conversation.

Two months before the July 30 cut, OpenAI engineers had already publicly disclosed a software-only optimization that halved inference costs for models it was applied to (The Information, June 30, 2026). The Luna cut extends that work productized: OpenAI's engineering leverage on inference economics now exceeds the bargaining leverage of any single cloud deal. Adjusted gross margin on OpenAI's inference reportedly improved from 33% in 2025 to 39% in Q1 2026 even as revenue tripled to a $5.7B quarterly run rate—implying the company is using productivity gains to subsidize price, not to harvest margin.

The compounding math is the key. A 20% kernel-level improvement and a 15% speculative-decoding improvement do not add; they multiply. Stacked, they leave roughly 1 × 0.80 × 0.85 = 68% of the original serving cost, and OpenAI chose to pass nearly all of that ~32% saving into the list price rather than keep half. For a company with a publicly disclosed path to ~$14B of 2026 losses, the calculus is rational: defend developer mindshare now, monetize later via the flagship tier (Sol stayed at $5/$30) and via Codex/ChatGPT auto-review, which was upgraded from GPT-5.4 to GPT-5.6 Luna on the same day—cutting that feature's costs by roughly 10×.

The price cut hurts the layer that resells OpenAI inference at fixed margins. Microsoft books Azure OpenAI Service as a take-rate on third-party model tokens; that take-rate just got squeezed on the lower tiers. Microsoft's FY26 Q3 capex-to-revenue hit 37.3% as the company front-loads AI capacity—capex that has to be amortized over a revenue base whose per-token economics OpenAI is now actively eroding. Return on invested capital for Microsoft sat at 5.4% in that quarter versus 6.9% a year earlier, evidence that capex is outrunning the model's pricing power.

Amazon's AWS is more diversified but equally exposed on Bedrock. The investment thesis for AWS in 2026 rests on Trainium 2 silicon and Anthropic model resale; OpenAI's small-model price war does not directly hit Trainium margins, but it raises the cost-curve bar any AWS-hosted model must clear. Capex-to-revenue at Amazon ran 27% in Q2 2026 with operating cash flow sales at only 20.8%, meaning every additional dollar of inference volume must clear an increasingly thin spread once OpenAI pulls the rug on price.

The pure-play inference hosts have the most direct exposure. CoreWeave's Q1 2026 revenue was $2.08B (+112% YoY) on a $99.4B backlog, but negative free cash flow and a debt-to-equity ratio of 7.4× make it the most rate-sensitive to a price war on the workloads it sells. Nebius, growing AI-cloud revenue at 841% YoY to $390M in Q1 2026, has sold out capacity and is still loss-making on an operating basis (operating margin -68.8%)—it needs the inference price floor to keep rising, not falling.

If the price war is being won on inference efficiency, the value flows to whoever supplies the silicon that runs the cheaper models. NVIDIA Blackwell GPUs now deliver inference at approximately $0.02 per million tokens according to SemiAnalysis InferenceX benchmarks (April 2026)—roughly 7× cheaper than H100's $0.14/M tokens, and several orders of magnitude below the B200's hourly rental rate of $3.50–$6.00. That gap between rental price and per-token cost is the moat: the silicon has a token-floor that GPU-hour pricing never sees.

The volume leg of the trade is going to lower-precision inference silicon. FP8 and INT8 paths cut memory bandwidth and energy per token, and OpenAI's own optimization play (Triton/Gluon kernel rewrites, speculative decoding) compounds on hardware that was already designed for it. NVIDIA's FY27 Q1 delivered $253.5B trailing-twelve-month revenue at a 74.1% gross margin and 63.0% net margin—pricing power that survives a software-driven price war precisely because hardware architecture is where efficiency gains land.

Downstream, Super Micro Computer's liquid-cooled AI server business captured much of the inference rack demand at sub-1× P/Sales (TTM 0.53×), the cheapest multiple in the AI hardware stack. The supply chain that feeds both—HBM and high-end DRAM—reported record results: SK hynix Q2 2026 revenue hit ₩79.3 trillion (+257% YoY), with HBM accounting for roughly 30% of DRAM revenue. The inference price war is bad for cloud token margins but is expanding the installed base of HBM-rich systems needed to run the cheap models.

Short-term (days to quarters): three catalysts dominate the tape. First, Anthropic is expected to follow OpenAI's lead—WSJ reported on June 10, 2026 that both vendors are contemplating further cuts ahead of OpenAI's IPO. Second, Microsoft Azure's FY26 Q4 earnings (late July) will show whether the 37% capex/revenue ratio is generating the inference revenue to amortize it. Third, CoreWeave's Q2 print (already guided to $2.45–2.6B vs $2.7B consensus) will reveal whether backlog conversion is slowing. Watch the inference revenue line inside cloud earnings—the first sign of margin compression will be hyperscaler AI services revenue growth decelerating while capex stays elevated.

Long-term (1–3 years): the structural read is that OpenAI has demonstrated a closed optimization loop—frontier model optimizing the serving stack for the budget tier—that no other vendor has matched. If the loop replicates across Anthropic, Google, and Chinese labs, the inference market becomes a software-economics arms race where the silicon floor keeps falling and the application layer keeps absorbing the savings. The winners are vertically integrated: NVIDIA (silicon + CUDA moat), SK hynix (HBM supply), and the cloud hosts that own proprietary accelerators (Alphabet TPU, Amazon Trainium). The losers are the pure-resellers who cannot pass through their own productivity gains—most exposed: Oracle on OCI's 14% gross-margin Nvidia cloud business.