The AI parts list: stop asking one chatbot to do everything

Suppose a pipe bursts in your kitchen, and your home is insured by Hippo, a US home insurance company. You call to report the damage. The first voice you hear is not a person. It is Clara, an AI voice agent Hippo put on its claims line in April. Clara takes down what happened and looks up your policy. You send photos from your phone, and image-recognition software turns them into an itemized list of what was damaged. Other AI models read the paperwork and draft a summary of the claim. Only then does a human claims adjuster review it and decide what Hippo will pay.

Hippo says Clara now handles 81% of these first calls about a new claim from start to finish, with no person involved, and that the company reaches a customer, on average, 1.2 business hours after a claim comes in. Those are Hippo's own figures. The design is the lesson. Hippo's claims line is not one AI model. It is at least five, each doing the job it was built for, with software and a person deciding what happens next.

Most companies still buy AI the other way. They license one capable chatbot and ask it to do everything: read the scan, transcribe the call, sort the ticket, check the photo, write the reply. That works in a demo. In production it is usually slower, more expensive and harder to control than it needs to be, because a general model is the most expensive part in the box and it returns prose where your software wanted a number.

This article is the parts list: the types of AI model a business can build with in late 2026, what goes into each and what comes out, what each costs at list price, and six concrete ways to put them together.

Why one model is the wrong unit

There are four reasons to build from specialist parts, and the first is price. AI models are mostly priced per token, a chunk of text about three quarters of a word long, and the gap between models that can do a given job is now around a hundredfold. On October 30, OpenAI's GPT-6 Astra listed at $10 per million input tokens and $50 per million output; its GPT-6 Luna, and Anthropic's Claude Haiku 5.5, at $0.10 and $0.50. Decision models charge about four cents per million input tokens and nothing for output. Reading a scanned page with a cloud OCR (optical character recognition) service costs $1.50 per thousand pages. Transcribing an hour of recorded audio costs about 21 cents.

The second is shape. A specialist returns what your code needs: a box around the dent, a probability that the claim is urgent, a vector for search, a transcript with a timestamp on every word, a forecast with a high and a low band. A general model returns sentences that contain those things, and somebody writes code to dig them out.

The third is speed, and the fourth is control. A part that does one job can be tested on that job, measured, priced and replaced without touching the rest. When a vendor changes a model or shuts one down, which happened twice in video generation in the past six weeks, you swap one part rather than rebuild the system.

The evidence is that companies are already moving this way. In a 2025 survey of 100 enterprise CIOs by Andreessen Horowitz, 37% were using five or more models, up from 29% a year earlier, and the most common reason given was that different models are better at different jobs. OpenRouter and Andreessen Horowitz, studying more than 100 trillion tokens of traffic, concluded that no single model dominates and that usage is becoming "model-agnostic and heterogeneous."

The parts list

Here is every family worth knowing, grouped by the job it does. Prices are list prices read on vendor pages on October 30, 2026; several are introductory and will rise.

The rest of this section takes the families one at a time; embeddings and RAG, and the generative models, get sections of their own after it. If you know a family already, skip it.

Models that read and write

Large language models are the general-purpose part: they read instructions and documents, and write text, code or structured data. Vendors sell them in tiers. Frontier models such as GPT-6 Astra, Anthropic's Claude Fable 5.1 and Opus 5.5, and Google's Gemini 4 Argon are for the hardest reasoning and long agent runs. Mid-tier models such as GPT-6.1 Terra, released October 26 at $1 and $5, Claude Sonnet 5.5 and Gemini 3.8 Flash are the everyday workhorses. Small models such as Haiku 5.5 and GPT-6 Luna take the high-volume routine steps.

Two things about language models catch budget owners out. "Thinking" or reasoning modes are billed as ordinary output tokens, including the reasoning you never see, so a reasoning model raises the bill through volume rather than rate; the effort setting is the dial. And open-weight models you can run yourself are not all open source. DeepSeek's latest is MIT licensed and Google's Gemma 4 is Apache 2.0, but Mistral Large 4, whose weights arrived on October 29, needs a separate license to run as a paid model service, and Meta's Muse Spark Open comes with "limited commercial terms."

Vision-language models read pictures: photos, scanned forms, charts, screenshots. You rarely buy one separately any more, because the major language models read images and PDFs. The difference is in the inputs. None of the current Claude or GPT-6 models accepts video; Google's Gemini reads video at one frame per second, up to about three hours. Every page and image is billed as input tokens, so a long scanned PDF can cost far more than its text alone.

Omni, or any-to-any, models take text, images, audio and sometimes video in one model, and answer in text or, in some cases, speech. One omni model can stand in for a chain of speech-to-text, a vision model and a language model, with less delay and fewer moving parts. Google's small Gemma 4 models hear audio as well as reading text and images, under Apache 2.0, and are the most downloaded models in Hugging Face's any-to-any category. Alibaba's Qwen3-Omni, also Apache 2.0, speaks its answers.

