Tokens fall 10x a year. Bills still go up. Both are true.
a16z calls it LLMflation: the price of a fixed level of LLM capability drops roughly 10x per year, and Epoch AI measures 9-900x depending on the task. Meanwhile frontier list prices have been flat for three years and reasoning tokens inflate per-task cost. A buyer's memo, with the changes we made to our own usage.
Someone in our beta group asked why their API spend went up in the same month three vendors cut prices. Fair question. The answer took me through two pricing studies and ended with us changing how we route our own requests, so here is the whole thing written down.
The half that is falling
a16z named the phenomenon LLMflation: GPT-3-level text generation cost $60 per million tokens in 2021 and $0.06 by late 2024. A thousandfold in three years. Epoch AI measured it independently across multiple benchmarks and got 9x to 900x per year depending on task difficulty, with about 50x for the median task.
Read the fine print on what is being measured, though. Neither study says “model X got cheaper.” They track the cheapest model that clears a fixed capability bar. The engine of the decline is replication: today's flash-tier models write at what was flagship level two years ago, at a few percent of the old price. The curve only pays out if you actually move down it.
The half that is not
Frontier list prices have been close to flat for three years. GPT-4 launched at $30 per million input tokens; each vendor's current strongest model still sits in the dollars-to-tens-of-dollars band (for a live cross-vendor sheet, models.dev maintains an open pricing table). The savings get channeled into the tier below, and the top shelf stays priced like a top shelf.
Reasoning models are the quieter problem. They bill their internal thinking as output tokens, and on a hard problem that can mean 5 to 20 times the consumption of a standard model. Unit price falls while tokens per task rise. For heavy users the second curve often wins, which is exactly what happened to the person who asked the question.
What we changed
After reading both studies we tiered our own routing. Continuation, companion chat, summaries: flash-class models, tenths of a dollar per million tokens. Flagships only for pivotal chapters and long-chain plotting. First month after the switch, same usage intensity, the BYOK bill came in at roughly half. Stack the cache discounts from the companion post and a 3-million-word novel, read end to end with continuations, costs about as much as a takeout meal. We checked that figure twice before publishing it.
Honest footnote: tiering has a failure mode. Flash models occasionally drop plot details in long arcs, and forgetting to switch up for an important chapter means redoing it. We still route by hand. If someone has a clean automatic policy for this, we have not found it.
The one-line conclusion
“Tokens are getting cheaper” is true for buyers who re-evaluate their model choice on a schedule, and false for buyers who default to the newest flagship and never read the bill. Re-evaluating requires being able to switch, which subscriptions prevent and BYOK makes trivial. Foreverse logs model, tokens and cache hits per request, so the audit is yours to run, not ours to claim.
FAQ
Where does the “10x cheaper per year” figure come from?
a16z's LLMflation analysis tracks the cheapest model that reaches a fixed benchmark score: hitting GPT-3-level MMLU (~42) cost $60 per million tokens in 2021 and $0.06 by late 2024, a 1,000x drop in three years, roughly 10x per year. Epoch AI's independent study across multiple benchmarks finds annual declines between 9x and 900x depending on task difficulty, with ~50x for the median.
Then why doesn't my AI bill go down?
Three reasons. Frontier list prices barely fall; the decline happens when new cheap models replicate old capabilities. Reasoning models emit large volumes of billed thinking tokens, raising per-task consumption 5-20x. And usage itself grows; agent workloads fire dozens of calls per task. The dividend goes to people willing to switch models, not to people who always buy the most expensive one.
What do prices look like in mid-2026?
Order-of-magnitude, per official pricing pages: flash-tier models (DeepSeek, Gemini Flash class) charge tenths of a dollar per million input tokens; frontier models from OpenAI, Anthropic and Google sit at a few to tens of dollars per million; every vendor's mini/flash tier runs 10-25x cheaper than its own flagship. Prompt-cache discounts cut effective prices by half or more on top.
How does a regular user capture the decline?
Tier your scenarios: run everyday continuation and casual chat on flash-tier models and reserve the flagship for pivotal chapters and complex reasoning. Flash tiers are already sufficient for most fiction work. And stay switchable with BYOK: the model market reshuffles yearly, and locking into a single subscription forfeits the 10x/year curve entirely.
Questions or ideas? Join our Discord →