Introduction
Three days after OpenAI shipped GPT-6 Astra, the useful questions are no longer about FrontierMath. They are about whether a production workload can actually move onto it this week — and OpenAI's developer documentation answers several of those questions in ways the announcement did not. The full specification sheet lives on our GPT-6 Astra model page; our launch coverage handled the benchmarks and the ARC-AGI-3 harness dispute.
This piece covers what that coverage left out: the endpoint Astra is not served on, a billing threshold that can double a bill for two thousand tokens, a price cut on the previous generation that OpenAI did not put in a headline, the rate-limit ladder, and the one third-party score that puts the price in perspective.
Astra Is Not on the Realtime API
OpenAI's model reference lists three endpoints for gpt-6-astra: Chat Completions, Responses and Batch. Against Realtime, Realtime translation and Realtime transcription, it says "Not supported" — alongside Assistants, fine-tuning, embeddings, moderation, and every audio, image and video endpoint.
If you run a streaming-voice product, there is no Astra migration to plan.
The instinctive next question is whether this is a regression from the previous flagship. It is not. The model references for gpt-5.6-sol and gpt-5.6-luna mark Realtime "Not supported" too. OpenAI's voice work has been a separate product line for some time, served by a dedicated gpt-realtime family — gpt-realtime-2.1, gpt-realtime-2.1-mini, gpt-realtime-2, gpt-realtime-translate and gpt-realtime-1.5 — and those models are the only ones on v1/realtime.
That split matters more than a missing checkbox, because the two lines are not converging. gpt-realtime-2.1 carries a 128,000-token context window against Astra's 1,050,000, and a knowledge cutoff of September 30, 2024 against Astra's April 30, 2026 — roughly twenty months of drift. Its audio pricing is its own scale as well, at $32 per million audio input tokens and $64 per million audio output tokens, with text at $4 and $24.
So a product with both a chat surface and a speech-to-speech surface does not get one upgrade. It gets a frontier chat model and a voice model that still thinks it is 2024, and any shared prompt library has to tolerate both.
The 272,000-Token Pricing Cliff
The rule in OpenAI's documentation is one sentence: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request."
The load-bearing words are for the full request. This is not a marginal rate on the tokens above the line — the entire call reprices.
| GPT-6 Astra, per 1M tokens | Input | Cached input | Cache write | Output |
|---|---|---|---|---|
| Up to 272,000 input tokens | $10.00 | $1.00 | $12.50 | $50.00 |
| Above 272,000 input tokens | $20.00 | $2.00 | $25.00 | $75.00 |
Run the arithmetic on the input side of a single request:
- 271,000 input tokens — billed at $10 per million: $2.71
- 273,000 input tokens — billed at $20 per million: $5.46
The 2,000 tokens that crossed the threshold cost $2.75, an effective rate of about $1,375 per million tokens. Nothing else in the price list is within two orders of magnitude of that.
Two consequences for anyone building on long context. First, if your prompt sits anywhere near 272K, trimming it below the line is worth more than any prompt caching work you could do — and note that cache writes double above the threshold too, so a cliff-crossing request punishes caching rather than rewarding it. Second, variable-length inputs need a token count before the call, not after: a retrieval step that usually returns 250K tokens and occasionally returns 280K has a bimodal bill, and the expensive mode is invisible in an average.
The same threshold and the same multipliers apply across the GPT-5.6 family, at each model's own rates.
OpenAI Cut GPT-5.6 Prices at the Astra Launch
This did not make the launch materials, and it changes the routing maths more than Astra's own price does.
| Model | Before | Now | Cached input |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 / $30.00 | $4.00 / $20.00 | $0.40 |
| GPT-5.6 Terra | $2.50 / $15.00 | $2.00 / $12.00 | $0.20 |
| GPT-5.6 Luna | $1.00 / $6.00 | $0.20 / $1.20 | $0.02 |
OpenAI's model reference calls these promotional and says they are available at least through November 21, 2026, so budget them as provisional rather than permanent.
Two things fall out of the new table. GPT-5.5, at $5.00 / $30.00, now costs more than its own successor — more than Sol on both input and output, while carrying a knowledge cutoff eleven weeks earlier. There is no remaining argument for starting work on it.
And Astra at $10 / $50 is exactly 2.5x Sol on both input and output. "Upgrade everything to the new flagship" is therefore a 2.5x line-item increase on every workload Sol was already handling correctly, before any long-context surcharge. Luna at $0.20 sharpens the same point from the other end: Astra's input rate is 50x it, which is what makes a tiered architecture — cheap models for extraction and routing, Astra only where a wrong answer is expensive — the shape that keeps the bill sane.