Models that perceive

Speech-to-text turns audio into a transcript, either after the fact or live while someone is talking. Batch transcription has become a commodity at roughly 15 to 30 cents per audio hour: AssemblyAI's Universal-3.5 Pro at $0.21, ElevenLabs' Scribe v2 at $0.22, Deepgram's Nova-3 at about $0.26. Live transcription costs more. One change matters to anyone who needs to know who said what. OpenAI is retiring its older transcription models on February 26, 2027, and its new live model returns no speaker labels or word timestamps. Speaker separation now comes from specialists or open models such as NVIDIA's Nemotron 3.5 Transcribe.

OCR and document AI read paper: invoices, receipts, IDs, forms, tables. Plain text costs the same $1.50 per thousand pages at Google, Microsoft and Amazon. The money is in structure. Pulling tables and key fields costs $15 to $50 per thousand pages on Amazon Textract and $30 for custom fields on Google and Microsoft. Mistral's OCR 4.1 returns tables, positions and a confidence score for every word at $4 per thousand pages. The confidence score is the useful part: it tells your workflow which fields a person should look at.

Detection and segmentation models say what is where in an image. Detection draws a labeled box ("forklift, 94%"); segmentation traces the exact outline of the dent or the pallet; tracking follows each object across video frames so you can count. Meta's SAM 3.1 finds every instance of a concept you name in a short phrase. These are the parts behind visual inspection. Procter & Gamble says vision models running beside its lines, built with Siemens, cut scrap by 10 to 20% depending on the product. Read the license before you build: the most popular detector family, Ultralytics YOLO, is AGPL-3.0, and Ultralytics says an Enterprise License is required for internal business tools. Roboflow's RF-DETR core models are Apache 2.0.

Image classification, depth and pose models answer simpler questions about a picture. Classification gives the whole image one label: defect or pass, product category, safe or unsafe. Zero-shot models in the CLIP and SigLIP families, such as Google's SigLIP 2, score a picture against any labels you write, with no training data, and also turn images into vectors for visual search. Depth models estimate how far away each part of a scene is: Depth Anything 3 gives distances in real-world units from a single photo, useful for measuring a room or a stack of pallets. Pose models such as ViTPose+ find a person's joints, for ergonomics, safety and sport. Check licenses size by size: in Depth Anything 3, the small, base and metric models are Apache 2.0, and the large and giant ones are non-commercial.

Models that judge and guard

Small classifiers and extractors are the older, smaller relatives of decision models, and Hugging Face is full of them: more than 127,000 text classification models and 31,000 entity extraction models on October 30. Most are encoder models of a few hundred million parameters that sort text into labels or pull out names, dates, account numbers and other entities. They run on an ordinary server or laptop, cost nothing per call once you host them, and give the same answer every time. The usual route is to fine-tune one on your own labeled examples, starting from a base model such as ModernBERT. Newer ones skip the training: Fastino's GLiNER 2.5 extracts whatever entity types you name in the request. OpenAI's open privacy-filter, released in April under Apache 2.0, finds and masks personal data and runs in a browser. The trade against a decision model is flexibility: a fine-tuned classifier needs labeled data and retraining when your categories change, while a decision model takes new categories in the request.

Guardrail classifiers are small models that sit before and after the main one and answer one question: is this input an attempt to hijack the model, is this output harmful, does this text contain personal data? OpenAI's moderation model is free; Microsoft's content safety service, which includes its Prompt Shields injection check, costs $0.375 per thousand text records; open models such as OpenAI's gpt-oss-safeguard follow a policy you write in plain English. They are one layer of defense, not a guarantee.

Decision models are the newest family, and the one most workflows are missing. A decision model does not write. You give it the situation and a set of typed questions, and it returns typed answers: which option, with a probability for each; a score on a scale you define; the probability that a statement is true. We covered the category and its economics in The four-cent decision. In the six weeks since TypeSafe's Jev started it, the category has filled up. OpenAI's Decisions API is generally available, runs on GPT-6 Luna at $0.10 per million input tokens with output free, and accepts images. Microsoft sells Microsoft-Decision-1 at Jev's price of $0.042. Cloudflare's Clef-omni takes audio and video as well as text and images. And AWS's Strands Labs released Strands Decider 2B, free under Apache 2.0 and small enough to run on an ordinary processor.

Jev is learning to see

The most interesting change in this family is at TypeSafe. When Jev launched in September it read text only. Its documentation now describes image input on both Jev 1.13 and Jev 1.14, the version released on October 14, and 1.14 adds a question type built only for pictures. That moves a decision model into territory that belonged to vision-language models and to the detection models above, while keeping what makes it a decision model. It still does not write, and every answer still comes back typed, with probabilities your code can act on.

An image can go anywhere text can go in a Jev request, and each place does a different job.

