LLM Price Index Up 24% in a Single Month — Yet Not One Model Raised Its Price

A seemingly contradictory statistic has sparked a debate about how to properly read an "average price" metric: ModelPriceWatch's LLM price index rose 24.2% in September — the largest single-month gain in the index's history, and its first-ever increase. Yet at the same time, not a single model in the index raised its own price. Both statements are true, and the gap between them reveals a trap lurking behind any "average price" metric.

What Happened This Month, at a Glance

ModelPriceWatch's LLM price index is an equal-weighted average of the "blended price per million tokens" for ten flagship models (weighted at 3:1 input-to-output, using list prices from each vendor's official website, with one model selected per lab).

September's movement was driven by two "seat changes":

• September 4: OpenAI's seat switched from GPT-5.6 Sol (blended price $8.00) to GPT-6 Astra ($20.00), pushing the index from $4.14 to $5.34.
• September 30: Anthropic's seat switched from Claude Opus 5 ($10.00) to Claude Opus 5.5 ($8.00), pulling the index back to $5.14.

The month closed at $5.14, a net gain of +24.2%. Measured from the index's first reading in February, the cumulative change went from −9.4% all the way to +12.5%.

Two other seats also swapped models (Muse Spark 1.2→1.3, Grok 4.6→4.7), but both successors carried over their predecessors' prices, contributing nothing to the index.

A note on methodology: "Equal-weighted average" and "constituent replacement" are two foundational concepts for understanding this type of index. Equal weighting means every model in the basket carries the same weight — if any one of the ten models doubles in price, the index rises by roughly 10%. "Seat changes" refer to the common index practice of swapping constituents: when a lab releases a new flagship model, that lab's representative in the index is updated. This is the same logic used by stock indices like the S&P 500 when they rebalance their components. The key issue is that a constituent change alone can move the index up or down, completely independent of any actual price adjustment by any individual model. This is why, when analyzing an index, it's important to distinguish between "price changes within a fixed basket" and "price changes caused by constituent replacement" — the former reflects genuine market pricing trends, while the latter mostly reflects the rules of index construction.

One Simple Test That Completely Rewrites the Conclusion

Take the ten models that were in the index on September 1 and reprice them using their October 1 list prices. The average comes out to $4.14 — identical to September 1, down to the cent.

In other words, not one of those ten models changed its price during September. Sort both baskets by price and nine out of ten numbers are exactly the same: the median for both dates is $3.00, and the cheapest nine models average $3.49 on both dates. The only difference is at the top end: $10.00 versus $20.00. GPT-6 Astra alone accounts for $2.00 of the index's $5.14 value.

So +24.2% is not "LLMs got more expensive in September." It's "one seat in the basket was replaced by a model in a higher pricing tier." This is the critical insight for reading any average price metric: movements can come from changes in composition, not from prices themselves moving.

In Fact, Every Successor Model Was Priced the Same or Lower

If you look at how each lab's new September releases priced against the models they replaced in the same product line, the picture is the complete opposite:

New Model	Replaces	Blended Price $/Mtok	Change
GPT-6 Luna	GPT-5.6 Luna	$0.45 → $0.20	−56%
GPT-6 Sol	GPT-5.6 Sol	$8.00 → $4.00	−50%
DeepSeekAd V4.1 Flash	DeepSeek V4 Flash	$0.66 → $0.525	−20%
Claude Opus 5.5	Claude Opus 5	$10.00 → $8.00	−20%
Fable 5.1, Sonnet 5.5, GPT-6.1 Sol, Grok 4.7, Muse Spark 1.3, Gemini 3.8 Flash	Their respective predecessors	Unchanged	0%

Among the ten successors, four cut prices and six held steady. Not one raised its price.

The only true "upward step" was GPT-6 Astra ($10/$50). It's not a successor to any existing model — it's a new tier above Sol that OpenAI designated as its default flagship, which is why the index slotted it into the OpenAI seat. In the same month, OpenAI also launched two cheaper Sol models (GPT-6 Sol and GPT-6.1 Sol, both at $2/$10). Had the index chosen the Sol line instead, the October 1 reading would have been $3.54, not $5.14.

That means $1.60 of the index's value is entirely a function of one judgment call: "which OpenAI model counts as the flagship?" Astra is a defensible choice (it's the officially designated default), but it's worth understanding how sensitive the headline number is to that single decision.

