Is GPT-6 out? Yes — and this article is about the version of that question developers actually mean. OpenAI released GPT-6 Astra on September 3, 2026 (announcement), and the search results are mostly launch news, AGI commentary, and ChatGPT subscription chatter. If you are deciding what the model costs to call over an API — as opposed to what a ChatGPT plan includes — the pricing is more layered than the “$10 / $50” headlines suggest, and most of the guides ranking right now predate the rate card that went live with the model.
So here is the API answer, with every figure captured on September 8, 2026: OpenAI prices GPT-6 Astra at $10 per million input tokens and $50 per million output tokens, with cached input at $1 and cache writes at $12.50 — and prompts longer than 272K input tokens reprice the entire request at 2× input and cache rates and 1.5× output. Through Modellix, an API gateway whose LLM price page lists GPT-6 Astra (as openai/gpt-6-astra) alongside 28 other language models behind one key, the same model bills at $9 input / $45 output / $0.90 cached read / $11.25 cached write on the ≤272K tier, with the same 10% OFF applied above 272K ($18 input / $67.50 output / $1.80 cached read / $22.50 cached write), and a Discount column that reads “10% OFF”. We run Modellix, so we have a commercial interest in the gateway numbers below; where a figure could not be verified against OpenAI’s own documentation, we say so.
This is a pricing guide for one model ID — what GPT-6 Astra costs, how the 272K tier and the cache columns change the bill, why a $10/$50 sticker is not automatically one bill, and how to audit the run afterward. It does not review capability, does not answer consumer questions about ChatGPT plans, and does not decide whether you should pin a concrete model ID or ride a moving alias — that decision is covered separately, and this guide assumes you call the model directly.
GPT-6 Astra pricing: the September 2026 rate card
All rates are USD per 1 million tokens. GPT-6 Astra is not flat-rate: it has two Input Context tiers split at 272,000 input tokens, and every dimension has its own price in each tier. The table below lists both tiers as shown on the Modellix LLM price page on September 8, 2026, in the page’s own “official list price → what you pay” formatting:
| Tier | Dimension (per 1M tokens) | Official list price | On Modellix |
|---|---|---|---|
| Input Context ≤ 272K | Input | $10.00 | $9.00 |
| Output | $50.00 | $45.00 | |
| Cached read | $1.00 | $0.90 | |
| Cached write | $12.50 | $11.25 | |
| Input Context > 272K | Input | $20.00 | $18.00 |
| Output | $75.00 | $67.50 | |
| Cached read | $2.00 | $1.80 | |
| Cached write | $25.00 | $22.50 |
The Discount column prints “10% OFF” verbatim for openai/gpt-6-astra, and the 10% applies to every dimension in both tiers — each Modellix price above is exactly 90% of its listed official price.
Two attribution notes before you use the table. First, the “official list price” column is Modellix’s label, but this time every number in it cross-checks against OpenAI’s own GPT-6 Astra model documentation, captured the same day: the docs state “Input $10.00 / Cached input $1.00 / Cache writes $12.50 / Output $50.00” per 1M tokens verbatim, including the cache-write rate. Second, for the >272K tier the OpenAI docs do not print a second table — they state that prompts over the threshold “are priced at 2x input and cache rates and 1.5x output for the full request.” Modellix’s listed second-tier figures ($20 / $75 / $2 / $25) line up with that multiplier exactly, which makes this the one row set where a gateway table and the vendor’s official wording reconcile item by item.
Two specifications deserve the same date-and-source treatment. The Context column on the Modellix price page reads 1M tokens (per the Modellix pricing page, September 8, 2026), while OpenAI’s model docs state a 1,050,000-token context window (per developers.openai.com, same day). We print both rather than pick one: which number governs depends on which surface you read, and neither source had reconciled the two as of this writing. Separately, OpenAI’s docs list 128,000 max output tokens — that is an OpenAI-specified limit, and the Modellix table does not list a max output at all, so we do not state it as a gateway commitment.
The 272K tier: what a long prompt does to the price
The tiering mechanism is not new — the GPT-5.6 family used the same 272K line, and our GPT-5.6 pricing guide explains how Input Context tiers work in general. What matters here is Astra’s specific arithmetic, and the part that surprises people: the higher rates apply to the whole request, not just the tokens past the line. OpenAI’s wording, quoted: “Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request.”
Two illustrative single requests make the consequence concrete (no cached reads, so the tier decision depends only on the input-token count — which is what the pricing page’s calculator keys on; it notes that “the input-token count automatically selects the matching Input Context tier”):
| Illustrative request | Input tokens | Output tokens | Official list total | On Modellix |
|---|---|---|---|---|
| Request A (≤ 272K) | 250,000 | 200,000 | $12.50 | $11.25 |
| Request B (> 272K) | 300,000 | 200,000 | $21.00 | $18.90 |
Request A bills at base rates: 0.25 × $10 input plus 0.2 × $50 output = $12.50. Request B sends only 20% more input tokens, but it crosses 272K, so the entire request moves to tier-2 rates: 0.3 × $20 input plus 0.2 × $75 output = $21.00 — roughly 68% more for the same output, because the output rate also rose 50% on a request that generated no extra tokens. On Modellix the same two requests come to $11.25 and $18.90, a straight 10% under each official total. If a prompt sits anywhere near the line, the tier — not the sticker price — decides your bill; the same input-token count that selects the tier is reported back in the request logs, which matters for the audit step later in this guide.