• The picture being judged. Put a photo in the situation and ask typed questions about it: is there a person at the door, what are they carrying, is the seal on this package broken. Every question is answered about the same image in one call.
• A picture inside the question. Ask whether the dog in the found-dog photo is the same dog as the one on the poster, or whether this signature matches the one on file.
• Pictures as the answers. This is the newest idea. A choice can be defined by example images instead of words, and a yes-or-no question can define "yes" with a photo. TypeSafe's own example defines yes as "correctly assembled," shown with a picture, and no as "parts missing or in the wrong place." That is visual inspection by example: no training run and no labeled dataset, just a reference photo of what good looks like. Jev is never fine-tuned on customer data, so examples in the request are how you teach it.

The new Box question asks where something is and returns a rectangle around it: the mug on the counter, the signature on the form, the person at the door. Finding things in a picture was the detection model's job, and the vision-language model's.

TypeSafe's worked examples show where this goes.

Video, one frame at a time. In its feeder camera example, Jev is sent one frame a second from bird feeder footage with five questions per frame, one of them a Box that finds the bird, and plain code turns the answers into a log of visits by species. On the 101-second clip TypeSafe used, Jev agreed with hand labels on whether a bird was present in all 101 frames, and named the right species in 132 of 133 sightings. That is the company's own test on one clip, not a benchmark, but it shows the shape: a camera feed becomes rows in a table. Swap the birds for vehicles at a loading bay or customers at a counter and the code barely changes.

Paper with no text layer. In a scanned vaccination record, Jev finds each vaccine with a yes-or-no question and a Box, then reads the year of each dose with a choice that includes "unreadable." Where a sticker covers a date, it answers unreadable instead of guessing.

Checking another model's work. In an invoice example, Jev draws a box around each value an OCR model read from a photo and checks the value against the page and against a close crop of the box. On TypeSafe's two sample invoices, the OCR model got three values wrong, and Jev flagged all three for a person to review.

The difference from a vision-language model is the shape of the answer. A vision-language model tells you what it sees, in sentences. Jev answers the questions you set, from the options you set, with a probability for each, and says where. For the many business questions about pictures that have a fixed set of answers (is it damaged, which product is this, is the label there, where is the serial number), that is the more useful shape, and it is priced like a decision. Images are billed as input tokens at $0.042 per million, with nothing charged for output. TypeSafe has not yet published how many tokens an image uses, so a cost per image cannot be worked out yet.

Three cautions. TypeSafe calls its image features "dreamer-ready": meant for prototypes, not yet up to its standard for production, switched on with a request header, and liable to change between releases. They take still PNG and JPEG images between 280 and 1,568 pixels a side, so video goes in as frames and audio not at all. And Jev is not alone. OpenAI's Decisions API reads images, Cloudflare's Clef-omni reads audio and video, and AWS's Strands Decider has an optional vision package. Decision models and vision models are converging, and for a buyer that means a cheaper answer to "which model looks at the photo?" whenever the question has a fixed set of answers.

Models that predict

Forecasting and tabular models do what spreadsheets and statisticians have done for decades, without training on your data first. A time-series foundation model, trained on millions of series, forecasts a new one with a range around it: demand, call volume, cash. Amazon's Chronos-2 is Apache 2.0, and Google's TimesFM runs inside BigQuery as a single SQL function. Two warnings. The licenses tightened this year: TimesFM 3.0's downloadable weights are non-commercial, and Prior Labs' tabular model TabPFN-3.5 forbids commercial use even of its outputs without a paid license. And classic machine learning still does most structured prediction. If you already have a gradient-boosted model forecasting churn, a foundation model is a challenger to test, not an upgrade to assume.

Models that talk

Voice agents hold a live spoken conversation and do things while they talk: look up the order, book the slot, take the claim. There are now three ways to build one. A chained agent runs speech-to-text, then a language model, then text-to-speech; you can read and check the text at every step and swap any part. A speech-to-speech model hears audio and speaks audio directly, which sounds more natural and handles interruptions better, with less visibility into what it is doing. And a newer design, OpenAI's GPT-Live, holds the conversation itself and hands the reasoning and tool use to a separate agent you already have. A typical production build costs 5 to 15 cents a minute: GPT-Live at $0.05 plus the backend, ElevenLabs Agents at $0.08 plus the language model, Retell's default build at $0.11. Premium speech-to-speech models run higher, and the phone carrier charges on top.

Under every voice agent sit three small models that rarely get named. Voice activity detection hears when someone is speaking; turn detection decides when they have finished, so the agent does not talk over them; diarization tells speakers apart. They are tiny and mostly open: Silero VAD is MIT-licensed, and Pipecat's Smart Turn has 8 million parameters. Read the license anyway. LiveKit's widely used turn detector may be used only inside LiveKit's own agent framework.

