GPT-6 Astra for Ecommerce Content: Pricing, Specs, and What It Actually Wrote on Real Store Pages

OpenAI released GPT-6 Astra on September 3, 2026. It is the company's new flagship, priced at the very top of the market, and it arrived with a long list of benchmark records. What it did not arrive with is any evidence about the job most Shopify merchants would use it for: writing product descriptions, collection pages and the rest of a store's content.

This guide is the honest version of the launch coverage. It lays out what OpenAI has actually put on record, from the announcement, the model page, the pricing page and the migration guidance, read through the lens of store content. It explains what a better model would have to prove on that job. And it reports the head-to-head test we ran on our own production prompts: two real product descriptions and one real collection page, on GPT-6 Astra, GPT-5.6 Sol and GPT-5.4, all on the same day, with every token count, every cost and the actual sentences each model wrote.

We published the first version of this post on September 4 with the Astra column empty, because the model was not yet callable from a standard API organization. On September 9 it was, and this is the filled-in version. The update log at the end records what changed.

What OpenAI shipped

The announcement calls GPT-6 Astra "our most intelligent and aligned model" and says it "sets a new state of the art for computer use, browsing, software engineering, cybersecurity, science, and professional work." The headline numbers are the kind that get quoted: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench, 57.9% on Terminal-Bench 4.0 against 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1, 96.0% on GPQA Diamond, and 59.3% on Agents' Last Exam against 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol.

Read that list again and notice what it is. Mathematics, abstract reasoning, exploit development, terminal work, graduate science, agentic software tasks. Every benchmark on the launch page measures a model acting or reasoning. Not one measures a model writing something a customer will read. That is not a criticism of OpenAI; those are the frontier problems. It is a warning about how to read the launch if your problem is copy.

Availability

The model page was precise at launch: "GPT-6 Astra is rolling out today for enterprises in our Trusted Access Program, with access through API and our Plus, Pro, Business and Enterprise plans coming in the coming days." The announcement adds Microsoft Azure and Amazon Bedrock as coming channels, says Astra usage in ChatGPT is included in existing subscription allowances, and notes that enterprise administrators must enable it for their workspace because access is off by default at launch.

On September 4 we checked every route we have. Our API organization got a model-not-found error for gpt-6-astra on both the Responses and Chat Completions endpoints. Bedrock in our AWS account listed the GPT-5.6 family only. Our ChatGPT account offered GPT-5.6 Sol and GPT-5.5. And OpenAI's own Playground, opened with the model preselected, showed the notice "GPT-6 Astra is rolling out over the next few days and may not be available" and fell back to GPT-5.5. Any launch-week article quoting Astra output on ecommerce copy was written by someone inside the access program, or was not written from Astra output.

On September 9 the model list at the /v1/models endpoint returned gpt-6-astra for our organization, with a created timestamp of August 27, 2026, and a one-line Responses API call completed. "Coming days" turned out to mean six. The rollout note has since been removed from the model page, which still lists a single undated model id. Everything in the results section below was run that afternoon.

The specs that matter for content work

Three of these rows matter more than the rest for a content pipeline.

Structured outputs and function calling mean the model can be held to a JSON schema, which is how a product-description generator guarantees it gets back a title, a meta description, an HTML body and alt text in separate fields rather than a blob of prose. Every serious content tool depends on this.

Image input means the model can look at the product photo while it writes, which is how a description mentions the actual colour and silhouette instead of guessing from the title. Images are billed as input tokens, and the model page lists image input as a supported feature at the standard rates.

The reasoning effort levels are the cost dial. A product description does not need "max". OpenAI's migration guidance says that if you were running earlier models at "none" or "minimal" effort you should "start with low and compare results." For copy, low is where we would begin.

Pricing, and what a description would cost

All figures per million tokens from OpenAI's pricing page, captured September 4, 2026.

Cache writes are $12.50 per million. Fast mode, per the announcement, "delivers up to 2x the speed of Standard processing at 2x the Standard price" and is not available with EU data residency. The long-context row applies to the entire request, not just the tokens past the threshold, so one oversized prompt doubles the input bill for everything in it.

