Artificial intelligence models explained: four families, four ways to pay
“Artificial intelligence models” is one phrase covering four different products, and the thing that separates them isn’t intelligence. It’s how you get the model, and how the bill is calculated. As of August 2026 you can rent Anthropic’s Claude Opus 5 at $5 per million input tokens, download DeepSeek’s 1.7 trillion parameter V4-Pro under an MIT licence, run Gemini 3.7 Flash for nothing on Google’s free tier, or pay Google $0.40 for one second of generated video. Those are not four tiers of the same product, though the phrase implies that they are.
So the useful way to sort the current families isn’t by benchmark score. It’s by the question a buyer actually has to answer: do you rent the weights, download them, or pay for something that never emits a token at all? Every figure below comes from a vendor page, model card or paper we opened in August 2026, and every one of them is dated, because most of them expire.
The four families, and the question that separates them
The first family is the closed frontier model, which is the one most people picture. You never touch the weights, you send tokens to an API, and you’re billed per million of them. The second is the open-weight model, so the file sits on Hugging Face, you download it, and your bill becomes the hardware it runs on. The third is the cheap tier that both families now carry, where the same vendor sells a smaller model at a fraction of the flagship’s rate. The fourth is everything that doesn’t produce text, which is priced per image, per second of video, or per song.
That split matters because the four families fail in different ways. A rented model can be deprecated out from under you. A downloaded one can’t be taken away, but then you own the GPUs. A cheap tier can be free until the vendor’s published schedule says otherwise. And a video model can be perfect and still cost more per minute of output than a frontier text model costs per book.
Family one: the frontier models you rent by the token
The three labs we priced here publish rates that sit closer together than the marketing suggests. Anthropic’s models overview lists Claude Fable 5 at $10 per million input tokens and $50 per million output, Claude Opus 5 at $5 and $25, and Claude Sonnet 5 at $2 and $10. OpenAI’s pricing page lists gpt-5.6-sol at $5.00 and $30.00. Google’s Gemini pricing puts 2.5 Pro at $1.25 and $10.00.
| Model | Input, per 1M tokens | Output, per 1M tokens | Context window |
|---|---|---|---|
| Claude Fable 5 | $10 | $50 | 1M tokens |
| Claude Opus 5 | $5 | $25 | 1M tokens |
| Claude Sonnet 5 | $2 | $10 | 1M tokens |
| Claude Haiku 4.5 | $1 | $5 | 200k tokens |
| gpt-5.6-sol | $5.00 | $30.00 | short context rate |
| gpt-5.6-terra | $2.00 | $12.00 | short context rate |
| gpt-5.6-luna | $0.20 | $1.20 | short context rate |
| Gemini 2.5 Pro | $1.25 | $10.00 | prompts under 200k |
Read that table as a starting point rather than a quote, because two of the three vendors charge more once the prompt gets long. OpenAI lists a separate long-context rate for gpt-5.6-sol at $10.00 input and $45.00 output, double and half again the short-context price. Google charges $2.50 input and $15.00 output on Gemini 2.5 Pro once a prompt passes 200,000 tokens. We worked through what that does to a real bill in our piece on frontier model pricing, and the short version is that the headline rate is the least reliable number on the page.
The catch buried deeper is that a token isn’t a fixed unit of text. Anthropic’s own tooltip on Claude Fable 5 says the model uses the tokenizer introduced with Claude Opus 4.7, and that compared with models before Opus 4.7, the same text produces roughly 30% more tokens. So two models at the same headline rate can hand you different bills for identical text, and that tooltip is the only place the difference is disclosed.
A price per token isn’t a price per job. Anthropic’s own note says the same text produces roughly 30% more tokens on Claude Fable 5 than on models before Opus 4.7.
