Every AI contract without a price-drop clause is a gift to OpenAI and Anthropic

On 26 June OpenAI put GPT-5.6 Luna on sale at $1 per million input tokens and $6 per million output. On 30 July it knocked 80% off, to 20 cents and $1.20. Anyone who spent July negotiating committed spend against that launch card agreed to pay five times the going rate for a price that lasted about five weeks.

I read Max Woolf's timeline of this summer's price war between OpenAI and Anthropic expecting a story about two labs dressing up their numbers before an IPO. What surprised me was how few launch prices survived at all. Anthropic launched Claude Sonnet 5 at $2 input and $10 output, billed as a temporary rate that would rise, and on 10 August made it permanent. OpenAI's Sol line went from $5 input and $30 output in June to $2 and $10, with cache reads at 10 cents, by 29 September. Woolf reckons that is cumulatively 75% cheaper. Claude Opus 5 launched at $5 and $25 on 24 July; two months later Opus 5.5 arrived at $4 and $20 with cheaper cache reads, a total cut of 40 to 60%, and Anthropic said plainly that the new model costs less to serve and "its pricing reflects that".

I believe Anthropic on that, and it is exactly why buyers should change how they sign. If serving costs keep falling and the two labs keep matching each other, the rate card on the day you sign is an opening bid. I don't think this summer was a one-off either. In June 2025 OpenAI cut o3 by roughly 80%, from $10 and $40 to $2 and $8, calling it the same model on a faster inference stack, and ARC Prize retested it and found no change in performance. Opus 4.1 cost $15 per million input tokens; Opus 4.5 came in at $5. Epoch AI puts the median fall in the price of hitting a fixed benchmark score at 50x a year.

Now look at what the sellers want. Woolf notes both labs are leaning on enterprise commitments to build a moat, and I would too in their position. A multi-year committed-spend deal converts a price war into an annuity. Outside that deal, switching is a toggle in a coding harness or a different endpoint string. There were 13 providers hosting DeepSeek V4 Flash on OpenRouter in May; 29 now host the 1017 update at around $0.15 input and $0.60 output. Moonshot's Kimi K3.1 Pro scores 55 on the Artificial Analysis index against 58 for Opus 5.5, costs $2 and $10, and its weights drop on 4 November, with more than 20 providers already hosting its predecessor. A commit is the one thing that stops you using any of that. Whatever discount the account team offers for committing is the price they are paying for your option to leave, and with list prices falling 75% in a quarter I think that option is worth far more than they are offering.

The sensible pushback comes from people tracking the data. Report AI's 2026 price index argues that frontier sticker prices fall more slowly than capability per dollar, and Epoch warns the fastest declines of the past year may not persist. An analysis on DEV Community points out that a 280-fold fall in inference prices did not shrink anyone's bill, because usage grew faster. I accept all three and they push me further towards pass-through. If declines slow, a clause that tracks list price costs the vendor almost nothing, so a vendor who fights it is telling you what it expects prices to do. If your volumes keep climbing, every cent of unit price is multiplied by more tokens each month, which makes the clause more valuable to you, and Opus prices have kept falling at the frontier regardless.

Here is how I would write the clause. It should reset your rate to the lower of the contracted price or any published list price for the same model or its successor, automatically, from the day of the announcement, with no renegotiation. It must name the cache read rate. When Anthropic released Fable 5.1 it kept input and output at $10 and $50 and cut cache reads from $1 to $0.25. Woolf calculates that cache reads make up 96 to 99% of token spend in agentic coding, so that one line cut the real cost by 70 to 75%. A clause pegged only to input and output prices would have captured none of it. Allowances need the same treatment: on the day GPT-6.1 Sol launched, OpenAI halved the usage allowance on its $200 a month Codex tier, citing its increasingly efficient models. Seat plans priced in loosely defined usage let the vendor keep the efficiency gain for itself.

Pull last month's invoice for your coding agents and work out what share of spend went on cache reads. If your price-protection clause does not mention the cache read rate, it protects the few per cent of the bill that is left, and you should have that fixed before the Kimi weights land on 4 November and the next round of cuts starts.

Prompted by OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?, Max Woolf's Blog.