How that compares with the models a store's content is written with today, and with the other vendor's top tier:

Now the arithmetic that matters, with measured numbers rather than projections. Our production product-description prompt for a floral wrap midi dress from a fashion boutique on our platform used 1,397 input tokens and 1,781 output tokens on GPT-5.6 Sol and cost $0.041. The same request on Astra used the same 1,397 input tokens, 1,445 output tokens, and cost $0.086. At low reasoning effort it used 1,185 output tokens and cost $0.073. So the "fewer tokens" claim is real, roughly 19% fewer at default effort and 33% fewer at low, and the description still costs about twice as much, because the per-token price is two and a half times higher. Our collection page went from $0.041 on Sol to $0.091 on Astra.

So the question for a store is not whether Astra writes better. At roughly twice the cost per page, it has to write enough better that the merchant stops editing the result, or stops regenerating it. Two rewrites of a four-cent description already cost more than one Astra description. That is the bar, and the rest of this post measures against it. If you want the background on how token prices turn into bills, we wrote a full guide on what AI tokens are and how to spend fewer of them.

What OpenAI claims that is relevant to copy

Strip the announcement and the developer guidance down to the statements that bear on writing product and collection content, and there are six.

It follows templates and matches your style. OpenAI says Astra "creates clear, well-structured documents, presentations, spreadsheets, and analyses that follow your templates and match your writing and visual style" and calls it "our best model for adhering to existing templates." Every ecommerce content generator is a template with facts poured in: heading, intro, features, specification table, care, call to action. Template adherence is the whole game.

It pulls only the context that matters. The announcement says Astra "is also trained to specifically pull only the context that matters into outputs, instead of repeating information unnecessary for the work at hand." A product prompt carries the merchant's existing description, tags, options, brand notes and a photo. Weaker models echo all of it back. A description that uses the relevant facts and leaves the rest is shorter, cheaper and better.

It makes better calls when instructions are ambiguous. "When instructions leave room for interpretation, GPT-6 Astra is better than previous models at making the right call. It uses context to fill in routine gaps and asks focused questions when the answer could change the outcome." Content prompts are full of soft instructions: include one honest consideration, keep the brand name if present, do not duplicate the title verbatim. How a model interprets those is where the quality lives.

It uses fewer output tokens. On Agents' Last Exam, OpenAI reports Astra "uses approximately 65% fewer output tokens than Opus 5" at the highest-scoring settings, and a customer quoted in the announcement reports "up to 20% fewer tokens than other models we've tested." Output is the expensive side of every bill, so if this holds for copy it takes some of the sting out of the $50 rate.

It makes fewer misleading claims. OpenAI's capability-hallucination evaluation shows "a substantial improvement over GPT-5.6 Sol." That evaluation is about the model's claims regarding its own abilities, not about product facts, but it points in the right direction. In copy, the failure that costs money is an invented specification: a fabric percentage, a dimension, a material the product does not have.

It stays inside its task. The alignment evaluation OpenAI built after its July incident found that GPT-5.6 Sol, run without production safeguards, went beyond the authorized target 48% of the time, while Astra did so in 0% of cases. For an agent that edits a live store, staying inside scope is not an abstract virtue. OpenAI's separate safety overview adds that Astra "is significantly more robust to prompt injections than GPT-5.6 Sol" and, in realistic browsing and workplace environments, "significantly less likely to perform misaligned and potentially destructive actions (for instance unauthorized transactions, data loss, excessive access, or circumvention of controls)." Prompt injection is the attack where a product page or a supplier document carries hidden instructions; a content agent that reads the open web to write copy is exposed to it every day.

One more line from the developer guidance is a caution rather than a claim. Astra "tends toward detailed, formatted responses." If you want two tight paragraphs, the guidance says to ask for "clear, concise paragraphs" explicitly. A model that defaults to more formatting is a model that has to be told, firmly, what a product description is not.

What OpenAI did not publish

There is no writing benchmark on the launch page. No marketing-copy evaluation, no readability study, no human-preference test on product or brand content. The Devin quote in the announcement mentions "writing" in passing, and the slide-deck demo shows template adherence on a presentation, not on a store page. That is the full published record on prose as of September 4.

It means that any claim you read this week about Astra being better at ecommerce content is either a guess extrapolated from the agent benchmarks or a test someone ran themselves. It is worth knowing which.

What "better at content" would actually mean

We write product descriptions and collection pages for Shopify stores for a living, and the failure modes are well known. A better model has to beat the current ones on these, in this order.

Fidelity to the source. The description must contain every specification the merchant already had, and nothing the merchant did not. The most common expensive error is a plausible invented detail.

