Falling LLM Token Prices, Rising AI Bills

Token prices collapsed by orders of magnitude, yet Uber burned a year of AI coding budget in four months

Published 27 October 2026 Updated 27 October 2026

Falling LLM token prices are real and steep, and they still may not decide what you pay. The price per million tokens shrinks almost every month, the menus get cheaper, and yet Uber burned through a full year's AI coding budget in four months. One thing decides your real bill, and it isn't the token price.

Cheap tokens, bigger bills

You pick a model from a pricing page, and the per-token number keeps going down. Then the invoice arrives bigger than last quarter's. That gap between the sticker price and the bill is the whole story of this article: the price of a unit fell, the number of units did not.

Falling LLM Token Prices: how far they really dropped

The crash is genuine. In March 2023, GPT-4 cost about $36 per million tokens. About a year and a half later, GPT-4o cost about $4. DeepLearning.AI's newsletter The Batch reads that as a cut of roughly 79% a year.

The fall is not even across tasks. Epoch AI measured what it costs to reach GPT-4-level performance and found that cost fell anywhere from 9x to 900x per year, depending on the task. On PhD-level science questions, it was roughly 40x a year. That is Epoch AI's measurement, not a vendor claim, and it means how much you save depends on what you ask the model to do.

The menus today are wide. OpenAI's own pricing page puts its cheapest model, GPT-6-Luna, at $0.10 per million input tokens, and its premium GPT-6-Astra at 100 times that. Anthropic's cheapest tier, Haiku, also starts at $0.10 per million input tokens. For most teams, picking the smallest model that does the job saves more than waiting for the next price drop.

One detail most charts skip: Google lists its flagship Gemini 4 Argon at a held price only through November 30, 2026. After that date the price doubles. Prices do not only fall.

DeepSeek's discount meets demand

What pushes a price back up is demand. In May 2026, DeepSeek made a roughly 75% cut on its flagship V4-Pro model permanent. It had started as a temporary promotion, and developers read the permanent version as a new floor.

Then the floor cracked. On July 31, DeepSeek released V4-Flash at $0.14 per million input tokens. Demand surged past what its capacity of roughly 20,000 H100-equivalent GPUs could serve reliably, and within a week the company warned developers to expect a significant price increase without saying how much.

That is the catch with the cheapest number on the board: it is a price, not a promise. If your product's margin rests on it, you need a plan for the day the discount is gone, or the day the API simply cannot answer.

Developers chase the cheap end anyway. OpenRouter's own routing data shows Chinese open-weight models grew from about 2% to 61% of the tokens it routes over roughly a year and a half, and those models settled 60% to 90% cheaper per token than comparable US closed models.

Race to the bottom, or too cheap to meter

Where this ends is contested. In October 2024, Marc Andreessen said LLM providers may be in a race to the bottom, comparing selling intelligence to selling rice: open models and a constant stream of new entrants, he argued, keep margins from holding.

Sam Altman tells it the other way, repeating OpenAI's goal of intelligence too cheap to meter. That phrase was coined in 1954 for nuclear power, and electricity never became too cheap to meter.

Under the slogan sits the accounting. The Register reported that OpenAI is projected to lose as much as $14 billion in 2026. Developers on Hacker News bring that worry down to their own bills: one widely read thread argues that flat-rate Claude subscriptions hand out far more usage than it would cost at API rates, a subsidy the labs cannot keep paying, while other commenters in the same thread push back by pointing at falling inference costs. Both of those are opinions from that discussion, not measurements. The practical reading: if your price is subsidized, you are building on somebody else's losses, and the first place developers feel that is their subscription.

Why your plan feels smaller

On August 28, 2025, Anthropic added weekly rate limits to Claude Pro and Max. The company said fewer than 5% of users would be affected, and it cited one user who consumed tens of thousands in model usage on a $200 plan. Both of those figures are Anthropic's own.

Then came the confusion. In a GitHub issue on Claude Code, more than 30 users reported that after a new Sonnet model shipped, their Pro weekly allowance fell from an advertised 40 to 80 hours to roughly 6 to 8. Those are user reports; Anthropic never confirmed in that thread whether the cut was intentional.

Notice what nobody in that thread was arguing about: the price of a token. They were arguing about how much they could use. At company scale, that same quantity becomes a budget line that explodes.

Agents eat the savings

Uber gave its engineers Claude Code in late 2025. By around April 2026, it had burned through its entire annual AI coding budget in four months, and it now caps spending at $1,500 per employee per month per tool.

Nobody's tokens got more expensive. The agent got good enough that everyone used it all day. Simon Willison calls 2026 a tokenmaxxing cycle: companies mandated AI adoption, and Uber said about 90% of its engineers used it monthly. The spending caps arrived months later, once total usage rather than the per-token price became the real cost.

Microsoft drew its own line, telling employees to stop using Claude Code by June 30, 2026 and move to GitHub Copilot CLI. The lesson is not about one tool. Big buyers steer spending toward whatever is cheapest in total, not cheapest per token.

Ukraine's sovereign option

Sometimes the question is not how cheap, but where your data goes. Around May 14, 2026, Ukraine's Lapathoniia platform opened a closed beta, giving developers an OpenAI-compatible API to two Ukrainian language models, MamayLM and LapaLLM. They run on De Novo's cloud, on hardware physically inside Ukraine. OpenAI-compatible means it speaks the same API format your code may already use.

It was co-founded by Bohdan Didenko and Vladyslav Melnyk, with academic support from Lviv Polytechnic's Department of Artificial Intelligence Systems. The stated goal is to let Ukrainian developers and state institutions use language models without sending data out of the country.

Now the catches, and there are several. The main one: it is a closed beta, so you cannot count on getting access today. The published material does not list its prices, its rate limits, a public sign-up address, or how its models score on any benchmark, so treat the cost question as unanswered rather than assume it is cheap or free; hosted models are served by somebody's hardware and somebody pays for it. You need no hardware beyond an API client, because the models run on De Novo's cloud rather than your machine.

Use it instead of a US cloud API when your data has to stay in Ukraine. Don't use it if what you need is the cheapest frontier model on the board.

What sets your bill

Volume decides your real bill, not the token price: how many tokens your workflow burns, and whether the price you were quoted survives everyone else burning them too. Per-token prices really did fall by orders of magnitude, and Uber still blew a year's budget in four months.

If you call a model from a product, take the smallest tier that does the job and keep a fallback provider, because DeepSeek showed that a cheap price can be pulled. If you live on a flat-rate subscription, treat the limits as the real price, because Anthropic has shown it will change them. If you run agents across a team, budget by usage per person, the way Uber now does, rather than by the price sheet. And if your data cannot leave Ukraine, watch Lapathoniia, knowing it is still a closed beta.

Tokens got cheap. Nobody ever buys just one.

Our Short on this: deepseek-ai/DeepInfer.

Sources