LLM Prices Fell Up to 80% in Months. Here's What to Change in Your Stack

Your LLM bill is probably wrong. Not because you can't count, but because the price list you budgeted against is months old.

Since late June, OpenAI and Anthropic have cut API prices repeatedly, by up to 80% on some models. Max Woolf (minimaxir) laid out the timeline this week in a post that hit the HN front page. I pulled the numbers. Prices below are $ per 1M tokens, as input / output / cache read.

Vercel's AI gateway data, cited in the post, shows aggregate token cost at 43.3% of April 2026 levels.

1. Open weights are the floor. DeepSeek Flash V4.1 lists at $0.15 / $0.60 / $0.003. Its cache read is about 2% of input price, versus 5-10% at the closed labs. Per the post, 29 providers host DeepSeek models on OpenRouter. When a model is hosted 29 ways, nobody gets to charge a premium for it.

2. Switching costs are near zero. If your integration is an OpenAI-compatible endpoint, moving providers is a config change. Gateways like OpenRouter make A/B testing a competitor trivial. The moat that used to justify pricing is gone.

3. Serving got cheaper. Anthropic said it directly: Opus 5.5 "requires less compute to serve than Opus 5, and its pricing reflects that."

4. Maybe, subsidy. Both labs have IPO talk and leaked financials showing billions in losses. The author flags the rideshare analogy: cheap now, expensive later. He doesn't buy a sudden spike, since competing providers profitable at these prices cap what anyone can charge. I'd hedge anyway (more on that below).

IBM CEO Arvind Krishna told The Register that open models now cost "anywhere from 25 percent to 10 percent of the token cost" of frontier ones at roughly "90 percent as effective." His pitch: "Would you like something that's 10 percent of the cost if it's 90 percent as good as the best?" He says most firms can get started with two or three racks.

Treat that as a vendor with something to sell. IBM's infrastructure revenue grew 5% YoY to $3.7B, slower than Dell, HPE and Lenovo. But the direction matches what the price list shows.

Take a made-up but realistic workload: 100M input tokens, 20M output, 60% of input served from cache.

# /// script # requires-python = ">=3.10" # /// # $ per 1M tokens: (input, output, cache_read) PRICES = { " gpt-6-sol ": ( 2.00, 10.00, 0.20 ), " gpt-6.1-sol ": ( 2.00, 10.00, 0.10 ), " deepseek-flash-v4.1 ": ( 0.15, 0.60, 0.003 ), } def monthly_cost ( model, in_m = 100, out_m = 20, cache_hit = 0.6 ): i, o, c = PRICES [ model ] return in_m * ( 1 - cache_hit ) * i + in_m * cache_hit * c + out_m * o for m in PRICES: print ( f " { m: 22 s } $ { monthly_cost ( m ): 8.2 f } " ) Enter fullscreen mode Exit fullscreen mode Output:

gpt-6-sol $ 292.00 gpt-6.1-sol $ 286.00 deepseek-flash-v4.1 $ 18.18 Enter fullscreen mode Exit fullscreen mode About 16x between the cheapest and the closed flagship, at list price. That says nothing about quality. Cheap and wrong is still wrong.

Token price is no longer the story. A year ago the question was "can we afford to use an LLM here?" Now it's "why are we paying flagship rates for a task a model at a fraction of the price can do?"

The teams that win the next year won't have the best model. They'll have the best routing, caching and evals, and the discipline to re-check the price sheet every month.

Sources: minimaxir, "OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?" and The Register on IBM's open-weight comments. Prices are list prices as reported there. Check current rates before you budget.