Template adherence. Heading, then intro, then features, then a two-column details table, then care instructions, then a close, in valid HTML, inside valid JSON. A model that improvises the structure breaks the storefront theme.

Keyword handling. The target phrase appears in the heading, the intro, the SEO title and the meta description, naturally, once each, and never stuffed.

Judgment on soft instructions. Our prompt asks for one honest consideration, bans a list of tired phrases, forbids prices and shipping policy in the body, and tells the model the photo is for visual grounding only and must never be referenced. Compliance with these separates copy that reads like a shop owner from copy that reads like a machine.

Cost and speed. Tokens in, tokens out, seconds per request, and above all how many attempts it takes to get a description the merchant does not edit.

The test

Because OpenAI has not published this, we did. The harness uses two of our production prompts unchanged, and every model receives the identical request.

Product descriptions use the same instruction set our app sends: extract and preserve every fact from the existing description, rewrite in the required HTML structure, include one honest note, avoid the banned phrases, never mention prices or policies, treat the image as visual grounding only, and return a JSON object with title, SEO title, meta description, URL handle, description HTML and alt text. Two real products from stores on our platform: a floral wrap midi dress from a fashion boutique (100% polyester, three sizes, an existing description with care instructions) and a hand-blown glass sculptural vessel by a named designer from a design store (an existing description with almost no specifications, which is the harder case, because the model has to resist filling a mandatory table with things it does not know). Each request carries one product photo at low detail.

Collection pages use our collection prompt verbatim: write SEO-grade HTML for a collection's description field from a research dossier and a content plan, allowed tags only, link products by handle, use only the image URLs supplied, no filler phrases, every factual claim traceable to the dossier. One real collection: Casual Dresses, eight products, four keywords, six planned sections totalling 640 words.

Each output is judged on the five criteria above, and the usage field is recorded for tokens, cost and time. Astra was run twice, at its default reasoning effort and at low, which is where OpenAI's guidance says copy work should start. GPT-5.4 and GPT-5.6 Sol were run at their defaults, which is how our app runs them. Everything below was run on the afternoon of September 9, 2026.

Results, September 9, 2026

Product description, floral wrap midi dress, one image at low detail:

Product description, glass sculptural vessel, one image at low detail:

Collection page, Casual Dresses, text only:

Four things to read off the tables before any prose.

The input gap is a GPT-5.4 quirk, not an Astra advantage. The same request with the same photo billed 4,488 input tokens on GPT-5.4 and 1,397 on both Sol and Astra, so roughly 3,000 tokens of difference sit entirely in how the older model counts an image. Astra inherits Sol's counting.

The fewer-tokens claim holds. At default effort Astra used 19% fewer output tokens than Sol on the dress and 37% fewer on the vessel; at low effort, 33% and 46% fewer. On the collection page Astra's 1,238 output tokens against Sol's 1,475 hides a bigger gap, because Sol's figure includes 285 reasoning tokens and Astra's includes none, and Astra wrote 671 words to Sol's 585. Per word of finished copy, Astra was the leanest of the three.

The price still wins. Fewer tokens at two and a half times the rate is more money: 2.1 times Sol on the dress, 1.7 times on the vessel, 2.2 times on the collection page, and 2.4 to 3.5 times GPT-5.4. Low effort saved 15% on the dress and 12% on the vessel, and nothing on the collection page, where the default run had already used no reasoning tokens.

Astra was faster than Sol on the descriptions, 25 seconds against 32 at default and 14 to 20 seconds at low, and level with it on the collection page at around 31 seconds. GPT-5.4 remains the fastest at 11 to 14 seconds, because it does not think before it writes.

What Astra actually wrote

The criteria, in the order we said they mattered.

Fidelity. All three models passed, on both products, at every setting. On the dress everyone kept the 100% polyester, the three sizes and the full care instructions. On the vessel nobody invented a dimension, a weight or a technique, and nobody leaked the store's internal shipping-zone tags, which sit in the product record as strings like "UK TO UAE210" and would read as a specification to a careless model. Astra used the photo for what is plainly visible and nothing more: "Pink, orange and yellow flowers with green foliage on a dark background" in the colours row of the dress table, and a "lime-green accent band" on the vessel. Both are allowed under our rule and both are correct against the photos.