Models that act

Computer-use models look at a screen and operate it. They read a screenshot, decide where to click or what to type, act, and look again, until the task is done. They are how an agent works in software that has no programming interface: an old claims system, a supplier portal, a government form. The major vendors offer computer use as a tool their language models drive, as in OpenAI's computer-use guide, and open models exist, such as Microsoft's Fara 1.5 under MIT and H Company's Holo4 under Apache 2.0. They are slower and less reliable than an interface built for software, and they act with whatever access their account has, which is why they need the identity and limits we described in our article on AI agents inside your systems.

Robotics and world models carry the same idea into the physical world. A vision-language-action model takes camera images and an instruction and outputs motor commands: NVIDIA's GR00T N1.7 for humanoid robots, Hugging Face's SmolVLA, small enough to train on a single graphics card, and Google DeepMind's Gemini Robotics 2. Physical Intelligence opened its robot models to customers through an API beta on October 13. World models generate simulated environments where robots and self-driving systems can be trained and tested before they touch the real thing; NVIDIA's Cosmos 3 is open for commercial use, and DeepMind's Genie 4 has a waitlist. For most businesses this family is still one to watch. Warehouses, factories and logistics are the likely first buyers.

Embeddings and RAG: how AI finds the right page

Most questions a business puts to AI are about its own material: this policy, that contract, last year's tickets, the product manual. A language model has read none of it. Retrieval-augmented generation, or RAG, is the standard fix. Find the few passages that answer the question, and hand them to the model along with the question. Embedding models are the part that does the finding.

What an embedding is

An embedding model reads a piece of content and returns a vector: a list of a few hundred to a few thousand numbers, placed so that content with similar meaning lands close together. "How do I get my money back?" and "refund policy" share no words, but their vectors are neighbors. That is what lets search match meaning rather than keywords, and it works across languages. The newest models embed images, audio, video and PDF pages into the same space as text. Google's Gemini Embedding 2 does all five, so a photo of a broken part can find the page of the manual that covers it.

Embeddings are among the cheapest calls in AI. OpenAI's text-embedding-3-small costs $0.02 per million tokens, and it is still OpenAI's current model, text only, two years after release. Voyage 4 runs $0.02 to $0.12 and Cohere's Embed 5, which reads images too, $0.08 to $0.12. At those prices, embedding a library of 10 million words costs less than two dollars. One warning: switching embedding models means re-embedding everything, because vectors from different models do not compare. Voyage's 4 series is an exception within its own family.

How a RAG pipeline works

A RAG system is eight steps, and only three or four of them are models.

1. Ingest. Connectors pull documents from SharePoint, Google Drive, the wiki, the ticket system. Software.
2. Parse. Files become clean text and tables. For scans, this is an OCR or vision-language model.
3. Chunk. The text is cut into passages. OpenAI's file search defaults to 800 tokens, overlapping by 400. Software, following rules you set.
4. Embed. Each chunk becomes a vector. A model.
5. Store. Vectors go into an index, next to the text, its source and who is allowed to see it. A database.
6. Retrieve. The question is embedded too, and the nearest chunks come back. A keyword search runs alongside and the two lists are merged, because exact terms such as part numbers and names are where vectors are weakest. Software.
7. Rerank. A reranker reads the question with each of the top 50 or 100 candidates and re-scores them, so only the best few go forward. A model. Cohere's Rerank 5, released October 21, costs $0.02 to $0.05 per million tokens.
8. Generate. A language model answers from those passages, and cites them. A model.

Each layer earns its place. In a retrieval study Anthropic published, plain embedding search missed the right passage 5.7% of the time. Adding a line of context to each chunk before embedding it cut that to 3.7%; adding keyword search, to 2.9%; adding a reranker, to 1.9%. Two thirds of the misses were gone, and none of the fixes involved a bigger language model.

Where the vectors live

A vector database stores vectors and finds the nearest ones fast. There are specialists, such as Pinecone, Weaviate, Qdrant and Milvus, and most databases you already run have added vector search: pgvector for Postgres, MongoDB Atlas, Elasticsearch and OpenSearch. Cloud storage has joined in. Amazon's S3 Vectors charges $0.06 per gigabyte a month, and AWS's own worked example puts 10 million vectors and a million queries a month at $11.38.

Size is the cost driver. A vector of 1,536 numbers takes about 6 kilobytes at full precision, so a million chunks is about 6 gigabytes. Storing each number in 8 bits instead of 32 cuts that by four and keeps about 99% of retrieval quality; storing it in a single bit cuts it by 32 and keeps about 96%, with a rescoring step, according to Hugging Face's measurements. Hugging Face puts 100 million full-size vectors held in memory at about $2,174 a month; in single bits, the same vectors need a thirty-second of the space.

If you would rather not assemble the pieces, the major platforms sell RAG as a service. OpenAI's file search charges $0.10 per gigabyte per day after the first gigabyte, plus $2.50 per thousand searches. Google's Gemini File Search charges only for embedding files when they are first indexed; storage and the embedding of questions are free. AWS's Bedrock knowledge bases include a managed option at $5 per gigabyte of source data a month and $1 per thousand retrievals.

