LLM prices are collapsing. Spending is still climbing.
OpenAI and Anthropic have cut prices by as much as 80% in months. Cheaper tokens are real, but the evidence points to better models, not lower bills, as what drives usage.
Photo by İsmail Enes Ayhan on Unsplash
The tidy story is that AI adoption has been held back by cost, and that OpenAI and Anthropic are now removing the barrier. The first half of the story is well documented. The second half is only partly supported.
Since June, the two labs have repeatedly lowered what they charge for their most popular models. The pattern looks less like a one-off promotion than a price war.
How steep the cuts have been
OpenAI's small model, GPT-5.6 Luna, launched at $1 per million input tokens and $6 per million output tokens. OpenAI later cut that by 80%, to $0.20 and $1.20.
Its flagship line followed. GPT-5.6 Sol started at $5 input and $30 output, then dropped to $4 and $20 for a limited three-month window. GPT-6 Sol arrived at $2 and $10. A later update, GPT-6.1 Sol, halved the cost of cached input again. One analysis of the timeline puts the cumulative saving for that model family at 75%.
Anthropic moved the same way. It made a temporary launch discount on Claude Sonnet 5 permanent. Claude Opus 5.5 then landed at $4 input and $20 output, with Anthropic saying the model needs less compute to serve than its predecessor. Total cost fell 40% to 60% against Opus 5, according to the same analysis, for a model that outperforms the pricier Fable 5.1.
Cached input matters more than it sounds. In agentic coding, where each step re-reads the whole conversation so far, cache reads make up 96% to 99% of token spend. So a 75% cut to that one line item works out to roughly a 70% to 75% cut in the total cost of using the model.
Cheaper tokens did not shrink the bill. They gave buyers a reason to use better models more.
Why now
The likeliest driver is competition, not generosity. Since agentic tools took off last November, swapping one model for another has become a settings toggle or a changed endpoint. When switching is that easy, brand loyalty counts for little.
Open-weight models add pressure from below. DeepSeek's V4 Flash launched in April at $0.14 per million input tokens and went viral, partly because OpenAI had no small model to answer it. The cut to Luna in July looks like a response, though that is an inference rather than something OpenAI has said.
The supplier base is widening. On OpenRouter, a marketplace that routes requests to hosts of open models, 13 providers served DeepSeek V4 Flash in May. The updated version now has 29. Several of them price within about 10% of DeepSeek itself, which suggests the margin for anyone selling the same weights is thin.
Quality is converging as well. Moonshot AI's Kimi K3.1 Pro scores 55 on the Artificial Analysis index against 58 for Opus 5.5, at half the input and output price. Cerebras, a chip maker, says it can serve Kimi K3, a model of about 3 trillion parameters, at 1,000 to 2,000-plus tokens per second. Open models are getting both cheaper and faster to run.
Does cheaper mean more use?
Here the evidence gets more complicated. Data from Vercel's AI gateway shows the cost per token down to 43.3% of its April 2026 level. Yet total spend on the gateway has kept rising, because models are improving faster than they are getting cheaper.
That points to a different mechanism from the one in the assignment's framing. If buyers were simply waiting for lower prices, spend might flatten as tokens got cheaper. Instead, buyers appear to be trading up to stronger models and running more work through them. The sources do not show whether new customers are entering because of price, so the claim that cost was the main barrier remains unproven.
Price cuts also do not always raise consumption. One heavy Codex user reported cutting his subscription tier to a fifth of its former size after a 75% effective price drop, because the same work now took far less. At the same time, OpenAI halved the usage allowance on its $200-a-month Codex tier, saying its models had become more efficient. If that pattern spreads, efficiency gains may be captured by the labs in subscription terms rather than passed on.
What it signals
First, the market is acting like a commodity. Both labs are leaning on enterprise commitments and personal agent products to build loyalty that raw model quality no longer guarantees.
Second, the margin question is open. OpenAI and Anthropic are rumored to be heading for public listings, OpenAI in 2027 and Anthropic within a couple of months. Leaked financials show both losing billions on research and scaling. Deep cuts to their main revenue line are an unusual move ahead of an IPO, and a prospectus would show how much margin remains on inference.
Third, the pressure is spreading to the edge. Block says its Dessa voice model runs locally on a consumer laptop, and TechCrunch notes this could let it power more products at lower cost. The model is not publicly available yet.
If competition among hosts holds, the floor for token prices is the marginal cost of serving them, not the price a lab would like to charge. That is good for buyers. It is much less clear that it is good for the companies selling the models.