Short answer

We found no public release entry for GPT-6.2 Astra in the official API changelog reviewed on November 7, 2026 (Asia/Shanghai). A third-party report says the string gpt-6.2-astra appeared in Codex usage, which is not an official launch and does not show that every account can call it. Price, context length and a release date are unconfirmed. Do not copy GPT-6 Astra’s specifications into those fields.

Where did the name come from?

Keep three layers separate:

  • Usage label: a string on a usage or billing view. It describes that recorded line.
  • Picker name: the product name a person can select. An older picker label does not identify which model wrote a given answer.
  • Public API model ID: a model value that the docs, the changelog and your own key agree on. Production traffic needs this layer.

On November 2, 2026, TestingCatalog (byline: Dwaybe Cubbins) reported that shared Codex usage records showed gpt-6.2-astra on October 31, while the picker still said GPT-6 Astra. The article cites social posts, says this is not an official announcement, and notes that a shared artifact is hard to attribute to one model. We did not open those original posts, so screenshot shares, output attribution and “launching soon” claims stay outside the fact table. A usage string is a third-party lead.

What is confirmed, and what is not?

The right-hand column is the documented model GPT-6 Astra. None of it is a GPT-6.2 Astra specification.

QuestionGPT-6.2 AstraPublished GPT-6 Astra
Release statusNo public changelog entry in this reviewAPI changelog records a release on 2026-09-03
Model IDThird-party records show gpt-6.2-astra; the docs URL we requested returned 404gpt-6-astra
Public APIRoute and account access unconfirmedDocs support Responses, Chat Completions and Batch
PriceUnconfirmedStandard, up to 272K input: input $10, cached input $1, cache writes $12.50, output $50 per 1M tokens
ContextUnconfirmed1,050,000 window; 922,000 max input; 128,000 max output tokens
Token KingNot listed as a callable model on this siteThe homepage price table has a GPT-6 Astra row; live access is whatever the console shows

Checked 2026-11-07. Sources: API changelog, GPT-6 Astra model page, a request to the gpt-6.2-astra docs URL, and TestingCatalog. A 404 means that URL had no public page at the time. It does not, by itself, prove that no internal test exists.

What are the release date, price and context length?

Release date: unknown. On the API changelog opened on 2026-11-07, the newest November entries were GPT-6-Cyber on November 5, GPT-Image-2.6 on November 4, and outbound SIP for the Live API on November 2. GPT-6.2 Astra was not among them. The same period on the product release notes covers those capabilities, not this model. The newest dated section we read in the Help Center model release notes was October 30, 2026, and it does not name GPT-6.2 Astra. The model catalog we opened still presented published flagships such as GPT-6 Astra and GPT-6.1 Sol.

API price: unknown. There is no official rate to cite. GPT-6 Astra’s $10 / $50 figures are that model’s Standard short-context text rates, not a 6.2 price.

Context window and max output: unknown. Window, maximum input and maximum output are different limits. A future window size would still not be a promise of reliable recall or of how much you can send in one request.

How do you check access and the model ID?

The official list-models method is GET /v1/models. The docs say it lists models currently available to the credential, with fields such as id and owned_by. The list is not a capability sheet, a price card or a public launch date. We did not run this call.

curl https://api.openai.com/v1/models \
  -H "Authorization: Bearer $OPENAI_API_KEY"

Keep the key in your own server or local environment. Do not put it in frontend code, screenshots or logs. Read the result in this order:

  1. Match the exact string. An extra dot or a missing hyphen is a different ID.
  2. If the ID is present, this key can see it right now. Open the official model page for that same ID and check endpoints, price conditions and permissions before sending traffic.
  3. If the ID is absent, this key cannot see it right now. That is not a worldwide availability ruling, and a failure is not evidence that the model exists.
  4. Check a gateway separately. A public official ID does not mean the gateway uses the same string or has opened a route.

Published models are the baseline. The gpt-6-astra page says inputs are text and image, output is text, and reasoning.effort accepts low, medium, high, xhigh and max. The changelog also says this model does not support none or custom temperature, top_p and logprobs, and that tool calling uses the Responses API. Those limits belong to GPT-6 Astra. Do not assume them for an unconfirmed ID.

How do you budget the models that are already published?

The arithmetic below uses official rates for published models, so a team can budget an evaluation. There is no rate to substitute for GPT-6.2 Astra.

