LLM Pricing Index
Short answer: token prices are down — but only on two of the eight rate cards we track, and both of those cuts are promotional. The index stands at 87.1 against its 2026-08-09 baseline of 100.0, a fall of 12.9% across 80 days and ten weekly checks. Six of the eight rates have not moved once. On the providers' published end dates the index climbs back to exactly 100.0 on January 1, 2027. Enter your own monthly volume below to read the index against your mix.
Read the index against your own volume
Two inputs: the input and output tokens you send each month, run on each of the eight tracked models. The basket never changes — only the rate cards do.
| Model | Rate today / 1M | Bill today | Bill at 2026-08-09 rates | Change vs baseline | Bill after scheduled change | Exposure / mo |
|---|
How the index is built
An index is only useful if the thing being measured never changes. Here that thing is a fixed basket of tokens, and the only variable is the published rate card.
Index level by checkpoint
Five readings: the baseline, the two real changes, and the two changes already published with a future date. The series ends where it started.
| Date | What changed | Basket | Index |
|---|---|---|---|
2026-08-09 Baseline | First verified snapshot of all eight rate cards | $40.30 | 100.0 |
2026-08-16 Real change | Gemini 3.6 Flash halved to $0.75/$3.75 | $38.61 | 95.8 |
2026-08-21 Real change · today | GPT-5.6 Sol cut to $4.00/$20.00 | $35.11 | 87.1 |
2026-11-21 Scheduled | GPT-5.6 Sol promo ends — back to $5.00/$30.00 | $38.61 | 95.8 |
2027-01-01 Scheduled | Gemini 3.6 Flash promo ends — back to $1.50/$7.50 | $40.30 | 100.0 |
The two real changes are worth −4.2 and −8.7 index points respectively, summing to the 12.9-point fall — a net change of -12.9% on the basket. The two scheduled changes are worth exactly the same amounts back, adding $1.69 – $3.50 a month per model line, so on the published dates the index closes a perfect round trip. Promotional end dates can be extended or withdrawn by the provider; we re-check them weekly.
Which rate cards actually moved
The reference basket priced model by model: what each line cost at the baseline, what it costs today, and what it costs if the published end date holds.
| Model | 2026-08-09 rate / 1M | Rate today / 1M | Basket then | Basket now | Change | If promo lapses |
|---|---|---|---|---|---|---|
GPT-5.6 Sol OpenAI · flagship | $5.00 / $30.00 | $4.00 / $20.00 | $12.50 | $9.00 | -$3.50 (-28.0%) | $12.50 from 2026-11-21 |
GPT-5.6 Terra OpenAI · mid | $2.00 / $12.00 | $2.00 / $12.00 | $5.00 | $5.00 | unchanged | — |
GPT-5.6 Luna OpenAI · budget | $0.20 / $1.20 | $0.20 / $1.20 | $0.5000 | $0.5000 | unchanged | — |
Claude Opus 5 Anthropic · flagship | $5.00 / $25.00 | $5.00 / $25.00 | $11.25 | $11.25 | unchanged | — |
Claude Sonnet 5 Anthropic · mid | $2.00 / $10.00 | $2.00 / $10.00 | $4.50 | $4.50 | unchanged | — |
Claude Haiku 4.5 Anthropic · budget | $1.00 / $5.00 | $1.00 / $5.00 | $2.25 | $2.25 | unchanged | — |
Gemini 3.6 Flash Google · mid | $1.50 / $7.50 | $0.75 / $3.75 | $3.38 | $1.69 | -$1.69 (-50.0%) | $3.38 from 2027-01-01 |
Gemini 3.5 Flash-Lite Google · budget | $0.30 / $2.50 | $0.30 / $2.50 | $0.9250 | $0.9250 | unchanged | — |
Two movers, six hold. GPT-5.6 Sol took 20% off input and 33.3% off output; Gemini 3.6 Flash took 50% off both. Everything else — both other OpenAI models, all three Claude models and Gemini 3.5 Flash-Lite — has held the same per-1M rate across ten weekly checks. Of the two movers, Sol carries $3.50 (38.9%) of exposure per line if its promo lapses and Gemini 3.6 Flash $1.69 (100.0%).
Six of eight rates have not moved in 80 days
The headline number is a 12.9% fall. The more useful number is the count behind it: two changes, six non-changes. List prices on frontier models are far stickier than the pace of model releases suggests — across ten weekly verification passes, only two cards moved.
One of the six non-changes was nearly a change. Anthropic had scheduled a September 1, 2026 increase on Claude Sonnet 5 to $3.00/$15.00 and then cancelled it; the introductory $2.00/$10.00 became the standard rate. Had it landed, that single line would have added $2.25 to the basket — 5.6 index points, more than Gemini's entire promotional cut is worth — and the index would read 92.7 today instead of 87.1.
That is the practical lesson from the series: for any given workload, one cancelled increase on the model you actually use outweighs every promotional cut on the models you do not. Budget against standard rates, then treat a promo as upside — see How LLM Token Pricing Works for the standard-versus-promotional distinction.