Four Index Points for 2.5x the Price
Artificial Analysis published Intelligence Index v4.2 on September 4, 2026, and scored both models on it: GPT-6 Astra (max reasoning) at 55, GPT-5.6 Sol (max) at 51 on that same v4.2 index. At medium effort the gap widens slightly, 52 against 46. Astra places second on v4.2 overall, behind Claude Fable 5.1 at 57.
Four points of a composite benchmark for 2.5x the price is not an argument against Astra — aggregate scores compress exactly the long-horizon agentic work Astra was built for, and computer use does not appear in the index at all. It is an argument against migrating by default. The v4.2 revision also retired GPQA Diamond and doubled the weight of private held-out evaluations, so v4.1 and v4.2 numbers are not comparable; a v4.1 score you remember for either model is a different measurement, not a smaller one.
Rate Limits and the Window You Actually Get
Astra's rate limits scale with your usage tier:
| Tier | Requests / min | Tokens / min |
|---|---|---|
| Tier 1 | 500 | 500,000 |
| Tier 2 | 5,000 | 1,000,000 |
| Tier 3 | 5,000 | 2,000,000 |
| Tier 4 | 10,000 | 4,000,000 |
| Tier 5 | 15,000 | 40,000,000 |
Tier 1's 500,000 tokens per minute is the number to plan against: it is less than one maximum-length request per minute. Long-context evaluation on a new account is rate-limit-bound before it is budget-bound.
On the window itself — the context window is 1,050,000 tokens, but maximum input is 922,000 and maximum output is 128,000, and those two figures sum exactly to the total. The window is partitioned, not shared. Reasoning tokens are billed as output and count against that 128,000, so a high-effort request can spend a large share of its output budget before emitting a visible answer.
Astra's knowledge cutoff is April 30, 2026. Anything after that — including this release — is not in the weights, so an agent reasoning about current model pricing needs it in context or in a tool.
What We Could Not Confirm: Astra Pro and Fast Mode
Our launch post described a GPT-6 Astra Pro variant for Pro, Business and Enterprise plans, and priced Fast mode at $20 / $100 per million tokens. Checking both against OpenAI's developer documentation today:
There is no Pro model ID. OpenAI's model catalog lists exactly one GPT-6 model, gpt-6-astra, with no Pro variant and no dated snapshot. The GPT-5.6 generation still has four entries. So whatever Astra Pro is on a ChatGPT plan, it is a plan-level tier rather than something the API can address, and a developer who goes looking for gpt-6-astra-pro will not find it.
Fast mode is a service tier, not a model. The documentation says only that "Fast mode is priced at 2x the applicable rates", and elsewhere that priority processing was renamed Fast mode on July 30, 2026, selected with service_tier: "fast" (or the older "priority"). At standard context, 2x the applicable rates is $20 / $100, which matches the figure in our launch coverage. The wording is worth reading closely, though: the multiplier attaches to the applicable rates, which above 272K input tokens are already doubled.
Conclusion
The benchmark argument for GPT-6 Astra was settled on day one. The migration argument is a different exercise, and on today's documentation it comes down to three things: whether your product has a voice surface that Astra cannot serve at all, whether your prompts sit near a threshold that reprices the entire request, and whether the workload actually needs a model that costs 2.5x a now-cheaper GPT-5.6 Sol for four points of index score.
For a large class of production traffic, the right response to this launch is not to migrate. It is to re-price against the new GPT-5.6 rates, move anything still on GPT-5.5 down to Sol, and reserve Astra for the long-horizon agentic and computer-use work it was actually built for. The full specification — endpoints, rate limits, every price tier and the safety restrictions on its cybersecurity capabilities — is on our GPT-6 Astra model page.
Sources
- GPT-6 Astra model reference — endpoints, pricing, rate limits, the 272K rule
- GPT-5.6 Sol model reference — reduced pricing and the promotional window
- GPT-5.6 Luna model reference — endpoint support for the previous generation
- gpt-realtime-2.1 model reference — the voice line's context window, cutoff and audio pricing
- OpenAI model catalog — the full model list, including the absence of a GPT-6 Pro entry
- OpenAI API pricing — standard and long-context rates, and Fast mode
- Announcing Artificial Analysis Intelligence Index v4.2 — methodology change, September 4, 2026
- GPT-6 Astra vs GPT-5.6 Sol — Artificial Analysis comparison