Between June and September 2026, the cost of using frontier LLMs from OpenAI and Anthropic collapsed. GPT-5.6 Luna dropped 80% in a single announcement. Opus 5.5 shipped 40-60% cheaper than Opus 5. GPT-6.1 Sol landed at 75% below its predecessor's price point just three months prior. The speed is unusual. The direction is not. The tit-for-tat timeline reads like a pricing duel: OpenAI releases Sol and Luna in June, Anthropic counters with Sonnet 5 at a lower price four days later, Anthropic launches Opus 5 in July at parity, OpenAI retaliates with an 80% Luna cut six days later. By September, both companies released new model families on the same day — GPT-6 Sol and Opus 5.5 — each undercutting its own predecessor. The coordination is competitive, not cooperative. The structural driver is not generosity. It is the open-weight inference market. DeepSeek Flash V4, released in April at $0.14/$0.28/$0.0028, went viral specifically because its cache read cost — the dominant cost driver for agentic coding at 96-99% of total token expenditure — sat at just 2% of input price versus 5-10% at OpenAI and Anthropic. That $0.0028 cache read forced the frontier labs to respond. When 29 distinct providers host the same open-weight model on OpenRouter, price discovery is ruthless. The cache read economics deserve particular attention. Agentic coding workflows hit cache 90-100% of the time. When Anthropic cut Fable 5.1's cache read cost by 75% in September, the effective cost reduction for typical agent usage was 70-75% — far larger than the headline token prices suggest. OpenAI matched this with GPT-6.1 Sol's cache price at 5% of input. These are not marketing discounts; they are structural repricing of the actual cost center. The Uber/Lyft loss-leader analogy that AI skeptics invoke fails on market structure. Uber and Lyft achieved a near-duopoly with high switching costs. LLM inference has 29+ providers for a single model, near-zero switching costs (change one API string), and open weights that anyone can host. Moonshot AI's Kimi K3.1 Pro, announced days ago with Opus-competitive performance at half the price, will be hosted by 20+ providers once weights drop November 4th. The barrier to entry is a GPU rack, not a regulatory moat. The Vercel data confirms the macro trend: cost-per-token fell to 43.3% of April 2026 levels by October. But aggregate spend rose because capability improvements outpace price cuts — users do more, not less, when tokens get cheaper. This is demand elasticity working as expected in a technology commodity market, not a bubble waiting to pop. What emerges is a market where frontier labs must justify premium pricing through genuine capability differentiation — Opus 5.5 reportedly outperforms Fable 5.1, and both companies are investing heavily in enterprise moats and personal agents like Dots and Duet. The price floor is set by the cheapest competent inference provider, and that floor keeps dropping. Both OpenAI and Anthropic appear to be reducing actual serving costs (Anthropic explicitly stated Opus 5.5 requires less compute to serve), meaning margins may be healthier than the sticker-shock price cuts suggest.