Template. Valid JSON and every required section, on every run, from every model. One real difference in how the rules were read. Our prompt puts specifications in the details table and says "never emit specs as bullets". GPT-5.4 and Sol wrote four or five feature bullets that restate specs anyway: wrap silhouette, tie waist, puff sleeves. Astra wrote three bullets that are benefits rather than specs: "A statement look for wedding celebrations." "An expressive choice for brunch with friends." "Easy to style with simple accessories." That is the more literal reading of our instruction, and it is the weakest section of copy any model produced. A merchant would rewrite those three lines. The model followed us and we asked for the wrong thing; the fix goes into our prompt, not into the model choice.

On the vessel, GPT-5.4 padded the mandatory table to ten rows with entries like "Visual Effect" drawn from the prose. Sol wrote seven. Astra wrote six and moved the maker's handmade-variation disclaimer into its own "Note" section under a heading, which is what our "preserve all sections from the original" instruction asks for. The other two folded it into a bullet or a table row.

Keywords. Every model put the target phrase in the H2, the intro, the SEO title and the meta description, once each, on both products. Astra's meta descriptions ran 149 to 160 characters and used the whole budget; Sol's and GPT-5.4's ran 127 to 152. On the collection page all three hit all four keywords with identical counts, the head term twice and the others once. That is worth a sentence, because a separate internal run of ours on six Swedish collection pages found Astra never echoing a target keyword verbatim, writing "ACME klassiska kaffekoppar i 6-pack" for the keyword "kaffekoppar 6 pack" while GPT-5.4 echoed the phrase as given. In English, with the keywords listed in the plan, Astra echoed every one. If you write for a non-English store, test this before you trust it.

Judgment on soft instructions. This is where the models separate, and it is the axis OpenAI's claims point at.

Our prompt asks for "one honest note or consideration". GPT-5.4 wrote: "the wrap styling gives you flexibility at the waist, but if you prefer extra coverage at the neckline, you may want to pair it with a subtle camisole or pin." A real consideration, and better than the disguised praise it produced in our September 4 run. Sol wrote: "the oversized, high-contrast pattern is designed to look bold rather than understated." Real, slightly abstract. Astra at default effort wrote: "Before heading out, check the wrap coverage while sitting and walking to find your preferred fit." That is the most useful sentence any model wrote about this dress. It is what a shop owner who has worn a wrap dress would say. Astra at low effort wrote: "A fit consideration: the waist is adjustable, but we don't have garment measurements to guide size selection." True, but it is a note about the listing's gaps rather than about the dress, and it landed in the closing paragraph. Lower effort produced a more literal, less helpful reading of the same instruction.

Then the product name. The merchant calls this dress "If You're Not the One Bubble Sleeve Midi Wrap Dress". Sol kept that name in the title. GPT-5.4 and Astra both replaced it with a generic "Floral Wrap Midi Dress" title, though Astra kept the original in the intro and in a "Style name" table row. Our prompt says to preserve the brand, and every model did; the quirk lives in the product name, which we never told them to protect. Sol read our intent. Astra read our words. That is another fix on our side, and a useful thing to know about the model: it does what the prompt says, precisely, and it will not guess at the parts you left out.

The collection page is where Astra pulled ahead. All three produced the six planned sections in order, three product links, three collection links, three image tags with the exact URLs from the dossier, and nothing outside the allowed tag list. The difference is what the copy is made of.

GPT-5.4's buying guide is generic: "Cotton is great when you want softness and breathability, while lightweight polyester can give you an airy drape and easy care." Sol's is a little more specific and adds a care question to the FAQ. Astra's names the products and their facts: "the Sunset Swing Tropical Shirt Dress uses lightweight polyester, while the Part Of My Story shirt dress is cotton," then "Blooming Affair has a flowy shape and abstract floral print," then "Living A Dream offers a sleeveless crochet look," and in the FAQ, "Hand washing is recommended for the Cup of tea lace midi; the cotton shirt dress lists machine washing cold." We checked every one of those claims against the eight product records in the dossier. All of them are there. That is the "pull only the context that matters" behaviour OpenAI described, applied to a page where it matters, and it is the kind of copy a merchant does not have to rewrite.

Astra's voice was also the closest to the brief, which asked for a warm, upbeat boutique owner talking to a friend: "Some days call for a dress and sneakers. Others need a little lace, a bright print, or a waist you can tie just how you like it." And later: "Brunch, a vacation dinner, or just Tuesday: there's room for a dress."

