By J. Wren · Published 2026-09-15 · Updated 2026-10-13
This site checks the published rates for 16 LLM and embedding models once a day and logs anything that changes. This article covers the log for Q3 2026 through 2026-09-14: what changed, when, and what it does to a typical monthly bill. It uses only what the log actually recorded — nothing here is inferred about prices from before tracking started, and nothing is guessed about vendors that did not show up in the log.
Four models changed price in the tracking window, across three distinct dates. That is eight log entries, because each model's input and output rate are recorded as separate entries. All eight are decreases — this log has not recorded a single price increase since tracking began. That is a fact about what happened to be tracked, not a claim that no vendor raised prices anywhere; a model this site does not track could have moved and it would not appear here.
| Date | Model | Vendor | Rate | Change |
|---|---|---|---|---|
| 2026-07-30 | GPT-5.6 Terra | OpenAI | Input | $2.50 → $2.00 ▼ 20% |
| 2026-07-30 | GPT-5.6 Terra | OpenAI | Output | $15.00 → $12.00 ▼ 20% |
| 2026-07-30 | GPT-5.6 Luna | OpenAI | Input | $1.00 → $0.20 ▼ 80% |
| 2026-07-30 | GPT-5.6 Luna | OpenAI | Output | $6.00 → $1.20 ▼ 80% |
| 2026-08-19 | Gemini 3.6 Flash | Input | $1.50 → $0.75 ▼ 50% | |
| 2026-08-19 | Gemini 3.6 Flash | Output | $7.50 → $3.75 ▼ 50% | |
| 2026-08-22 | GPT-5.6 | OpenAI | Input | $5.00 → $4.00 ▼ 20% |
| 2026-08-22 | GPT-5.6 | OpenAI | Output | $30.00 → $20.00 ▼ 33% |
The log holds eight entries across four models in Q3. This site only started tracking on 2026-07-25, so there is no history before that date to compare against, and the sixteen tracked models are a narrow slice of the market. Anthropic and Google's Gemini Pro/Flash-Lite tiers recorded no changes at all in this window — that could mean stable pricing, or it could mean a change happened before tracking started. Treat this as a log of what was observed, not a complete picture of the market.
Percentages are easy to state and hard to feel. Here is the same eight entries converted into
dollars for one defined usage pattern: 2,000,000 input tokens and 500,000 output tokens
per month — roughly the volume of a small production feature doing a few thousand
short exchanges (around 500 tokens each) in a month. The math is
(input_M × in_rate) + (output_M × out_rate),
using the before and after rates from the table above.
| Model | Before | After | Monthly difference |
|---|---|---|---|
| GPT-5.6 | $25.00 | $18.00 | -$7.00 (-28%) |
| GPT-5.6 Terra | $12.50 | $10.00 | -$2.50 (-20%) |
| GPT-5.6 Luna | $5.00 | $1.00 | -$4.00 (-80%) |
| Gemini 3.6 Flash | $6.75 | $3.38 | -$3.37 (-50%) |
At this usage level the absolute savings are small — a few dollars a month — because the volume itself is small. The percentage is what carries over to bigger workloads: scale the same usage pattern up 10x, to 20,000,000 input and 5,000,000 output tokens a month, and the GPT-5.6 change alone goes from a $250.00 bill to a $180.00 bill, a $70.00 monthly difference. The percentage change does not depend on volume; the dollar amount does.
GPT-5.6 Luna's 80% cut on both input and output is the largest single move in the log. A workload built around Luna at its old rate would be paying five times more than necessary at today's rate — worth checking if anything in a running budget still assumes the July numbers.
A vendor's own pricing page only shows what a model costs today. It does not say whether that number changed last week, or six months ago, which makes a few ordinary questions hard to answer without watching closely: is the tier a project was budgeted around still the same price, did a provider get cheaper since the last time it was evaluated, is a "new" low price actually new. An automated daily check and a running log answer those without requiring anyone to have been paying attention at the time.
One guard worth naming: a rate that moves more than 50% in a single day is treated as a probable data error rather than a real price cut and is held back from publishing until checked by hand. GPT-5.6 Luna's 80% cut tripped that guard on 2026-07-30; the resulting failure went unnoticed for five days before anyone checked it, and only after it was confirmed against OpenAI's own pricing page by hand was it published, on 2026-08-04. The guardrail caught the cut correctly — the five-day delay before someone looked was a separate monitoring gap, since fixed by the failure alerting described on the about page.
The table above shows what changed; it is not a current rate card. For today's numbers across all 13 tracked chat models, use the LLM API cost calculator, or go straight to a vendor: Claude, GPT, Gemini. The full, continuously updated version of this log — including any changes recorded after this article was published — lives at the pricing change log.
No rate changes were recorded between 2026-09-15, the day this article was published, and the check on 2026-10-13. The table above is still the whole log: 8 entries on four models, the most recent on 2026-08-22 (GPT-5.6), 52 days before the 2026-10-13 check. Nothing in the figures above has been revised.
A scheduled job checks the rates once a day. Runs are occasionally delayed; from 2026-09-15 to 2026-10-13 a run was recorded on each of the 29 calendar days (UTC) and none failed. How the data is sourced and checked, including the 2026-07-30 failure described above, is on the methodology page. The pricing change log shows the date of the latest check.
Does this cover every LLM price change in Q3 2026? No. It covers only the 16 models this site tracks, and only from 2026-07-25 onward. A change to an untracked model, or one that happened before tracking started, will not appear here.
Why no price increases in the log? Because none were recorded in this window. That is a description of the data, not a claim that prices only ever go down.
Will this page get updated as Q4 changes come in? This article is a snapshot of Q3. New changes continue to accumulate on the live pricing change log, which is the page to check for anything after 2026-10-13.