The LLM Price Index rose 24% in September. No model in it got more expensive.

The LLM Price Index rose 24.2% in September. That's the biggest monthly move it has ever made, and the first time it has gone up.

Not one model in it got more expensive.

Both of those statements are true, and the gap between them says a lot about how to read any "average price" for LLMs.

The month in one table

The LLM Price Index from ModelPriceWatch is the equal-weight average of ten flagships' blended price per million tokens (3 parts input to 1 part output, list prices as printed on each vendor's own pricing page), one model per lab.

Net for the month: +24.2%. Since the first reading in February: from −9.4% to +12.5%.

Two other seats changed hands (Muse Spark 1.2 → 1.3, Grok 4.6 → 4.7), both at their predecessor's price, so they moved nothing.

The check that changes the story

Take the ten models that were in the index on September 1 and price them at their October 1 list prices.

They average $4.14. Exactly what they averaged on September 1.

None of them changed its own price during the month. Sort both baskets by price and nine of the ten numbers are identical: the median was $3.00 on both dates, and the nine cheapest averaged $3.49 on both. The only difference is the top price: $10.00 then, $20.00 now. GPT-6 Astra alone is $2.00 of the $5.14.

So the +24.2% isn't "LLMs got more expensive in September." It's "one seat in the basket moved up a tier."

Meanwhile, every successor launched at the same price or cheaper

Here's what September's releases looked like when they replaced a model in the same line:

*Against GPT-5.6 Sol's promotional $4/$20.

Ten successors from the index's labs: four cheaper, six flat, none higher.

The only step up was GPT-6 Astra at $10/$50. It isn't a successor to anything. It's a new tier above Sol, and OpenAI calls it its default flagship, which is why the index seats it for OpenAI. In the same month OpenAI shipped two cheaper Sol models (GPT-6 Sol and GPT-6.1 Sol, both $2/$10). Had the index seated the Sol line instead, October 1 would have read $3.54, not $5.14.

That's $1.60 of the index level riding on one judgment call about which OpenAI model counts as the flagship. Astra is the defensible choice, since it's what OpenAI points you to, but it's worth knowing the number is that sensitive to it.

The practical version: if you moved to the newest model in your line, you paid the same or less in September. If you're still calling GPT-5.6 Sol at $4/$20, GPT-6 Sol is half that.

Anthropic and OpenAI cut cache prices instead

The index blends input and output list prices. It ignores cached input. In September that's where Anthropic and OpenAI cut:

• Until September 1, no Anthropic or OpenAI model on the tracker priced cache reads below 10% of its input rate. (DeepSeek V4 Pro already did, at 3.3%.)
• Claude Fable 5.1 launched at 2.5% ($0.25 vs $10 input).
• Claude Opus 5.5: 5% ($0.20 vs $4).
• GPT-6.1 Sol: 5% ($0.10 vs $2).
• By October 7, Anthropic had halved Sonnet 5.5's cache rate to the same 5%.

If you run an agent or a chat product that resends a long system prompt and history on every turn, the cache rate can move your bill more than the headline input price. None of it shows up in a blended-price index.

The cheapest GPT-4-class model: the number fell, but read the footnote

The index has a companion line, the floor: the cheapest model that clears a fixed capability bar (GPQA Diamond ≥ 70, externally measured, roughly GPT-4-class reasoning). It went from $0.113 to $0.090 per million tokens on September 30, and the gap to the flagship ceiling widened from 36.6x to 57.1x, the widest on record.

The footnote: most of that is a change in which listing is tracked, not the market. DeepInfra lists DeepSeek V4 Flash twice: an undated entry at $0.09/$0.18 that held all of September, and a dated "0731" entry that it calls the official release. On September 30 the floor switched to the dated entry, now $0.06/$0.18. DeepInfra did cut that entry at some point; the cut can only be placed between August 15 and September 30. Without the switch, the gap would have been 45.5x.

The part that isn't a footnote: DeepSeek retired V4 Flash from its own API on September 10. So the cheapest GPT-4-class capability on the market is now a third-party host's copy of a model its maker no longer sells. The September report called that fragile. It's more fragile now. If DeepInfra dropped it, the floor would jump to about $0.13, and the cheapest model a lab sells itself that clears the bar is $0.45.

What to take from this

• Ask what an "average price" is averaging. A 24% jump can be one seat changing occupant. Price a fixed basket before you believe a trend.
• Check whether you're on an old SKU. Same-line successors launched flat or cheaper all September. GPT-5.6 Sol to GPT-6 Sol is a 50% list-price cut, if the newer model works for your use case.
• Price your workload, not the rate card. If your prompts repeat, the cached-input rate is where September's discounts went.
• Know who sets your floor. If your cheapest model is a host's copy of a retired model, have a fallback priced in.

Every figure above is as of October 1, 2026, and is fixed in the dated report at modelpricewatch.com/reports/state-of-llm-pricing-2026-10. The live index is at modelpricewatch.com/price-index, with the methodology and the full price history dataset on Hugging Face (CC-BY-4.0). Every price links to the vendor page it came from, with a capture timestamp.

Disclosure: published by ModelPriceWatch.

If you build on OpenAI: when GPT-6 Astra became the default, did your usage move up to it, or stay on Sol? We're trying to work out which of the two numbers ($5.14 or $3.54) is closer to what teams actually pay.