GPT-6 Astra and GPT-6.1 Sol Standard text rates below apply when input is at most 272K tokens. Amounts are US dollars per million tokens. Cached input, cache writes and output are separate charges. Batch, Flex, Fast, Ultrafast, regional processing and tool calls are excluded from the examples.

Published modelInputCached inputCache writesOutput
GPT-6 Astra$10$1$12.50$50
GPT-6.1 Sol$2$0.10$2.50$10
GPT-6.2 AstraUnconfirmedUnconfirmedUnconfirmedUnconfirmed

Sources: the Astra model page, the Sol model page, and the September 29, 2026 Sol release note. The pages print $12.5 and $2.5; those are the same dollar amounts as $12.50 and $2.50.

Short-context example: one request with 100,000 uncached input tokens and 20,000 output tokens, no cache writes and no tools. This is rate arithmetic, not an invoice.

  • GPT-6 Astra: 0.1 × 10 + 0.02 × 50 = $2
  • GPT-6.1 Sol: 0.1 × 2 + 0.02 × 10 = $0.40

The official pages say that when input exceeds 272K tokens, the full request is priced at 2× input and cache rates and 1.5× output. Do not apply the short-context unit rates to a 1M-token prompt. Under that long-context rule, 1,000,000 uncached input tokens plus 200,000 output tokens would be $35 for Astra and $7 for Sol: 2 × input rate × 1 + 1.5 × output rate × 0.2. Equal token counts are not equal task cost. Reasoning, retries, human rework and tools all change the bill.

The Token King homepage price table currently lists configured models such as GPT-6.1 Sol and GPT-6 Astra. Live rates and access are on the console. Neither the price table nor the model cards list GPT-6.2 Astra. Site prices are not the official rates above. Compare them on the homepage price table before you integrate.

What should a migration review record?

This is an engineering checklist, not a model evaluation we ran. Until a new model has docs and your key can actually see it, run the process on a model you are already allowed to use, and keep the model ID in configuration.

  1. Fix the tasks. Use real work with acceptance checks, such as a small repository change or a research summary that must cite each source. One demo is not that set.
  2. Fix the conditions. Record the exact model ID, date, reasoning.effort, input and output tokens, tool permissions, retries, latency and final cost.
  3. Separate the billing thresholds. Check whether input exceeds 272K, whether the request writes cache, and whether any tool has a per-call fee before comparing models.
  4. Read quality and cost together. Record first-pass success, how often a person has to edit the result, and the total cost of each accepted result. Count failed retries.
  5. Keep a fallback. If the new ID is missing or the quality bar fails, the request can return to a documented model. One error code does not prove that a rumored model exists or that it has never appeared anywhere.

For long context, confirm the maximum input, the billing threshold and your own truncation rules first. Vision and computer use are directions on the GPT-6 Astra docs; keep a person in the loop for sensitive actions on external systems. For a closer look at published Sol and Astra rates, see the GPT-6.1 Sol pricing and API guide. That article is about released models.

FAQ

Has GPT-6.2 Astra been released?

As of this review on 2026-11-07, the official API changelog has no public release entry. A third-party usage report is not an announcement.

Is gpt-6.2-astra a callable model ID?

The string appears in a third-party report about usage records. This review found no official public route and did not check access with any account. Use the models list, then the provider’s docs.

Why would usage show a name the picker does not?

TestingCatalog reported that difference. A test route or an internal split is a possible explanation, not an official one, and it does not identify which model produced a given artifact.

Is it the same model as GPT-6 Astra?

Keep them separate. GPT-6 Astra has a model page and a 2026-09-03 changelog entry. In this review, GPT-6.2 Astra has third-party leads only. Parameters, prices and access stay with the model that documents them.

Should we wait, or build now?

Build the evaluation and the budget on a model you can already call and that has documentation. Keep the model ID configurable. Compare on the same tasks after a changelog entry, a model page and your own models list agree.

What if the request says the model was not found?

Check spelling, endpoint, project and what that key is allowed to see, then compare with the models list. We did not reproduce that error. One error neither proves worldwide access nor proves the ID has never appeared anywhere.

Can Token King call GPT-6.2 Astra today?

Do not treat the rumor as availability. This site does not currently list the model. Check the models and prices that are listed, then confirm the live ID and rates in the console.

Sources to check first

Public pages change quickly. If the changelog later adds a formal entry, use that entry’s model ID, price conditions and access rules in place of today’s “unconfirmed.”