OpenAI and Anthropic cut LLM API prices up to 80% since June, and budgets are stale

What happened

Since late June, OpenAI and Anthropic have cut API prices repeatedly, with reductions reaching 80 percent on some models. A dev.to post assembles the timeline, crediting an analysis by Max Woolf (minimaxir) that reached the Hacker News front page. All prices below are US dollars per million tokens, listed as input, output and cache read.

The sequence, as reported: GPT-5.6 Sol launched on June 26 at $5/$30/$0.50, alongside Luna at $1/$6/$0.10. Claude Sonnet 5 followed on June 30 at $2/$10/$0.20, and Claude Opus 5 arrived on July 24 at $5/$25/$0.50. On July 30, Luna was cut 80 percent to $0.20/$1.20/$0.02. Sol dropped 20 percent on August 21 to $4/$20/$0.40. Fable 5.1 launched September 1 at $10/$50/$0.25, with cache reads 75 percent cheaper. GPT-6 Sol landed September 22 at $2/$10/$0.20 while Opus 5.5 came in 40 to 60 percent cheaper, and on September 29 GPT-6.1 Sol halved cache pricing again, to $0.10.

Aggregating across models, Vercel's AI gateway data, cited in the post, puts token cost at 43.3 percent of April 2026 levels.

Why prices are falling

The post identifies four drivers. First, open weights set a floor: DeepSeek Flash V4.1 lists at $0.15/$0.60/$0.003, and its cache reads cost about 2 percent of input price versus 5 to 10 percent at the closed labs. With 29 providers hosting DeepSeek models on OpenRouter, nobody can charge a premium for identical weights.

Second, switching costs are near zero. If an integration targets an OpenAI-compatible endpoint, moving providers is a configuration change, and gateways make A/B testing a competitor trivial. The moat that used to justify premium pricing is gone.

Third, serving got cheaper. Anthropic said directly that Opus 5.5 "requires less compute to serve than Opus 5, and its pricing reflects that."

Fourth, possibly subsidy. Both labs are surrounded by IPO speculation and leaked financials showing billions in losses. The author invokes the rideshare analogy — cheap now, expensive later — but doubts a sudden spike, since competitors already profitable at these prices cap what anyone can charge.

The open-weight pitch is going enterprise

IBM CEO Arvind Krishna told The Register that open models now cost anywhere from 25 percent down to 10 percent of frontier token cost while being roughly 90 percent as effective, and that most firms can get started with two or three racks of hardware. The dev.to post flags the obvious conflict of interest: IBM's infrastructure revenue grew 5 percent year over year to $3.7 billion, slower than Dell, HPE and Lenovo. Still, the direction matches what the price lists show.

What a bill looks like

The post models a representative workload: 100 million input tokens, 20 million output tokens and a 60 percent cache hit rate. At list prices that comes to $292 per month on GPT-6 Sol, $286 on GPT-6.1 Sol and $18.18 on DeepSeek Flash V4.1 — a gap of roughly 16 times between the cheapest option and the closed flagship. As the author notes, that comparison says nothing about quality: a cheap model that gets the answer wrong is still wrong.

What to change in your stack

The post's recommendations are concrete. Re-price the stack immediately, because anything budgeted before September is stale, and restructure prompts with large stable prefixes to exploit cache reads, which fell hardest. Route by task: extraction, classification and summarisation rarely need a flagship model, so send those to a cheap or open model and reserve the expensive one for hard calls. Build evals before switching providers, since a chief executive's 90-percent-as-good claim is not a benchmark on your own data. And stay portable: keep model names in configuration rather than scattered through code, and avoid long contracts signed at old prices.

Why it matters

Token price is no longer the deciding factor in most LLM deployments. A year ago the question was whether a given use case could afford a model at all; now it is why flagship rates are being paid for tasks a fractionally priced model can handle. The competitive edge shifts from model choice — which changes monthly — to routing, caching and evaluation discipline. Teams that treat provider pricing as a moving input, and preserve the ability to swap models in an afternoon, are positioned both to capture the current floor and to hedge if prices climb after the rumored IPOs.