AI Models That Make Images, Video and Speech: 180 Counted

Most talk about AI models is about chat. But in the snapshot of OpenRouter’s public model catalog we took on September 25, 2026, 180 of the 625 routes produce something other than text: images, video, speech, transcripts and the number lists that power search. They come from 35 vendors, and they are being added faster than chat models.

A route here means one model offered through the catalog’s single API, so a model with a batch or free variant counts more than once. The count below is what a developer or marketer can reach through one gateway, not a census of every model in the world. It also shows a practical gap: for 39 of these routes, including every video model, the catalog prints no usable price.

The pace is the clearest signal. Non-text listings ran at 29 in July and 27 in August, and 18 more arrived in the first 25 days of September. For comparison with language models, our census of context and output limits covers the text side of the market.

Some output types need a word of explanation. Embeddings turn text into lists of numbers so software can find similar content; rerank models sort search results by relevance. Transcription turns speech into text and speech turns text into a voice. Audio covers two Google music models and two OpenAI voice-chat models. Decisions is a newer type that returns a structured choice instead of prose; our post on TypeSafe’s Jev explains it.

Fifteen routes produce text as well as another output, mostly Google and OpenAI image models that can reply in words and pictures. Each counts once under its non-text type. Among speech routes, 16 of 20 list their voices: Deepgram’s Aura-2 lists 90, Kokoro 54 and MiniMax’s Speech 2.8 models 45 each, while Microsoft’s MAI Voice 2 lists four. Our ranking of text-to-speech models compares their quality and vendor prices.

Most of the 29 video routes turn a prompt or a still image into a clip. Nine also accept video, for editing, extending or upscaling footage, and six also accept audio as an input.

One row shows why a catalog is a starting point, not an authority. OpenAI’s deprecations page scheduled the removal of its Videos API and the Sora 2 models, including sora-2-pro, for September 24. Our snapshot a day later still listed openai/sora-2-pro. The catalog also shows an expiry date of November 11, 2026 on bytedance/seedance-1-5-pro, the only non-text route carrying one.

For chat models, a catalog price is a cost per million tokens and can be compared directly. For other outputs it cannot. Every video route lists zero in every price field, so the catalog gives no way to budget a clip. Image routes mostly carry their cost in image fields rather than the usual input and output fields: 37 of 57 list zero for both of those.

Transcription shows the unit problem most sharply. Its raw input prices run from 0.00000125 to 0.36 across 24 routes, a spread that reflects different billing units, such as tokens, seconds or minutes, rather than a real price gap. The catalog does not say which unit each route uses.

A catalog is a starting point, and the sections above exist because each of its columns can be read as something it is not. The four illustrative readings below each take one column at face value. They are drawn from the snapshot’s own rows, not from teams we have advised.

Each reading would have produced a confident number and a wrong plan. The record in the next section writes the unit of each column down before the number is used.

The conclusion’s matrix says where the catalog stops and the vendor’s page begins, and a record per candidate holds the hand-off. The route: the catalog ID, its listed input and output types, its listing date and any expiry date, with the snapshot time they were read at. The vendor’s page: the price, the billing unit and the availability status, read on a dated visit, since the catalog cannot supply the first two for video and states the third only as a listing. The unit conversion: for transcription, every price turned into cost per audio hour; for speech, the per-minute cost beside the voice list and languages; for embeddings, the per-token price that already compares with chat models. The test: one small paid run, with what the invoice actually counted beside what the page said it would. And the decision, with the catalog’s role recorded as where the candidate was found rather than as the source of its price.

An illustrative record for a text-to-speech choice reads: route, a speech model with its voice list and no expiry date; vendor’s page, a per-minute price and a current status, read the same day; conversion, the per-minute price multiplied out for a month of expected audio; test, ten minutes of generated speech, with the invoice confirming the unit; decision, shortlisted against two other speech routes on quality, with the price now a known quantity.

The record is also what keeps the snapshot honest over time. Routes are added and removed after the moment the catalog was read, and a record that carries the snapshot time and the vendor visit date can be re-checked, where a bare count cannot.

Two of the catalog’s own dates belong in the record for the same reason. The listing date is the catalog’s, not the vendor’s release date, so a route that was listed in September may describe a model that shipped months earlier. The expiry date, which only one non-text route carried in our snapshot, is a scheduled removal that the record should carry forward to the plan, because a pipeline built on a route with an expiry has a known date by which it needs a replacement.