The price timeline is stark. On June 26, OpenAI previewed GPT-5.6 Luna at $1/$6/$0.10 per million tokens -- already positioned as a cheaper alternative to the flagship Sol model. Anthropic launched Claude Sonnet 5 on June 30 at $2/$10/$0.20 with temporary launch pricing intended to rise later. By July 30, OpenAI had slashed Luna by 80% to $0.20/$1.20/$0.02. In August, Anthropic permanently kept Sonnet 5 at its cheaper launch pricing after mixed feedback on Opus 5's value proposition. In September, Anthropic's Fable 5.1 launched with cache reads cut 75%. OpenAI's own GPT-5.6 Sol took another 20% cut in August for a three-month window ($4/$20/$0.40).

The driver, according to the analysis, is not loss-leading -- it is structural. Open-weight models and secondary inference providers now serve as an "economic check" on closed-model pricing power. When a trillion-parameter hybrid MoE like Mistral Large 4 (weights due Oct 31) and specialized coder models like Nanbeige4.3-33B compete on equal footing with proprietary offerings, the price premium for black-box models narrows significantly. Even if Anthropic's and OpenAI's IPOs were catastrophic, the industry would survive -- the pricing floor is set by architecture diversity, not lab benevolence.

For Indian enterprise AI planners, the practical implications cluster around three areas. First, API costs become so dramatically compressed that deploying production-grade agent systems on frontier APIs approaches marginal-cost economics -- the constraint shifts from token bills to governance, evaluation, and integration engineering. Second, the sovereignty lane gains urgency: MeitY's IndiaAI Mission 10,000-GPU tender and domestic hosting options (TCS Hyderabad DC, HCLTech-Odisha DC with Sarvam) matter less for raw inference when cloud-hosted open-weight models deliver comparable performance at fraction of the cost. Third, vendor procurement arguments based on "cheaper AI tools" weaken when every major provider has undercut itself 80% -- differentiation must rest on safety guarantees, compliance posture, and ecosystem lock-in, not price.

The compute-financing counter-narrative adds depth to the price story. While API unit costs plummet, infrastructure concentration accelerates. SpaceX's $40B GPU debt raise -- backed by Starlink cash flow and investment-grade credit -- demonstrates that capital flows preferentially to firms with diversified revenue streams independent of AI dependency alone. Musk's prediction of 200-plus neocloud deaths implies a future where compute capacity consolidates into fewer, better-capitalized pools. Data center tenants now face late-delivery penalties of $5M or more per day -- up fivefold from earlier this spring -- as hyperscaler demand creates steep leverage for capacity-constrained developers.

The India-specific mirror: the same pricing pressure applies to Indian IT service providers negotiating AI-led deals for global clients. If API costs collapse while delivery overhead remains fixed, margin compression becomes inevitable unless delivery teams leverage cheaper models to offer lower-price tiers. Simultaneously, the Krutrim story (today's Economic Times reports 60+ linguistics and semiconductor leaders departed as funding dried up) shows what happens to AI-native startups when the dual forces of pricing disruption and funding contraction collide: ambitious full-stack platforms shrink on A-level exits while competitors who master the open-weight/inference layer capture the resulting market share.