Anthropic, Claude models overview, August 2026
One more thing the rented tier gives you that the download tier doesn’t: a stated knowledge cutoff. Anthropic publishes two of them, and the gap between them is the honest part. Claude Haiku 4.5 is listed with a training data cutoff of July 2025 but a reliable knowledge cutoff of February 2025, which Anthropic defines as the date through which a model’s knowledge is most extensive and reliable. Claude Opus 5 carries May 2026 on both, and Claude Sonnet 5 carries January 2026 on both.
Family two: the open weights you download and run
The download tier stopped being the budget option some time ago. DeepSeek’s V4-Pro-0813 card describes a 1.7T parameter model released under the MIT licence, reporting 87.9 on Terminal Bench 2.1 and 60.0 on HLE with tools, and says it’s “broadly competitive with the strongest proprietary models available”. That’s the vendor’s own claim on its own card, not an independent evaluation, and it should be read that way.
What isn’t a claim is the download counter. DeepSeek-V4-Flash-0731, a 304B parameter model also under MIT, logged 2,547,549 downloads in the month to 20 August 2026 on its Hugging Face page. That’s a measured number, and it says more about which family is winning than any leaderboard does.
| Model | Parameters | Licence named on the card | Native context |
|---|---|---|---|
| DeepSeek-V4-Pro-0813 | 1.7T | MIT | not stated on the card |
| DeepSeek-V4-Flash-0731 | 304B | MIT | described as million-token |
| Qwen3.8-2.4T-A95B | 2.4T total, 95B activated | qwen3.8-max | 262,144 tokens |
| Qwen3.8-27B | 27B | apache-2.0 | 262,144 tokens |
Notice the licence column, because “open weights” hides a spectrum. Both DeepSeek models above carry MIT. Qwen3.8-27B carries apache-2.0. But Qwen’s largest model, the 2.4T, ships under a bespoke licence its card names “qwen3.8-max”, which is a different legal object from Apache even though the weights download the same way.
Mistral is blunter about the split than most. Its pricing FAQ says you can self-host its models anywhere, then adds that open-weight models such as Mistral 7B “are Apache 2.0 licensed for research/individual use; while commercial deployments require a Mistral license with separate terms for derivatives and production use”. Meta’s developer site takes a third position again, describing Muse Glimmer as “a 30-billion-parameter open model optimized for local agent workflows” and naming its Muse Spark models without attaching a licence to them. If you’re weighing the two routes against each other, we’ve costed them out in open weights or closed API.
Family three: the cheap tiers, and what a free AI model costs
Inside every hosted family there’s now a small model priced like a rounding error. OpenAI’s gpt-5.6-luna runs at $0.20 input and $1.20 output, against gpt-5.6-sol at $5.00 and $30.00. That’s 25 times the output price for the flagship, by our calculation from the two published rates, and both models come from the same lab and the same generation. Anthropic’s spread is narrower, because Haiku 4.5 at $1 and $5 sits against Fable 5 at $10 and $50, a factor of 10.
Free is a real option too, and it’s worth being precise about which kind of free. Google’s Gemini pricing page marks the free tier “Free of charge” on its text models, which covers Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash, the Flash-Lite tiers and Gemini 2.5 Pro. The same table carries a row labelled “Used to improve our products”, which reads Yes for the free tier and No for the paid tier. That’s the price, stated plainly by the vendor, and it’s the one figure that never appears in dollars.
The paid rates on that page carry an expiry as well. Gemini 3.7 Flash is listed at $0.75 input and $3.75 output “through December 31, 2026”, then $1.50 and $7.50 “starting January 1, 2027”. So a cost model built on today’s Flash rate has a doubling scheduled into it, in writing, at the turn of the year. Whether the cheap tier is enough for your task is a separate question, and one we tested in flagship or entry tier.
