GPT-6 Astra Pricing: 48x Your Chatbot’s Reply Cost

Two things are true this week, and a WordPress site owner cannot act on both. The first is that large language models keep getting cheaper: the frontier token price index sits 84% below its 2023 base, and a dozen trackers spent the summer publishing charts of that decline. The second is that OpenAI’s newest model, gpt-6-astra, prices a single reply from this site’s chatbot at 48 times what the model we run today charges for the same reply.

Neither number is wrong. They are measured from different things, and only one of them is a bill. This post prices GPT-6 Astra the way an invoice would: against a real WordPress chatbot workload, measured on this install, with every rate read from the vendor’s own card rather than from the coverage.

What OpenAI actually published

Astra is on OpenAI’s public rate card as of 5 September 2026. The standard text rate is $10.00 per million input tokens and $50.00 per million output tokens, with cached input at $1.00 and cache writes at $12.50. The model card gives a 1,050,000-token context window, a 128,000-token maximum output and a knowledge cutoff of 30 April 2026.

A note on dating, because this site has been caught by it before: several secondary write-ups give the launch as 3 September. I could not confirm that at the origin, so this post makes no launch-date claim. The prices below are what the vendor’s own pages served on 5 September, and that is all they are.

What matters more than the headline pair is that the card publishes six different ways to be billed for the same model. Three multipliers stack on top of the standard rate, and the differences between them are larger than the differences between most competing models.

Six lanes, five distinct prices — Batch and Flex land on the same number by different routes. That distinction is not cosmetic, and it is the first place a cost estimate goes wrong.

What that costs on a real chatbot reply

Every figure from here on runs against one measured profile. On 2 September we queried this install’s own stored transcripts rather than its configuration, and the result was 3,760 input tokens and 170 output tokens per reply: a 3,478-character system prompt, a retrieval context averaging 11,426 characters across six samples, and a user message averaging 135 characters across fourteen. The replies themselves averaged 681 characters across ten. Those are small samples and we are not going to pretend otherwise, but they are measurements of a live bot rather than a guess.

Multiply that profile through the six lanes and the spread is wider than the model catalogue it sits in.

Two of those six bars are grey because a visitor-facing reply can never be billed at them.

The Batch lane is asynchronous. It accepts a file of requests and returns results later, which is a perfectly good way to embed a knowledge base and a completely useless way to answer someone who is waiting in a chat widget. We made this same point about OpenAI’s Batch API on this workload three days ago and it has not stopped being true: the 50% discount is real and it reaches the reply path at no vendor.

The long-context lane is unreachable for the opposite reason. It triggers at 272,000 input tokens. Our prompt is 3,760 — 1.38% of the threshold. It is a price that exists for a workload this is not, and it belongs in the chart only because a reader scanning a rate card will otherwise wonder why nobody mentions it. (The same shape turned up on xAI’s long-context rate, where our prompt reached 1.9% of the trigger.)

Flex is the honest ambiguity. OpenAI’s model card prices it at 50% of standard and does not say on that page what Flex is; the documentation page that would define it returned a 404 when we went looking. It is drawn as reachable here because it is a synchronous service tier rather than a batch queue, but a 50% discount whose service guarantees you cannot read is not a discount you should budget against. Treat the $23.05 row as a ceiling on savings, not a plan.

That leaves a reachable spread of $23.05 to $92.20 per 1,000 replies — 4x, on one model, on one day, decided entirely by which tier flag your integration sets.

Where Astra sits in OpenAI’s own catalogue

The comparison that matters to a site owner is not Astra against Claude or Gemini. It is Astra against the OpenAI model they are already running. Here is the current GPT-5.6 line and Astra, priced identically on the same 3,760-in / 170-out reply.

Astra is 2.5x Sol, which is the number every summary led with. It is 48.2x Luna, which is the number that describes what happens if a WordPress site owner reads “the new model is out” and changes a dropdown. Luna’s $0.956 is not a rate-card estimate either; it is the corrected figure we published on GPT-5.6 Luna’s cost per 1,000 replies after discovering we had originally priced the bot off its configured ceiling instead of its measured behaviour.

Note also that Astra is not the top of OpenAI’s own card. gpt-5.6-cyber, a security-specialised variant, is dearer. The frontier is not a single price and “the newest model” is not the same claim as “the most expensive model”.

On this workload, model choice is an input-rate decision

A retrieval chatbot is a lopsided thing. Of the 3,930 tokens in an average exchange here, 95.67% are input — system prompt plus six retrieved knowledge-base chunks — and only 170 tokens come back out. Astra’s output rate is five times its input rate, which sounds like the number to optimise, and it is not: because output is 4.3% of the tokens, input is 81.56% of the Astra bill.

This has a practical consequence that cuts against how the model is being benchmarked. Reasoning-effort settings, which the benchmark sites are busy charting as cost-per-task, mostly change how many output tokens a model produces. On a workload where output is four cents in every twenty-two, effort barely moves the invoice. The same conclusion held for Grok 4.6 and for DeepSeek’s peak-hours change; Astra makes it four vendors in a row.