The practical takeaway: If you upgraded to the latest model in your product line in September, you're paying the same or less. If you're still on GPT-5.6 Sol ($4/$20), GPT-6 Sol costs half the price.

Anthropic and OpenAI Slashed Cached Input Prices

The index only blends input and output list prices — it ignores cached input pricing. And the real price cuts in September from both Anthropic and OpenAI landed squarely there:

• Before September 1, no model in the tracked list from either company had a cache read price below 10% of its input price. (DeepSeek V4 Pro had already reached 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 also halved Sonnet 5.5's cache price to the same 5%.

If you're running an agent or a chat product that re-sends long system prompts and conversation history on every turn, cached input pricing can have a bigger impact on your bill than the headline input price. And none of this is visible in a blended price index.

What is cached input pricing? Cached input is a pricing mechanism in LLM APIs: when the same text (typically a system prompt or long context) appears repeatedly across multiple requests, the provider temporarily caches it and charges a lower "cache read price" on subsequent calls instead of the full "input price." For applications that re-send large amounts of fixed content on every request — such as AI assistants with complex system prompts, multi-turn chat products that pass full conversation history, or agents that repeatedly read long documents — cached input pricing can have a far greater impact on actual costs than the stated input price. For example, if an application sends an average of 100K cached tokens per request and the cache price drops from 10% to 2.5% of the input price, that portion of costs falls by 75%. These savings are completely invisible in an index that only blends input/output list prices, which is why "the index didn't fall" and "your bill dropped significantly" can both be true at the same time.

The Cheapest GPT-4-Class Model: The Number Fell, But Read the Footnotes

The index also tracks a companion metric — the floor price: the cheapest model that meets a fixed capability threshold (GPQA Diamond ≥ 70, measured externally, roughly equivalent to GPT-4-era reasoning). It dropped from $0.113 to $0.090 (September 30), widening the spread between the cheapest qualifying model and the flagship ceiling to 57.1× — the widest gap on record.

The footnote: most of this wasn't a market change, it was a change in "which listing to track." DeepInfra listed DeepSeek V4 Flash twice — one entry with no date that showed $0.09/$0.18 throughout September, and one dated "0731" entry marked as the official release. On September 30, the floor switched to the dated entry ($0.06/$0.18). Without that switch, the spread would have been 45.5×.

The part that isn't a footnote: DeepSeek pulled V4 Flash from its own API on September 10. So the cheapest GPT-4-class capability currently available is a third-party-hosted copy of a model the original developer no longer sells. This was already a fragile situation — it's now even more so. If DeepInfra delists it, the floor price would jump to roughly $0.13, and the cheapest model that clears the threshold while being sold directly by a lab is $0.45.

What is GPQA Diamond? GPQA Diamond (Graduate-Level Google-Proof Q&A) is a multiple-choice benchmark designed by PhD-level experts, covering hard sciences including physics, chemistry, and biology. The questions are deliberately constructed to resist direct lookup via web search, testing a model's genuine reasoning and knowledge synthesis abilities. A score of 70 is a commonly used industry reference point, roughly corresponding to the reasoning level of flagship models from the GPT-4 era. The "floor price" tracks the cheapest model that clears this threshold, answering a practical question: what's the minimum cost to get GPT-4-level reasoning? This metric has direct value for developers making model selection decisions — if a task has a minimum reasoning requirement, the floor price is the starting point for budget planning. But as the article reveals, when the cheapest qualifying model comes from a third-party host and the original developer has already discontinued it, the sustainability of that price becomes a risk factor that needs to be built into the decision.

Key Takeaways

• First, ask what the "average price" is actually averaging. A 24% jump might just mean one constituent got swapped out. Before trusting a trend, reprice the fixed basket yourself.
• Check whether you're still on an old SKU. Every September successor in the same product line launched at the same price or lower. If a new model fits your use case, the move from GPT-5.6 Sol to GPT-6 Sol is a 50% price cut.
• Price for your workload, not the price list. If your prompts repeat, September's real price cuts are hiding in cached input pricing.
• Know who's setting your floor price. If your cheapest model is a third-party-hosted copy of a retired model, have a priced fallback option ready before you need it.

All data in this article is as of October 1, 2026, published by ModelPriceWatch and fixed in their dated report. Disclosure: this article was published by ModelPriceWatch.