The index level is not one number
The two cuts were not symmetrical. GPT-5.6 Sol's output rate fell 33.3% while its input rate fell only 20%, so where your tokens sit decides how much of the fall you actually see.
| Workload shape | Basket at 2026-08-09 rates | Basket today | Index |
|---|---|---|---|
Input only 1M in, 0 out per model | $17.00 | $15.25 | 89.7 |
Reference mix 1M in, 250K out per model | $40.30 | $35.11 | 87.1 |
Output only 0 in, 250K out per model | $23.30 | $19.86 | 85.2 |
20x volume 20M in, 5M out per model | $806.00 | $702.25 | 87.1 |
Output-heavy workloads gained the most, and volume changes nothing at all: 20x the tokens produces 20x the dollars and the identical 87.1. Read your own shape with the calculator at the top of this page. The split between your input and output tokens is the single biggest driver of that reading — see Input vs Output Tokens for how to measure it.
The most expensive model changed hands
A price index says nothing about the spread underneath it, and the spread moved in a way the headline does not show. At the baseline GPT-5.6 Sol was the dearest line in the basket at $12.50 and GPT-5.6 Luna the cheapest at $0.5000 — a 25.0x gap. Sol's cut dropped it to $9.00, behind Claude Opus 5 at $11.25.
So the dearest model in the panel is no longer OpenAI's flagship: it is Claude Opus 5, and the spread has narrowed to 22.5x. Both revert when the promotions lapse — Sol returns to $12.50, Opus 5 goes back to second place and the gap reopens to 25.0x.
| Point in the series | Dearest line | Cheapest line | Spread |
|---|---|---|---|
| 2026-08-09 baseline | GPT-5.6 Sol · $12.50 | GPT-5.6 Luna · $0.5000 | 25.0x |
| Today | Claude Opus 5 · $11.25 | GPT-5.6 Luna · $0.5000 | 22.5x |
| After both promos lapse | GPT-5.6 Sol · $12.50 | GPT-5.6 Luna · $0.5000 | 25.0x |
Model choice still dwarfs every price move in this series. The 12.9% fall is real, but the difference between the cheapest and dearest line for the same work is 22.5x — see LLM API Pricing Comparison and How to Reduce LLM API Costs.
What the index does not measure
A short series with a fixed panel has sharp edges. These are them.
Pricing index questions
Are LLM API prices going up or down?
Down, but narrowly. Against our first verified snapshot on 2026-08-09 the index stands at 87.1, a fall of 12.9% over 80 days. Only two of the eight rate cards moved: GPT-5.6 Sol fell from $5.00/$30.00 to $4.00/$20.00 per 1M on 2026-08-21, and Gemini 3.6 Flash halved from $1.50/$7.50 to $0.75/$3.75 on 2026-08-16. Both are promotional, and on the providers' published end dates the index returns to exactly 100.0 on 2027-01-01.
How is the index calculated?
One fixed basket: 1,000,000 input tokens and 250,000 output tokens, run once on each of the eight models we verify and summed. That basket cost $40.30 on 2026-08-09 and is defined as 100.0. The identical basket costs $35.11 today, so the index reads 87.1. Nothing in the basket changes between checkpoints except the published per-1M rates.
Which models have actually changed price?
Two of eight. GPT-5.6 Sol on 2026-08-21 and Gemini 3.6 Flash on 2026-08-16. The other six — GPT-5.6 Terra, GPT-5.6 Luna, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5 and Gemini 3.5 Flash-Lite — have held the same per-1M rates across ten weekly checks and 80 days.
Does the index level depend on my workload?
Yes. The cuts were not uniform across the two axes. An input-only basket reads 89.7, the reference 1M-in/250K-out mix reads 87.1 and an output-only basket reads 85.2 — output-heavy workloads gained more, because GPT-5.6 Sol's output rate fell 33.3% while its input rate fell 20%. The index is scale-invariant: 20x the volume moves the dollars 20x and leaves the level unchanged.
What happens when the promotions end?
GPT-5.6 Sol returns to $5.00/$30.00 on or after 2026-11-21, putting $3.50 back on its basket line — a 38.9% rise on that line. Gemini 3.6 Flash returns to $1.50/$7.50 on 2027-01-01, doubling its line from $1.69 to $3.38. Together they take the basket from $35.11 to $40.30, a rise of 14.8%, and the index back to 100.0.
Why is DeepSeek not in the index?
Because the index is a fixed basket of exactly the eight rate cards in our data file, each re-checked against official provider documentation every week; adding a card mid-series would break the series. DeepSeek is now published as a separate rate card on the DeepSeek API Pricing page because its peak/off-peak and cache-hit/miss tiers do not fit the single-rate-card model, but it is not part of the weekly-verified panel. Prompt-caching, batch and long-context rates are excluded for the same reason — see Methodology.
Does this page upload my token volumes?
No. Your input and output token counts stay in this browser tab. There is no form, no account, no API call and no storage — the arithmetic runs against the same published per-1M rates the rest of the site uses.
Where these numbers come from.
The basket is fixed at 1,000,000 input tokens and 250,000 output tokens per model and never changes. Basket cost is the sum over the eight tracked models of input tokens ÷ 1,000,000 × the input rate plus output tokens ÷ 1,000,000 × the output rate; the index is that basket divided by the same basket priced at the 2026-08-09 baseline, times 100. The baseline card and the two scheduled cards are stored as history in the same data file as the live rates and are never mixed with them: current rates alone drive every other page on this site. Current rates are re-verified weekly against official provider documentation — OpenAI 2026-08-23, Anthropic and Google 2026-08-09 to 2026-08-16. Caching, batch, long-context surcharges, free tiers, subscriptions, taxes and volume discounts are excluded throughout.
Official sources: OpenAI pricing, Anthropic pricing, Google Gemini API pricing.