Illustrative Modellix artwork, not a product screenshot: two Input Context billing lanes split at the 272K threshold. No dollar figures are shown.
Same price tag, different bill: GPT-6 Astra vs Claude Fable 5.1
Here is the comparison no single-vendor pricing page can show you. Standing at the Modellix price table on September 8, 2026, two rows share an identical headline — openai/gpt-6-astra and anthropic/claude-fable-5.1 both list at $10 input / $50 output and both bill at $9 / $45 on Modellix — yet their bills diverge the moment caching or long inputs enter:
| Dimension (per 1M tokens, Modellix actuals) | openai/gpt-6-astra | anthropic/claude-fable-5.1 |
|---|---|---|
| Input (official list / on Modellix) | $10.00 / $9.00 | $10.00 / $9.00 |
| Output (official list / on Modellix) | $50.00 / $45.00 | $50.00 / $45.00 |
| Cached read | $0.90 | $0.225 |
| Cached write | $11.25 | $13.72 |
| Pricing structure | Two tiers at 272K | Flat-rate (no tiers) |
The cached-read gap is the starkest: $0.90 vs $0.225 per million — 4× — on identical $10/$50 stickers. A cache-heavy month that reads, say, 10 million cached tokens (an agent reusing a large prefix across turns, each request staying under the 272K line) spends $9.00 on that line with Astra and $2.25 with Fable. Cached writes invert the comparison — Astra’s $11.25 sits below Fable’s $13.72 (the latter is Modellix’s listed figure for Fable; Anthropic’s own pages do not quote a cache-write rate, as our Claude Fable 5.1 price guide documents). And Fable is flat-rate while Astra splits at 272K, so a request that crosses the line on Astra quietly adds 2× input and 1.5× output to a workload whose Fable equivalent would not.
Read that table as billing structure only. It compares rate cards, not models — we are not judging Astra against Fable 5.1 on quality, and we are not telling you which to choose. The point is narrower and easier to miss: “the same price” between two vendors is not the same bill once cache multipliers and tiering exist, and that difference is only visible from a table that carries both vendors’ rows at one date — exactly what a 29-model gateway table is.
What GPT-6 Astra costs on Modellix — and how to reproduce it
The gateway delta is real and quotable: the Modellix price page’s Discount column reads “10% OFF” for openai/gpt-6-astra on September 8, 2026, and both worked examples above reproduce on the page’s own cost calculator, which prices from the same displayed rates. Fill in Request A (250,000 input, 200,000 output, 0 cached) and it returns $11.25 against an original price of approximately $12.50 — about 10% less; Request B returns $18.90 against approximately $21.00. Treat the 10% as a same-day snapshot, not a promise: this catalog has already moved discount tiers within 24 hours once this week (the Anthropic family went from 3% to 10% OFF overnight in early September), so re-check the price page before you budget.
Two structural facts make the gateway price worth more than the discount line. First, the same key that reaches Astra reaches 28 other language models — including Fable 5.1 in the row above and DeepSeek’s V4 Flash at $0.396 per million input tokens — so “try a cheaper model” is a one-string change in the request body, not a new account; our cross-vendor LLM API pricing comparison holds a full 28-model snapshot taken September 3, 2026 (the catalog has since grown by one row — GPT-6 Astra joined it this week). Second, one pricing footgun deserves its own warning: Modellix also serves the stable alias ~openai/gpt-latest, and as of September 8, 2026 it points at openai/gpt-6-astra — meaning clients who call the alias are now on Astra’s rate card without changing a line of code, and OpenAI itself offers no -latest alias (only the single gpt-6-astra snapshot). Whether to pin a concrete model or ride the alias is the decision our GPT latest model guide takes on; the numbers above assume you call openai/gpt-6-astra directly.
Calling it is a standard OpenAI-compatible request: point an OpenAI SDK or any Chat Completions / Responses client at https://llm.modellix.ai/v1 with your Modellix key and "model": "openai/gpt-6-astra" — no Modellix SDK required, and what “OpenAI compatible” does and does not mean is covered in its own guide. One boundary to keep straight: this is a text gateway — image and video generation run on the separate media API at api.modellix.ai (async tasks), and the text host takes text prompts, so GPT-6 Astra’s text/image input modality only matters on surfaces that accept images. If the media side is what you are comparing, that is the OpenAI video-generation API topic, not this rate card.
