416 AI Models Priced Daily: The USTA AI Price Index

The price of talking to an AI model is a posted string on a marketplace, not a rumor. On August 24, 2026 the USTA AI Price Index sealed 416 models from 53 providers as they were listed that day, and this page is a census of those listings — not a ranking of which model to buy.

Models listed publicly on OpenRouter with their posted per-token prices, plus the Hugging Face trending list, as captured by US Tech Automations’ sealed daily AI-economics snapshots. This is a census of one marketplace’s listings, not of every AI model in existence.

Prices below are quoted per million tokens, converted in code from the verbatim per-token decimal strings the source publishes. Nothing is rounded to make a point. The July AI Price Index is a separate sealed day; this August page does not subtract from it and does not write a trend.

A price index is not a recommendation; it is a snapshot of what a marketplace posted.

416 AI models from 53 providers were priced on August 24, 2026.

Key Takeaways

• 416 models from 53 providers on the August 24, 2026 freeze.

• 392 listings carry a paid price, 19 are free, and 5 are variable.

• Median paid prompt is $0.51 per million tokens; median paid completion is $2.00.

• Budget tier holds 184 models (median prompt $0.15); frontier holds 39 (median prompt $10).

• Cheapest paid listing is IBM: Granite 4.0 Micro at $0.017; most expensive is OpenAI: o1-pro at $150 prompt and $600 completion.

One Marketplace, One Sealed Day

The number is deliberately literal. It is not an estimate of how many models really exist, and it is not a curated shortlist of the ones worth using. It is a count of distinct public listings on a single marketplace on a single day — a floor you can stand on, not a ranking you have to trust.

The 53 providers behind those 416 listings range from household-name labs to single-model open-source shops, and the index treats them all the same way: whatever price a provider posts is the price we record. No provider is weighted, promoted, or filtered for prominence.

This edition is cross-sectional: a single sealed day, with no trend claim attached. A prior June edition of the AI Price Index used the same methodology, and so did the July AI Price Index. We make no cross-edition comparison here — prices move daily, and two snapshots are not a trend.

The tier cutoffs later in this piece are our own methodology choices, not the marketplace's. Everything else — model counts, posted prices, context windows — is copied from the source. Where we make a judgment call, we say so and put the cutoff in the sealed display set so it can be checked.

Paid, Free, and Variable Listings

Not every listing carries a fixed price. Of the 416 models, most post a flat paid rate, a handful are free, and a few use variable pricing that changes by request.

392 of the 416 models carry a posted paid price.

The 392 paid models are the ones this index can rank on cost; the 19 free and 5 variable listings are noted for completeness but sit outside the price tables. That distinction matters — a free model still costs compute somewhere, and a variable price cannot be pinned to a single per-token number on a given day.

Put differently, 392 of the 416 listings are directly comparable on cost. The remaining free and variable entries are not. That the comparable set is so large is why an index built on posted prices can say something useful about the marketplace rather than dissolving into caveats.

Budget, Mid, and Frontier Bands

Sorting the paid models by their prompt price splits them into three bands. The cutoffs — up to $0.50 per million, then up to $5.00 — are ours; the counts and medians are the marketplace's.

The frontier tier holds 39 models at a $10 median.

The $0.50 and $5.00 cutoffs are our methodology choices, not the marketplace's.

The spread is the story. The budget tier's median prompt price is $0.15 per million tokens; the frontier tier's is $10. That is a wide gap for what is nominally the same task — sending a prompt and reading a reply — and it is why which model is a budget decision, not just a quality one. The 184 budget models outnumber the 169 mid listings, and the frontier tier is a thin 39-row band at the top.

The shape of that distribution is a useful signal for anyone building on these models. The bulk of the marketplace has settled into the budget and mid bands, where a $0.15 or $1.00 median prompt price makes high-volume automation affordable. The frontier tier, at a $10 median, is where you pay for the hardest reasoning — and where the index earns its keep by making that premium explicit.

The Cheapest and Most Expensive Listings

The extremes make the range concrete. The overall paid median sits at $0.51 per million prompt tokens and $2.00 per million completion tokens, but the endpoints are far apart.

The median paid model costs $0.51 per million prompt tokens.

From $0.017 per million at IBM: Granite 4.0 Micro to $150 prompt and $600 completion at OpenAI: o1-pro, the marketplace spans several orders of magnitude. A workflow that would cost pennies on Granite 4.0 Micro could cost real money on o1-pro for the identical volume of tokens. The median is a useful anchor, but it hides how far the tails reach.

It is also worth separating the two prices most models charge. Prompt tokens (what you send) and completion tokens (what the model writes back) are often priced differently — the median completion rate of $2.00 per million sits well above the $0.51 prompt rate. For output-heavy workflows, that completion price, not the headline prompt price, is the number that governs the bill.

Context Windows on the Same Freeze

Price is only half the decision; context window is the other half. The median model accepts 262,144 tokens of input, and the ceiling is far higher.

