OpenAI and Anthropic cut LLM prices up to 80%, reshaping AI feature economics

What happened

OpenAI and Anthropic have cut prices on their language models by as much as 80% since July 2026, according to a breakdown by Max Woolf published on 25 October and summarised in a 30 October article on dev.to. The dev.to piece walks through what those cuts change in practice, and its core argument is that the headline number hides the part of an API bill that matters most.

Per Woolf's figures as cited by dev.to, OpenAI's small "Luna" tier fell roughly 80% in July and then halved again with the release of GPT-6 Luna in September. OpenAI's flagship "Sol" models got cheaper in stages through August and September, and GPT-6.1 Sol cut cache-read prices in half again. Anthropic made a launch discount on Sonnet 5 permanent, reduced cache-read pricing sharply on Fable 5.1, and priced Opus 5.5 at 40–60% below Opus 5 overall.

Woolf attributes the cuts to competition: open-weight models and a wide field of inference providers now offer quality so similar that price becomes the main remaining differentiator.

Cache reads dominate agent bills

The article's central point is that input and output list prices are the wrong place to look for many workloads. Citing Woolf, it reports that in agentic coding, cache reads make up 96–99% of token spend: agents repeat a largely identical payload of system prompt, tool definitions and file context on every turn, and nearly all of it is served from cache. For that kind of workload, to use the article's illustration, a 75% reduction on cache reads matters far more than a 20% reduction on output tokens; for a one-shot summariser with no shared prefix, the balance reverses.

The recommended first step is instrumentation. Both major APIs report per-request usage, and logging input, output, cache-read and cache-write tokens per feature turns an opaque blended bill into per-feature unit costs — the difference between knowing what "the AI" costs in total and knowing what a ticket-triage feature costs per thousand tickets.

Designing prompts so the cache hits

If cache reads are now nearly free, the corollary is that requests have to be cacheable in the first place. The article's rule is to place stable content at the start of a prompt and per-request content at the end. Anthropic's Messages API marks the stable prefix explicitly with a cache-control marker; OpenAI applies caching automatically to long prompts, but the same ordering principle applies. One timestamp placed at the top of a system prompt can invalidate the cache for everything that follows. Cache-breakers the article flags include dynamic values in system prompts, tool definitions serialised in a different order on each call, and JSON context re-serialised with unstable key order.

Routing, portability and repricing

With small models now priced in cents per million tokens, the article argues against sending every request to a flagship model out of habit. A routing layer that maps tasks such as classification, extraction and summarisation to a small model — and reserves larger models for reasoning and code — covers most products, provided teams benchmark the small model on a few hundred real inputs before committing. It also recommends hiding providers behind a single interface, with model IDs and fallbacks kept in configuration, so a price change can be answered by editing an environment variable and re-running evals rather than rewriting application code.

Two caveats come with the advice. Cheaper does not mean automatically good enough: for many narrow tasks, plain code — pattern matching, a conventional classifier, a database lookup — still outperforms an LLM, and price cuts do not change that. And subscription plans may not track API prices; according to dev.to, Woolf cut his own subscription quota and expects plan allowances to tighten before API costs climb again.

Why it matters

A fall of this size changes the unit economics of features that were rejected on cost only months ago, and the article suggests repricing shelved ideas at current rates — with a stress test of whether each one still makes sense at double today's prices. More broadly, the cuts are evidence that inference is commoditising. With quality converging across providers, prices can move in either direction, and the developers who capture the savings will be those who know exactly which tokens they are paying for and can change models by editing configuration.