Verify the bill after the run: the request-log endpoint
A rate card predicts what a request should cost; the step every launch-week pricing post skips is checking what it did cost — which tier your requests actually landed in and whether the cache hits you assumed happened. On the Modellix LLM gateway every request is logged, and the Get LLM logs documentation specifies the read endpoint: its OpenAPI block lists the operation as GET /logs against the gateway’s /v1 server base, so the full request URL is https://llm.modellix.ai/v1/logs:
1 | curl -sS "https://llm.modellix.ai/v1/logs?start_time=1757116800&end_time=1757203200" \ |
start_time and end_time are required UNIX seconds and may span at most 30 days; mdlx_user_id optionally filters by the end-user ID you sent as X-Mdlx-User-Id on inference. Each returned request carries prompt_tokens, completion_tokens, and cached_tokens — the two numbers that decide the 272K line, plus the tokens that actually billed at the cached-read rate — along with cost and the model as { provider, model_name }. The cost unit is sub-pennies, where 1 = $0.0001: divide by 10,000 to get dollars, so a request logged at cost: 1125 billed $0.1125. With that, the audit is arithmetic: check a request whose logged input count crossed 272K against tier-2 rates, confirm cached reads billed at $0.90 rather than $9.00, and sum the run. (The documentation does not state how long logs are retained, so pull what you need within the 30-day query window.)
Illustrative Modellix artwork, not a product screenshot: the log gives you the token buckets that decide the tier, and the cost to check against the rates above. No dollar figures are shown.
GPT-6 Astra on the Modellix LLM Gateway
Read the LLM API guide and request-log reference for model IDs, the /v1/logs fields, and billing notes in the Modellix documentation.
View DocsFrequently Asked Questions
Is GPT-6 out?
Yes — GPT-6 Astra was released on September 3, 2026. The API model ID on OpenAI’s side is gpt-6-astra; on Modellix it is openai/gpt-6-astra, and it has been in the catalog since the week of release. Some search results still ask whether GPT-6 will exist — those pages predate the launch by days.
How much does GPT-6 Astra cost?
On OpenAI’s rate card, captured September 8, 2026: $10 per million input tokens, $50 per million output tokens, $1 cached input, and $12.50 per cache write — and prompts above 272K input tokens bill the full request at 2× input and cache rates and 1.5× output ($20 / $75 / $2 / $25 list). On Modellix, the same model is $9 / $45 / $0.90 / $11.25 on the ≤272K tier — and $18 / $67.50 / $1.80 / $22.50 above 272K, the same 10% OFF applied to both — with a Discount column reading “10% OFF”.
What context window does GPT-6 have?
OpenAI’s model documentation states 1,050,000 tokens; the Modellix price page’s Context column reads 1M tokens (both as of September 8, 2026). We list both because the sources disagree and neither had reconciled the number as of this writing. Max output is 128,000 tokens per OpenAI’s docs; Modellix does not list a max output.
Is GPT-6 Astra free?
No. It is a paid API model billed per token from usage, and there is no free route to it on Modellix: the $1 signup credit ended on August 19, 2026 (trial credit is now by email request), and the two $0-priced rows on the LLM price page — modellix-ai/free-llm and zai/glm-4.7-flash — are unrelated to OpenAI and carry no usage documentation.
Is GPT-6 Astra the same as ChatGPT’s GPT-6?
The model is OpenAI’s; the meter is not. ChatGPT subscription plans bundle usage on OpenAI’s chat and coding surfaces, while the API bills per token with no cap and no plan. This guide covers only the API side — if you are asking which ChatGPT plan includes GPT-6, that is a subscription question, and the answers there say nothing about per-1M-token cost.
Does Modellix’s 10% OFF apply to both tiers?
Yes — on September 8, 2026 the discount applied to every dimension of both the ≤272K and >272K tiers (list $10/$50/$1/$12.50 and $20/$75/$2/$25 each billed at 90%). Discounts on this table have changed within 24 hours before, so verify the current Discount column before committing spend.
How do I know which tier my request hit?
From the request logs: GET /v1/logs returns each request’s prompt_tokens, cached_tokens, and cost. If the logged input count crossed 272,000, the whole request billed at tier-2 rates — compare the cost against the tier arithmetic above to confirm.
Check Current GPT-6 Astra Rates on Modellix
Log in to see the live listed rate for openai/gpt-6-astra, run the price-page calculator on both tiers, and pull per-request cost logs on one key.
LoginPricing and model availability change without notice. Figures above were captured September 8, 2026 from the Modellix LLM price page, from OpenAI’s GPT-6 Astra model documentation, and from the OpenAI launch announcement; the Fable cache-write figure shown is Modellix’s listed rate, which Anthropic’s own pages do not quote. Modellix is an API aggregator with a commercial interest in this guide and is not affiliated with OpenAI; the “10% OFF” is a same-day snapshot, so verify current rates against the live sources before committing spend. Access language models from OpenAI, Anthropic, Google, DeepSeek, Qwen, and more through a single API key at modellix.ai.