The identical twin nobody is comparing it to

Four days before Astra appeared, Anthropic shipped Claude Fable 5.1. Its published rate is $10.00 input and $50.00 output — the same two numbers, to the cent. Its five-minute cache write is $12.50, also to the cent.

There is exactly one meter where the two cards disagree.

The gap comes from a footnote. Anthropic’s pricing page carries the line that cache hits on Fable 5.1 “are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier.” Astra uses the industry-standard 0.1x. Fable 5’s predecessor used 0.1x too, at $1.00 — so this is a change Anthropic made at the 5.1 release and put in small type under a table.

On a retrieval chatbot, the cache-read rate is the one meter you would pick if you could only see one. Input is 95.67% of the tokens and a fixed slab of that input never changes.

What a 4x footnote is actually worth

Here is where the honest answer disappoints, and where most comparison posts would quietly stop.

The system prompt is 870 tokens, 23.14% of every input, and it is byte-identical on every single reply. That is the part you can cache with confidence. Doing so on Astra takes the bill from $46.10 to $38.27 per 1,000 replies, a 17.0% saving. Doing the same on Fable 5.1 gives $37.62.

So a 4x difference on the rate card is worth 1.7% on the invoice.

The bottom row is a ceiling, not a plan. 76% of every input is retrieval context that changes with the question, so it can only be served from cache on an exact repeat of a previous query. Some fraction of real traffic is exactly that — people ask the same three things — but we have not measured what fraction on this site, and quoting the ceiling as though it were the practice would be the same error we made pricing the bot off its configuration instead of its transcripts.

The transferable rule: a multiple on a rate card is not a multiple on your bill unless the meter it applies to carries most of your tokens. Cache read is a 4x difference on a meter that touches 23% of one side of the ledger, and 4x of 23% of 81.6% is a rounding error.

“84% cheaper” and “48x dearer” are both true

Which brings us back to the two sentences at the top.

The frontier price index that fell 84% is a median across a widening catalogue. Its constituents change every month, and this year they have overwhelmingly widened downward: Luna at $0.20 input, Haiku-class models at $1.00, open-weight models at pennies. The median of that set falls even while the ceiling rises. The index is measuring how cheap a good enough model has become. It is not measuring, and has never measured, what the newest flagship costs.

Yesterday we published a version of this same trap at the level of a single model pair, where the same Qwen model swap came out as −20%, +35.6% or +60% depending only on which price of the older model you measured from. This is the market-wide edition. A percentage change is not a number until somebody names what it was measured from, and an index quietly names a constituent set that nobody reads.

You can watch the same failure happen live on the search results for this model. One page reports Astra as “2.5x GPT-5.6 Sol’s current prices across the board”. A page two results below reports it as “75% more expensive than GPT-5.6 Sol at max effort”. Both are describing the same launch. The first is rate-per-token; the second is cost-per-task, which folds in how many reasoning tokens each model burns to finish a job. They differ by a factor of two, they sit one rank apart, and neither says which it is in its headline.

What this means if you run MxChat

Three practical answers, and the first one is the least exciting.

You cannot switch to Astra today. We checked rather than assumed: the MxChat plugin source contains zero gpt-6 strings. Its OpenAI catalogue currently offers gpt-5.6-luna, gpt-5.6-sol and gpt-5.6-terra among the older families, and this site is configured on Luna. Astra will appear in a future release; it is not a dropdown you can change this afternoon.

When it does appear, the question is not “is it better” but “is it better at answering questions about your own documentation”. A retrieval chatbot’s job is to read six chunks of text you supplied and paraphrase the relevant one. That is not the task the frontier benchmarks measure. At $46.10 against $0.96 per 1,000 replies you are paying a 48x premium for reasoning depth on a task with very little reasoning in it. There are sites where that is right — a legal or medical bot that must refuse carefully, a sales bot doing multi-step configuration — and there are many more where it is not.

The lever that actually moves your bill is the retrieval context, not the model. Six chunks at 11,426 characters is 76% of every input on this install and it is set by rag_sources_limit. Dropping from six sources to four cuts roughly a quarter of the input on every reply at every vendor, permanently, and unlike a model swap it costs nothing to test. Our guide to chatbot running costs works through the arithmetic, and the MxChat documentation covers where the setting lives.

The short version

Frontier prices going up while the median goes down is not a contradiction, and it is not a scandal. It is what happens when a market stratifies. The only mistake available here is reading a statistic about the middle of the catalogue and applying it to the top — and then reading a statistic about the top and applying it to your invoice.

If you want the same treatment applied to the other vendors: GPT-5.6 Sol’s dated cut, the Sonnet 5 rise that was cancelled, and DeepSeek’s peak-hours pricing each priced the same measured reply. The MxChat add-ons that generate content run on a separate model setting from the chat model, which is worth remembering before you attribute a bill to the wrong dropdown.