Why not just send everything?

Language models now read a million tokens, about 2,500 pages, in one request. So why retrieve at all?

Cost, first. Sending a million-token document set with every question costs about $2 to $4 per question at mid-tier prices, before any caching discount. Retrieving 8,000 tokens of relevant passages costs about two cents. That is our arithmetic at October list prices, and at 10,000 questions a month the gap is tens of thousands of dollars.

Accuracy, second. Models get worse at finding things as the pile grows. Chroma tested 18 models and found that every one did worse when the relevant material sat inside a prompt of about 113,000 tokens than when it was handed over in a focused prompt of about 300. In the NoLiMa benchmark, 11 of 13 models fell below half their short-prompt score at just 32,000 tokens, when the question and the answer shared few words.

Long context does sometimes win on quality. A June 2026 study found answers from the full document set more accurate than RAG, 73.1% to 65.4%, at 26 times the token cost. The working rule from the vendors' own guidance: a small, stable collection, under about 200,000 tokens or 500 pages, can go straight into the prompt with caching. A large, changing or permission-controlled collection with many questions a day needs retrieval.

Newer patterns

Agentic retrieval lets a model plan several searches, read the results and search again, the way a researcher would, instead of searching once. GraphRAG builds a map of the people, products and relationships in a collection, so it can answer questions about the whole of it ("what are the recurring complaints?") that passage search handles badly; Microsoft's LazyGraphRAG cut the indexing cost to 0.1% of the original method. Page-image retrieval, pioneered by ColPali, embeds a picture of each page instead of its parsed text, so charts, tables and layout survive. And permission-aware retrieval checks, at search time, whether the person asking may see each passage.

Where RAG goes wrong

RAG fails in plain ways: the answer is not in the collection, the right passage is not retrieved, the passage is retrieved but dropped, or the model ignores it. A study of three production systems counted seven such failure points. Retrieval does not end made-up citations, either. A Stanford study found that leading legal research tools built on RAG and marketed as hallucination-free still produced hallucinations 17% to 33% of the time. And retrieval inherits your permissions: an assistant that searches everything a user can open will surface every overshared file. Measure retrieval and answers separately, whether the right passage came back and whether the answer is faithful to it, with open tools such as RAGAS.

Done well, the return is large. OCBC, the Singapore bank, describes a system for its corporate credit analysts that combines keyword and vector search with a reranker; in a paper written with Georgia Tech researchers, it reports that across more than 800 analysts, document review fell from several hours to about three minutes.

Embeddings beyond search

The same vectors do other work. Cluster a year of support tickets or survey comments and the recurring problems appear without anyone reading them all. Find near-duplicate products, invoices or customer records. Recommend the article or product closest to what someone just read. Flag the transaction or the sensor reading that sits far from all the others. Each is an embedding call and a little arithmetic, at a fraction of a cent per item.

Generative AI: the models that make things

Every language model above is generative: it writes. This section is about the models that make everything else: pictures, video, voices, music and on-screen presenters. They move faster than any other family on the list, they carry the most legal strings, and their real cost is rarely the price per item. It is the number of tries and the human review each finished piece needs.

Images: create, edit, keep it on brand

An image model does three different jobs, and vendors are not equally good at each.

Creating an image from a prompt is the job everyone has seen. Editing is the one businesses use most: replace the background, remove an object, put the product in a kitchen instead of a studio, resize for another placement, translate the words on a poster. The newer models edit by instruction ("make the jacket navy and change nothing else"), by a mask you paint over the area to change, or, in Black Forest Labs' FLUX 3, by a box you specify. Keeping things consistent, the same product, character or style across hundreds of images, is the hardest job. OpenAI's own documentation still lists consistency for recurring characters and brand elements as a limitation. The tools for it are reference images (FLUX 3 accepts up to ten) and style models trained on your own approved images; FLUX 3's small Klein version can be fine-tuned and run on your own servers.

Sources: OpenAI, Runway, Google, Google Cloud, Black Forest Labs, Recraft. Midjourney, still the favorite of many designers, has no public API, which makes it a tool for people rather than a part for a workflow.

Video: start from an approved still

Video models make clips of a few seconds up to about 20 seconds per generation, now usually with speech, sound effects and music generated in the same pass. They start from a text prompt, from an image (sometimes a first and last frame), or from an existing clip to restyle or extend. For business work the dependable pattern is to start from a still that someone has already approved, so the product is right in the first frame, and let the model add the motion.

Sources: Google Cloud, Runway, Black Forest Labs, Luma.