Two slips. In the FAQ Astra wrote "Yes" followed by an em dash and then "try swapping sneakers for strappy heels", once at default effort and twice at low, and the low run switched to curly apostrophes. Neither GPT-5.4 nor Sol wrote a dash anywhere across all twelve outputs in this test. Our house style bans the em dash in merchant copy because readers have learned to read it as machine-written, so on our platform that sentence would have been caught and rewritten before it reached a page. And at default effort Astra placed the product images as bare image tags and put the product links inside the headings rather than in the paragraphs. Both are allowed by the prompt. Both are a slightly different shape from what the other models produced, and a theme that styles figures would render them differently.

Where that leaves the price

On product descriptions, the honest summary is a same-day tie on everything that can be checked, a small edge to Astra on judgment at default effort, a small edge to Sol on reading intent, and a bill that is twice as large. Sol's dress description would ship with one edit. Astra's would ship with one edit too, the feature bullets. Same edits, double the cost. On this job Astra has not earned its price, and low reasoning effort does not rescue it, because the saving is 12% to 15% and the honest note got worse.

On the collection page the answer flips. Astra's page is the one we would publish with the fewest changes: strip one em dash. GPT-5.4's would need its buying guide rewritten to say something specific about the products, and Sol's sits in between. The difference in cost is five cents. A collection page is written once and read by everyone who lands on it from search, so five cents for a page that needs less editing and says more true things about the products is the easiest purchase in this post. If we were choosing a model for collection pages alone today, it would be Astra, with a keyword instruction and a dash check added to the prompt.

How to run it yourself

If you want to reproduce this rather than take our word for it, OpenAI's guidance gives most of the recipe and our runs add three notes.

Set the model to gpt-6-astra and remove the temperature, top_p and top_logprobs parameters, which the model does not accept. Use the Responses API. Ask for structured output against your JSON schema. Put the static instructions first and the product data last so prompt caching, which is on by default for prompts of at least 1,024 tokens, can reuse the instruction block at one tenth of the input price. Run anything a human is not waiting for through the Batch endpoint at half price.

Our notes. First, start at default effort for copy, not low. OpenAI's guidance says to start at low and compare; in our runs low saved 12% to 15% on descriptions and made the one judgment-heavy sentence worse, and it saved nothing on the collection page. Second, tell the model explicitly what the page is not, because the guidance is right that it "tends toward detailed, formatted responses", and tell it not to write em dashes, because it will. Third, hold every output against the source data before you look at the prose. Astra passed that test every time here; the model you are replacing probably does too, and the differences you are paying for are further down the list. Fidelity first, style second, cost third.

What this means for a store using an AI content tool

If you use an app rather than the API, the question is how the app chooses models. Inside our platform, every model lives in one registry with its verified price and its measured quality, and a new model enters as a candidate that cannot be routed to until it has been measured on the actual jobs. GPT-6 Astra went through that door on September 9. It is now a priced candidate in the registry. It is not the default for product descriptions, because on that job it tied the current model on correctness and cost twice as much. It is the model we are most likely to offer next for collection pages, once two things this test surfaced on our side are fixed: a prompt line that asks for the exact keyword phrase once, because of the Swedish paraphrasing described above, and a check that strips the em dash it likes to write. We would rather be a week late than wrong about the thing your customers read.

For the free version of that decision, our product description generator and collection description generator run on the current models, and our guides on how to write product descriptions and AI collection page generation for Shopify cover the parts no model changes.

Update log

September 9, 2026: GPT-6 Astra became callable from our standard API organization. Re-ran the full test on GPT-5.4, GPT-5.6 Sol and GPT-6 Astra (default and low reasoning effort) on the same afternoon, so every number in the results tables is same-day. One product substitution: the Heather Forest glass vessel used in the September 4 baseline is no longer listed on the store, so the test uses the same maker's Grey Magenta vessel, which has the same spec-free description shape. Re-checked pricing against OpenAI's pricing page on this date; unchanged. Rewrote the title, summary, availability section, cost arithmetic, results, verdict and FAQs. The September 4 baselines for GPT-5.4 and GPT-5.6 Sol were within a few percent of the September 9 runs on every metric, which is normal run-to-run variation for these models.

September 4, 2026: Published one day after the announcement. Specs, pricing, availability and claims taken from OpenAI's announcement, model page, pricing page and model guidance on this date. Baseline results run on GPT-5.4 and GPT-5.6 Sol. GPT-6 Astra returned model-not-found on our API organization and was absent from Bedrock, ChatGPT and the Playground for our accounts.