Family four: the models that never produce a word
Text pricing has trained everyone to think in tokens, but a large part of the current catalogue isn’t billed that way at all. Google’s own model list runs to video generation, music, speech translation across 70 or more languages, embeddings, computer use and embodied reasoning for robots. Each one carries a different meter, so the meter is the thing to look at, not the model name.
| Model | What it makes | Paid rate |
|---|---|---|
| Veo 3.1 Standard | Video with audio | $0.40 per second at 720p and 1080p, $0.60 at 4k |
| Veo 3.1 Lite | Video with audio | $0.05 per second at 720p |
| Gemini 3.1 Flash Image | Images | $60.00 per 1M image tokens, given as $0.067 per 1K image and $0.151 per 4K image |
| Lyria 3 Clip Preview | Music, up to 30 seconds | $0.04 per song |
| Lyria 3 Pro Preview | Music, full song | $0.08 per song |
| Gemini Embedding 2 | Vectors, not prose | $0.20 per 1M text tokens, $0.45 per 1M image tokens |
Run the arithmetic on the top row and the shape of this family becomes obvious. Veo 3.1 Standard at $0.40 per second is $24 for a minute of video, which is roughly what a million output tokens costs on Claude Opus 5. That’s a minute of footage against something close to a novel, so the money matches but the unit of work doesn’t. The Lite tier at $0.05 per second changes that ratio by a factor of eight without changing the category.
Embeddings deserve a separate mention because they’re the family people forget they’re already using. Gemini Embedding 2 is described on Google’s model page as a multimodal embedding model mapping text, images, video, audio and PDFs into a unified embedding space. But it doesn’t answer anything. It turns things into vectors so that a search index can compare them, which is why how a multimodal model reads an image matters more than the demo suggests.
Why parameter counts stopped telling you how big a model is
Look back at the open-weight table and one entry refuses to make sense. Qwen3.8-2.4T-A95B has, in its card’s words, “2.4T in total and 95B activated”. The card also lists 92 layers, 512 experts, and 10 routed experts plus 1 shared expert active per token. So the advertised 2.4 trillion is the size on disk, and roughly 4% of it does the work on any given token.
That’s the mixture-of-experts design, and a February 2026 survey by Dong Pan and colleagues puts the mechanism in one line: “MoE model only activate a subset of experts based on a routing network.” The paper, submitted to arXiv on 8 February 2026, argues this sparse conditional computation is what let parameter counts keep climbing at comparable computation cost. Which means a parameter count now describes your storage bill, not your inference bill. We unpacked the architecture separately in mixture-of-experts explained.
The gap between big and small is narrower than the numbers imply, too. Qwen’s own cards report GPQA Diamond at 92.6 for the 2.4T model and 89.2 for Qwen3.8-27B, and SWE-bench Pro at 67.7 against 61.7. By our calculation the larger model carries about 89 times the total parameters for 3.4 points of GPQA Diamond. Those scores are self-reported by the same vendor on the same benchmarks, we didn’t run them, and vendor-reported numbers are exactly the ones to treat with suspicion. Still, they’re the vendor’s own admission of its own scaling return.
What would change this picture
Three things, though none of them are benchmark results. The first is regulation on compute rather than behaviour. Article 51 of the EU AI Act presumes a general-purpose model has high impact capabilities, and therefore systemic risk, when the cumulative training compute exceeds 10^25 floating point operations. That’s a threshold measured in FLOP rather than in downloads, which means it reaches the open-weight family exactly as hard as the hosted one.
The second is deprecation. Anthropic’s overview already files Opus 4.8, 4.7, 4.6, Sonnet 4.6, Sonnet 4.5 and Opus 4.5 under legacy models. Google’s pricing page goes further, warning that the Imagen 4 models are “deprecated and will be shut down on August 17, 2026”, a date that had already passed when we read the page. A downloaded checkpoint under MIT can’t be withdrawn from you, and that asymmetry is the strongest argument the open-weight family has. It isn’t a capability argument at all.
The third is the licence trend. Right now the two DeepSeek checkpoints carry MIT and Qwen’s 27B carries Apache, while the largest model in our table carries a bespoke licence. If the flagship open models keep moving toward custom terms, then “open weights” stops being a category and starts being a marketing word, and this piece’s second family collapses into the first. That’s the thing worth watching, and it’s visible on the model cards long before it shows up on a pricing page.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