An eight-second clip therefore costs from about $0.40 to $3.20 per attempt, and a usable clip usually takes several attempts. This is also the most volatile family on the list. OpenAI removed Sora 2 and its Videos API on September 24 with no direct replacement, and on October 22 Google took its Veo preview out of the Gemini API, pointing developers to a new model, Gemini Omni Flash, or to Veo on Google Cloud. One hedge is an aggregator. Runway's developer platform sells access to about 15 other companies' video models alongside its own, and offers model routers that pick one for you, billed at the rate of whichever model they choose.

Voice: narration, cloning and dubbing

Voice models do three jobs.

Narration (text-to-speech) reads a script aloud in a chosen voice, in dozens of languages, with control over pace and emotion. ElevenLabs' Eleven v4 costs $0.08 per thousand characters, roughly six cents per spoken minute, and its Turbo version, built for live use with about a tenth of a second of delay, costs half that. Google's Gemini 3.8 Flash text-to-speech costs about 1.4 cents a minute until December 31 and double that from January. OpenAI is retiring its standalone text-to-speech models on January 6 and moving speech to its live-conversation API, which is a change of workflow, not just of model name.

Voices can be designed from a written description, picked from a library, or cloned from a recording of a real person. Cloning is now gated by consent. OpenAI and Google both require the speaker to record a fixed sentence stating that they own the voice and consent to the clone. ElevenLabs' highest-quality cloning goes further: you can clone only your own voice, even with someone else's permission, and they must create and share it themselves.

Dubbing chains the voice models together: transcribe the original, translate it, and speak the translation, often in a clone of the original speaker. Live speech translation is now a single model too; OpenAI's costs $0.034 a minute. The economics can be striking. The small marketing team at Lounge by Zalando now generates ad voiceovers with AI across 25 markets, which its manager says would cost over a million a year the traditional way.

Music and sound

Music models compose instrumental or vocal tracks from a prompt, for ad beds, hold music, sonic logos and video soundtracks; sound-effect models make the door slam and the crowd noise. Quality is no longer the hard question. The hard question is whether the output is cleared for commercial use.

ElevenLabs says its music model was built with labels and publishers and is cleared for nearly all commercial uses, at $0.15 a minute; its sound effects cost $0.12 a minute. Google's Lyria 3.5 costs $0.08 a song. The consumer favorites are still in court. Suno settled with Warner but is still being sued by Universal and Sony, and Udio, which settled with Universal and Warner on terms that require a new model trained on cleared music, was sued again by Sony in July. For anything you publish, use a vendor whose license covers commercial use in writing.

Avatars and digital presenters

Avatar models put a realistic presenter on a script: a stock character or a "digital twin" of a real employee, lip-synced, and translated into many languages. Synthesia and HeyGen are the best-known vendors, and the main business use is training, onboarding and internal communication, where a script changes more often than anyone can reshoot it. The newest versions are live: a face on top of a voice agent, which Runway prices at about 20 cents a minute. A digital twin of a real employee is a likeness right. Get written consent, and decide in advance what happens to the twin when that person leaves.

3D

3D generation turns a photo or a prompt into a textured 3D model, for product viewers, augmented-reality try-ons, games and digital twins. Microsoft's TRELLIS.2, under MIT, is the most downloaded image-to-3D model on Hugging Face, and Meta's SAM 3D rebuilds objects from a single photo. Licenses vary sharply here too: Tencent's Hunyuan3D 2.1 license does not apply in the European Union, the UK or South Korea.

Using generated media safely

Four checks belong in every generative workflow.

Disclose. In the European Union, Article 50 of the AI Act has applied since August 2. A business that publishes realistic generated images, audio or video of people, places or events, a "deep fake" in the Act's words, must disclose that it is artificial, with a lighter touch for work that is evidently creative or satirical. Fines reach €15 million or 3% of worldwide turnover.

Keep the provenance. The major generators now mark their output: Google applies its invisible SynthID watermark to generated video, and the industry standard C2PA "content credentials" record in a file's metadata that AI made it. Editing tools and upload pipelines often strip that metadata. Make sure yours do not.

Own what you need to own. The US Copyright Office says that prompts alone generally do not make an output copyrightable. A fully generated jingle or ad image may be one a competitor can copy. Where you need to own the asset, keep real human creative work in it, and keep a record of that work.

Know who stands behind it. Google Cloud offers customers an indemnity covering generated output, and some other vendors make similar promises in enterprise contracts. Read the exclusions: indemnities typically do not cover outputs you have modified, or prompts that deliberately imitate someone else's work.

A field guide to Hugging Face

Hugging Face is the largest public library of AI models: 3,199,469 of them on October 30, filed under 52 task labels. If you browse it, almost every label maps to a family above.

A few labels are older jobs that general language models have largely absorbed: summarization, extractive question answering, table question answering, fill-mask. You will still find thousands of models under each, and some are worth it for cost or privacy, but for most business uses a small language model now does the job. The reinforcement learning label holds nearly 80,000 models, many of them coursework.