137 models advertise a 1,000,000-token context or larger.

A 262,144-token median means the typical listed model can already hold a long working document in memory. At the top, SpaceXAI: Grok 4.20 Multi-Agent advertises a 2,000,000-token window, and 137 models now clear the 1,000,000-token floor. Long context is no longer a frontier-only feature; it has spread well down the price ladder.

That spread has a practical edge: a workflow that needs to reason over a long document is no longer forced into the $150 tier just to get the window it requires. With 137 models at 1,000,000 tokens or more, context length and cost can now be chosen independently rather than bundled together.

The same sealed-snapshot discipline sits behind our Los Angeles building-permit report. Different dataset, same rule: publish the freeze, do not invent a trend from one day.

Hugging Face Attention, Not Price

Alongside the price census, the index captures the Hugging Face trending list — a rough read on what the open-model community is downloading right now, independent of price.

The top three trending entries on the sealed day were Qwen/Qwen3.8-27B, unsloth/Qwen3.8-27B-GGUF, and orcarouter/Qwen3.8-27B-Uncensored-MLX. Trending rank is attention, not endorsement: it reflects downloads and buzz, not a quality or safety judgment from us. We list it because attention often precedes availability — today's trending open weights are frequently tomorrow's hosted, priced listings.

All three names sit in the same Qwen3.8-27B family. That is a popularity cluster on another surface, not a price signal, and it is not a claim about OpenRouter listing order.

How the Index Is Sealed

All figures are computed directly from US Tech Automations’ sealed daily AI-economics snapshots; nothing is estimated, modeled, or extrapolated. Prices are per million tokens, converted in code from the verbatim per-token decimal strings the source publishes. The tier cutoffs are our methodology choices and appear in the display set.

This edition is cross-sectional only — a single sealed day. The AI-economics clock reads the OpenRouter public model listing and the Hugging Face trending list daily and content-hashes each sealed edition, so any figure here can be traced back to the exact listing that produced it. No change claims until the clock holds multiple monthly observations on a dedicated movement page.

1. Fetch. Read the OpenRouter public model listing and the Hugging Face trending list on the clock day.

2. Convert. Keep the verbatim per-token strings the source publishes and convert them in code to per-million prices.

3. Tier. Apply the sealed cutoffs ($0.50 budget max, $5.00 mid max) without moving a listing by hand.

4. Seal. Content-hash the edition so every published figure traces to one immutable daily capture.

Frequently Asked Questions

Q: How many AI models does the index track?
A: 416 models from 53 providers as of August 24, 2026, drawn from OpenRouter's public listings. Of those, 392 carry a paid price, 19 are free, and 5 use variable pricing. It is a census of one marketplace, not every model in existence.

Q: What does the median model cost?
A: The median paid model is $0.51 per million prompt tokens and $2.00 per million completion tokens. Prices run from $0.017 per million at IBM: Granite 4.0 Micro up to $150 prompt and $600 completion at OpenAI: o1-pro.

Q: How are the price tiers defined?
A: The tiers are our own methodology, drawn on prompt price: budget up to $0.50 per million (184 models), mid from $0.50 to $5.00 (169 models), and frontier above $5.00 (39 models). The cutoffs appear in the sealed display set so they can be checked.

Q: Is this a trend or a snapshot?
A: A snapshot. This edition is cross-sectional — a single sealed day — and makes no trend claim. Prices are converted in code from the verbatim per-token decimals the source publishes; nothing is estimated or modeled.

Q: How much context do these models offer?
A: The median context window is 262,144 tokens, and the largest is 2,000,000 tokens (SpaceXAI: Grok 4.20 Multi-Agent). 137 models advertise a 1,000,000-token window or larger.

Q: Do Hugging Face trending ranks tell you which model is cheapest?
A: No. Trending is popularity on another surface. The three names on this freeze — Qwen/Qwen3.8-27B, unsloth/Qwen3.8-27B-GGUF, and orcarouter/Qwen3.8-27B-Uncensored-MLX — are attention labels, not prices.

Put Cost Data to Work

An operations manager pricing an AI-in-the-loop workflow can treat the $0.51 median prompt and $2.00 median completion as the planning anchors, then decide whether a given job belongs in the 184-model budget band or the 39-model frontier band. Model choice is often the largest variable cost in the run; this freeze exists so that choice is made on posted numbers.

A founder comparing a cheap listing to a famous one can put IBM: Granite 4.0 Micro at $0.017 next to OpenAI: o1-pro at $150 / $600 and ask whether the job actually needs the frontier tail. Most automations do not. An analyst watching context can use the 137 models at 1,000,000 tokens as proof that long windows are no longer bundled with the $150 tier.

US Tech Automations builds AI automations, and model choice is a cost decision. This index is the same sealed-snapshot discipline as our permit research. See the agentic workflow platform.

Source: US Tech Automations Research — computed from the sealed daily AI-economics snapshot, August 24, 2026.