Key takeaways
- Mid-tier and budget model pricing fell 35.8% year over year to a median of $1.00 per million input tokens as of 24 July 2026, across 130 tracked models.
- Frontier pricing is reported to have doubled since January 2026. Agent workloads overwhelmingly run on frontier models.
- Blended token cost fell to roughly $1.16 to $1.18 per million in early August 2026, down 43% from $2.04 at the end of May. That average hides the split.
- The cheapest listed paid API dropped to $0.02 per million input tokens. A fifty-fold gap between floor and frontier now exists inside the same market.
- If your bill is rising while the headlines say prices are falling, both are true. Check which half of the market your agents actually use.
Does ChatGPT recommend your competitor instead of you?
Check how AI assistants describe your company, and whose name they give when someone asks for a recommendation in your category. Run an audit today, from $19.
Between the end of May and early August 2026, the blended cost of processing a million tokens fell from $2.04 to roughly $1.17. A 43% drop in ten weeks, and the number every summary quoted.
It is also close to meaningless for anyone running agents, because the average is describing two markets that moved in opposite directions.
What actually happened to prices?
The floor collapsed and the ceiling rose.
| Tier | Reported price per million input | Direction in 2026 | What runs there |
|---|---|---|---|
| Cheapest listed paid API | $0.02 to $0.03 | Falling | Classification, extraction, routing |
| Median across 130 models | $1.00 | Down 35.8% year over year | Summarisation, drafting, simple tools |
| Frontier | Substantially higher | Reported doubled since January | Multi-step agents, tool selection, planning |
Read the last column. The work agents actually do sits in the row that got more expensive.
Why does the average look so good?
Because a blended average tracks where the volume went, not what any given call costs.
An enormous amount of new inference volume in 2026 is cheap, small-model work: classification, extraction, embeddings, routing. That volume drags the mean down whether or not a single vendor cut a single price. Meanwhile a fixed multi-step agent workload, unchanged in shape, can cost more this quarter than last.
Both statements are true at once. Only one of them is on your invoice.
The 80% cut that is not a general trend
Late in July 2026 OpenAI reduced GPT-5.6 Luna pricing by about 80%, to $0.20 per million input and $1.20 per million output. Real, large, and specific to one model family. Reductions like this get reported as evidence of a falling market. They are better read as a vendor moving one product into a different tier.
Why does an agent bill grow even on cheaper tokens?
Three multipliers sit between the rate card and the invoice, and none of them appears in a price comparison.
- Frontier dependency. Tool selection and planning are exactly the steps that degrade when you downgrade the model, so they stay on the expensive tier.
- Induced demand. Cheaper tokens make longer runs affordable, so teams run longer. The per-token price falls and the token count rises faster.
- Waste. Retries, coordination between agents and re-sent context all multiply tokens per completed task. A price cut applies to all of them equally, including the ones that produced nothing.
The third is the one you can act on this quarter without changing a model.
What should you actually measure?
Cost per completed task, split by step. Not cost per million tokens, and not a monthly total.
| Metric | Answers | Hides |
|---|---|---|
| Price per million tokens | What the vendor charges | Everything about your workload |
| Monthly spend | Whether finance will call | Whether the money bought anything |
| Cost per completed task | Unit economics of the service | Which step is responsible |
| Cost per step, per completed task | Where to route to a cheaper model | Nothing useful, this is the one |
Once you have the last row, model routing stops being an architecture debate and becomes arithmetic. Steps whose output is checked by a later step can usually run on a cheaper model. Steps whose output nobody checks cannot.
The question worth asking at your next cost review
Not “can we get cheaper tokens”. Ask: what share of last month’s frontier spend went to steps whose output was never read by a human or verified by another step? That number is usually larger than anyone expects, and it is the only one a price cut cannot fix.
Frequently asked questions
Are AI token prices going up or down in 2026?
Both, depending on tier. Trackers report mid-tier and budget models down 35.8% year over year while frontier pricing has doubled since January 2026. The blended average fell 43% between the end of May and early August, which reflects volume shifting to cheap models rather than any single price falling.
How much does a million tokens cost in September 2026?
The reported median across 130 tracked models as of late July 2026 was $1.00 per million input and $4.00 per million output. The cheapest listed paid options sat near $0.02 to $0.03 per million input. Frontier models cost substantially more and have been rising.
Why is my AI bill rising when token prices are falling?
Three reasons compound. Agents run on frontier models, where prices rose. Cheaper tokens invite longer runs and larger context. And retries, coordination and re-sent context multiply the number of tokens a single completed task consumes.
Should you move agent workloads to cheaper models?
Only where the task tolerates it. The usual pattern is a small model for high-volume, well-specified steps and a frontier model for the reasoning that decides the outcome. Route by step, not by workload, and measure cost per completed task rather than per call.
Method and sources
- LLM pricing trackers publishing median API prices across 130 models as of 24 July 2026: $1.00 per million input, $4.00 per million output, with mid-tier and budget models down 35.8% year over year.
- Reported blended token processing cost of roughly $1.16 to $1.18 per million in early August 2026, against $2.04 at the end of May 2026.
- Reported OpenAI pricing reduction on GPT-5.6 Luna in late July 2026 to $0.20 per million input and $1.20 per million output.
- Lowest listed paid API prices: Qwen3.7 Flash at $0.03 input and $0.13 output per million as of 3 September 2026, and Llama 3.1 8B Instruct at $0.02 per million input as of 12 September 2026.
- All figures come from public pricing trackers and 2026 market write-ups. We have not independently verified vendor rate cards and cite these as reported. Prices change without notice.