A model's page carries a lot that is not its task. What the labels mean:

• Base or instruct. A base model has only been trained to continue text. An instruct, chat or "-it" version has been tuned to follow instructions. You almost always want the instruct version.
• Size. "7B" is seven billion parameters. "35B-A3B" means 35 billion in total with 3 billion active for each word, a "mixture of experts" design: the total sets how much memory you need, the active count sets the speed and cost. Google's "E4B" means four billion effective parameters.
• Quantized copies. Files marked GGUF, MLX, AWQ or 4-bit are compressed versions that run on smaller hardware with some loss of quality. Hugging Face holds more than 216,000 GGUF repositories, mostly uploaded by the community rather than the original maker. The original license still applies to a copy.
• The model tree. Each page shows what was built from the model: fine-tunes on new data, LoRA adapters (small add-ons that specialize a model without changing it), merges of several models, and quantized copies. Google's Gemma 4 E4B page lists more than 440 adapters, 440 fine-tunes and 425 quantized versions.
• The license. Apache 2.0 and MIT are permissive. A license marked "other" means the real terms are in a file in the repository, and we found every kind of string this month: non-commercial use only, whole regions excluded, revenue caps, use only inside the maker's own software, and different terms for different sizes of the same model. Read the license file, not the tag.
• Gated. You must accept the maker's terms, or be approved, before you can download.
• Custom code. A model that needs "trust remote code" runs Python from its own repository when you load it. That is a security review, not a checkbox.
• Downloads and likes. These measure popularity, not quality, and "trending" favors whatever is new.
• Spaces and inference providers. Spaces are demos other people built with a model. Some models can be called through Hugging Face's inference providers without hosting them yourself.

Six ways to put the parts together

The value is in the combinations. Here are six that real companies run, with the parts in order and where the person sits.

1. The claims desk

For insurance, warranty and any business that pays out against evidence.

1. A voice agent or web form takes the first report.
2. Detection and segmentation find the damaged items in the customer's photos, and a vision-language model turns them into an itemized list.
3. Document AI reads the receipts, estimates and reports, with a confidence score per field.
4. A decision model answers narrow questions with a probability each: is the damage consistent with the reported cause, is this a likely total loss, does anything need a specialist. Code routes the confident answers to a fast track and the rest to an adjuster.
5. A language model drafts the claim summary and the customer letter.
6. An adjuster decides.

This is close to Hippo's design. Our addition is step 4: a decision model in the middle is what lets code, rather than a person, sort the easy claims from the hard ones, at a few cents per million tokens.

2. The support desk that answers from your own documents

1. Speech-to-text if the customer called.
2. A guardrail classifier screens the message, and a decision model reads intent and urgency.
3. An embedding model finds candidate passages in your manuals and policies, and a reranker keeps the best few.
4. A small or mid-tier language model drafts the answer from those passages only.
5. A decision model checks that each claim in the draft is supported by a passage. Below your threshold, a person sees it first.
6. Text-to-speech, or a voice agent, speaks the answer, and the text is logged.

Verizon runs a platform built on Google Cloud that it says handles the majority of its inbound consumer calls and chats. Its chief technology officer describes the design as modular on purpose: "we can plug a different language model without changing everything else."

3. Paperwork in, decisions out

For accounts payable, underwriting, lending and customer onboarding.

1. Document AI extracts the fields, with a confidence score for each.
2. Code checks what code can check: totals add up, the purchase order exists, the dates are in range.
3. A decision model that reads images checks each extracted value against the part of the page it came from, and flags the ones that look wrong.
4. A rules engine or decision model settles the clear cases. The complex ones go to a language model copilot that assembles the file, summarizes it and suggests a decision with the evidence.
5. A specialist decides, and the copilot drafts the explanation to the customer.

Life insurer PolicyMe runs this shape in underwriting. An automated engine decides most applications in seconds; complex ones go to an in-house copilot that turns long medical records into one structured summary and suggests a decision. The company says turnaround on complex cases fell 59%.

4. The content studio

For catalogs, campaigns and localization.

1. A language model writes the brief and the shot list from the product data.
2. Image generation, given real product photos as references, makes the stills and variants.
3. Video generation turns the best stills into short clips.
4. Text-to-speech reads the voiceover in each market's language, over a licensed music bed.
5. A checker compares the generated images with the real product photos for color, shape and detail, and a decision model that reads images checks the brand rules.
6. A human creative approves, the content is labeled where the law requires, and the provenance metadata is kept.

At Zalando's partner day in September, executives said around 70% of the content on display was made with AI, WWD reported, and Zalando is building a "product truth" engine that compares generated content with the original product imagery and flags inaccuracies. Its subsidiary About You said generating product presentation instead of photographing it cut that spend by about 90%, saving around €6 million of a €7 million budget line. At Mondelēz, a six to eight second social video that took up to 10 weeks can now be generated in under five minutes, though the human checks still take days.

5. Search what the camera sees

For retail, field service and spare parts.

