A bigger model does not make a better reel. Here is what it does.
What GPT-6 Astra does inside an AI reel pipeline (brief, script, hooks) and what it cannot touch (pictures, cuts, timing), with the real cost per script.
The hero of our site says "Powered by GPT-6 Astra", and the most common question that follows is some version of "so is that why the reels are good". The honest answer is: partly. A frontier model writes a better hook than a small one. It also has no influence on about half of what decides whether a reel gets watched, because that half is not writing.
So what does the model actually do inside an AI reel pipeline, what does it not do, and when is paying for the biggest model worth it? Here is where GPT-6 Astra sits in Ultim, stage by stage, with what each stage costs.
The jobs a language model has in a reel pipeline
Ultim assigns models per task. The admin page lists five: the main model that writes every reel, the hooks on the free analysis, the brand brief, the planner note, and QA. Each slot can hold any model that returns structured output, and the shortlist for the writing jobs includes GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Sol, Gemini 3.8 Flash and DeepSeek V4 Pro. The lighter jobs get cheaper, faster models.
That design is the first answer to the question. If one model were the product, there would be no slot.
The brief
The model reads what the crawler fetched from the website and writes the brief: the facts, each with its quote, the tone, the topics. This is a reading and summarising job, and the quality that matters is restraint. A brief that adds a claim the site never made poisons every reel after it. The model is told, in the product's own words, that if a claim is not on your site the writer is not allowed to use it.
The script and the hooks
This is where the model earns its price. From the brief it writes a hook of 3 to 8 words that breaks a belief, 6 to 8 beats of 2 to 4 seconds, a scene template for each beat, a one-sentence caption and an end card. The five hooks you see after pasting a URL on the landing page are the same job in miniature.
The difference between models shows up here as judgement: which fact to lead with, whether "676 shops, zero accounts" is a better opening than "the largest database of US manufacturers", whether the sentence is something a person would say out loud.
QA
Mostly not a model job. The product's own description of the QA slot is "reserved for model checks; most QA is deterministic code". Whether every headline appears in the spoken line, whether every number exists in the brief, whether the reel is inside 20 to 30 seconds: code checks these, and a script that fails is sent back to the writer, up to three times.
The jobs no language model does
| Decision | Who makes it | Why not the model |
|---|---|---|
| Which photos and screenshots appear | Writer picks from the brand's assets; code validates | Only real site material is allowed |
| Scene length and cut points | Voice timestamps and the music beat grid | Must match the spoken words to the frame |
| Crops | Content-aware crop in the renderer | Needs pixels, not text |
| Loudness and mix | FFmpeg, -14 LUFS target | Measured, not judged |
| Posting hour | Scheduler, 10:00 to 18:00 local | Set by data from 4.5 billion posts |
| What the next reel tests | Planner, from 24 and 48 hour reads | One variable at a time, from measured results |
None of these are opinions, which is why none of them are prompts. In the motion design style the video is an HTML page rendered by a headless browser, and the model never sees a frame. We described that in how Claude makes motion design.
The ceiling that no model can raise
Based on our analysis of 4.5 billion TikTok posts (2014 to 2026), a model trained on 6 million posts using only metadata (length, caption, hashtags, hour, sound type, hook type) reached an AUC of 0.74. The top 10% of posts by predicted score had about a 3.5% chance of landing in the top 1%, against a baseline of 1%. That is the whole metadata lever: roughly 3x.
A writing model operates on a slice of that lever (the hook type, the caption) plus the part that metadata cannot see: whether the first sentence makes someone stop. It does not operate on the account's history, on the platform's test pool, or on the 38 attempts the 90th percentile creator needed before a first hit. Switching from a small model to a frontier one moves the quality of each attempt. It does not move the number of attempts, which the data says matters more.
What the model costs, and what the reel costs
Via OpenRouter, read on 11 October 2026, GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens. Claude Sonnet 5.5 is $2 and $10; Claude Haiku 5.5 is $0.10 and $0.50. A frontier writer is ten to a hundred times the price of a fast one per token.
In a reel that difference is still small in absolute terms. On 13 October 2026 five production reels cost $0.155 to $0.291 each, of which the script was $0.100 to $0.202 including QA retries, the voice $0.050 to $0.086 and the render $0.003 to $0.004. The script is the largest line, and it is a dime or two. The full breakdown is in what one AI reel really costs.
So the decision is not "can we afford the big model". It is "does the big model's hook earn more than twenty cents of extra attention". The way to know is to measure, which is why the admin page shows results per model: reels written, published, views, likes, shares and cost over the last 90 days. A model keeps the writer slot by its numbers, not by its name.
Where a bigger model can help
There are three places to look when comparing models in the writer slot: whether it follows the facts-with-quotes rule without retries; whether its hooks break a belief rather than describe a topic; and whether it keeps the spoken line and the on-screen headline the same sentence, which the QA reports as "headline not spoken" when it fails. All three show up as QA retries, so a better writer partly pays for itself through a cheaper script. The rest has to be earned in shares per view, which is what the results-per-model page is for.
It does not help with the length lever (61 to 120 seconds, lift 3.08), the sound lever (own voice 1.09 against someone else's music 0.79) or the posting hour. Those are defaults in the pipeline for every model.
What this means for your reels
Judge a tool by where its model sits, not by which model it names. If the model writes from your website's facts and code handles the timing, the crops, the mix and the schedule, the model name is a quality dial on the writing. If the model is being asked to generate the pictures, decide the cuts and guess the posting time, no name on the badge will save the reel.
And budget for attempts before you budget for tokens. An account with 100 or more posts has a 39% chance of at least one top 1% hit; one that stops at four has 1.6%. The model makes each test better. The loop makes there be a hundred of them.
Ultim's hero names GPT-6 Astra for the writer slot, the alternatives are a click away in the admin, and the results per model are in the open. The cheapest way to see its writing is the free analysis: paste your site into /tools/reel-hooks and read the five hooks.
FAQ
What does GPT-6 Astra do in Ultim?
It is the model Ultim names for the main writer slot: the brand brief, the scripts and the five hooks on the free analysis. It does not generate pictures, cut the video or time the scenes; those are code.
Is a more expensive model worth it for reels?
Only for the writing. The script is $0.10 to $0.20 of a $0.16 to $0.29 reel, so the model is the biggest line, but no model changes the length, caption, sound or posting hour levers, which are set by data.
Can I use a different model?
Yes. The writer slot is chosen per task, and the alternatives include Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6 Sol. Results per model are tracked by reels, views, likes, shares and cost.