1. A small detection model on the phone finds the object in view.
2. An embedding model turns what the camera sees into a vector.
3. Vector search matches it against the catalog.
4. A language model answers questions about the match.
5. If nothing matches, a vision-language model describes the scene in words and searches on that.

That is Amazon's Lens Live, available to tens of millions of US customers, which matches the camera view against billions of products and hands follow-up questions to its shopping assistant. The same five parts let a field technician photograph a worn part and get the part number and the manual.

6. The localization line

For training, marketing and product video that has to work in many languages.

1. Speech-to-text, with speaker labels, transcribes the original.
2. A language model or translation model translates it, held to a glossary of your product and brand terms.
3. Text-to-speech speaks each translation, in a licensed voice or a consented clone of the original speaker.
4. Where a face is on screen, an avatar or lip-sync model matches the mouth to the new language.
5. A linguist in each market reviews tone, terms and timing, and the dub is labeled as AI-voiced where required.

3Play Media, which sells this kind of service, says its AI-voiced dubs have had more than 30 million hours watched on YouTube, and that over two months in eight languages they held viewers 24% longer on average than the original English tracks. That is a vendor's own data, and the company credits the human linguists in its process for the result.

How to wire them without regret

Put code in charge. Each model is a function call with an input and a typed output; your code decides what happens next. Keep arithmetic, dates and business rules in code. Thomson Reuters' new CoCounsel for tax and accounting writes and runs code for its calculations rather than letting a language model do the sums.

Use the cheapest part that can do the job, and escalate on low confidence. Snowflake built this into its AI query engine: a small model handles most rows and passes the uncertain ones to a large one. In its published results, the small model alone lost 18.8% of the large model's accuracy; with escalation, the loss was 4.3%, and the job ran 2.9 times faster. The trade is real, though. In a benchmark LangChain ran with NVIDIA, routing an agent's calls between a small and a frontier model cut cost by 74% and accuracy by about six points, and made the bill per run less predictable. Routing pays when the price gap between the parts is wide and the cheap part is right most of the time.

Keep every part behind your own interface. Vendors retire models on their schedule, not yours. Since late September, OpenAI has removed its video API and Google has moved its video preview, and OpenAI has set end dates in January and February for its standalone text-to-speech models and its older transcription models. A thin layer of your own between the workflow and each vendor turns a forced migration into a configuration change.

Test each part, then measure the whole. Every part needs its own test set: the 300 to 500 real cases with known answers that tell you whether the cheaper part holds up. But the business result belongs to the chain. The Permanente Medical Group's AI scribes, which transcribe a visit and draft the clinical note for the doctor to approve, saved an estimated 15,791 hours across 2.6 million visits. At Beth Israel Lahey Health, a study of 1.6 million visits found notes took 1.2 minutes less each, with no gain in after-hours work or productivity. The same parts can produce very different returns, depending on how the work around them changes.

Put the person where the decision is. In every example above, a human approves at the point where money, health or the brand is at stake, and the models do the reading, sorting and drafting before that point. That is also where your controls and your audit log belong.

The fine print

Most outcome numbers are the companies' own. Hippo's, PolicyMe's, Zalando's and Mondelēz's figures come from press releases or executives on stage. The health system evaluations and the published benchmarks are stronger evidence. Treat the rest as directional, and measure your own.

Open weights come with licenses. The strings are in the detector, forecasting and tabular families this year: AGPL on YOLO, non-commercial terms on TimesFM 3.0 and Salesforce's Moirai 2.0, a ban on commercial use of TabPFN-3.5's outputs, and a $20 million monthly revenue cap on Mistral's new Timestral. Your legal team should read the license of every open model before it reaches production.

Talking to an AI must be disclosed too. Article 50 of the EU AI Act also requires that people be told when they are talking to an AI, including on a phone line. In the United States, the FCC ruled in 2024 that AI-generated voices count as artificial voices under the robocall rules, so outbound AI calls need the recipient's prior consent. The rules for generated images, video and music are covered in Using generated media safely, above.

Where the list is heading

Three things are worth watching.

The families are merging. A year ago, "read this photo" meant a vision model, "sort this ticket" a classifier and "transcribe this call" a speech model. Now decision models look at photos, embedding models put text, images, audio and video in one space, and omni models hear, see and speak in a single network. The lines on this list will keep blurring, and the useful question for each step of a workflow will be less "what kind of model is this?" than "what shape of answer does this step need?"

Prices fall and the shelf turns over. The cheapest language models fell to ten cents per million input tokens this year, and decision models to about four cents. In the same weeks, OpenAI left video and set end dates for its older speech models, and Google moved its video preview. Any part on this list is likely to get cheaper, and any part can disappear.

The hard part moves from choosing to composing. Any competent team can pick a good model. What separates the claims desk that answers in an hour from the one that answers in a week is the chain: which part goes where, what code checks between the parts, where low confidence sends a case to a person, and how the whole is measured. That is what this list is for.