DualView

MODEL PROFILE · OpenAI

GPT-6 Astra

A hosted model to evaluate for reasoning and code workflows, with text output and documented image-input support.

Source reviewed: · Language model

Provider documentation summary. No hands-on results or independent ranking are claimed.

Documented facts

The following fields come from the official provider source, reviewed on the date above.

API identifier
gpt-6-astra
Inputs / outputs
Text and image inputs; text output. Native audio and video are not listed as supported.
Context / maximum output
1,050,000 / 128,000 tokens
Standard token prices
USD $10 input and $50 output per million tokens. Cached input is $1 per million.
Long-context caveat
Above 272,000 input tokens, the documentation applies 2× input/cache and 1.5× output rates to the full request.
Access
Hosted API. Downloadable weights are not documented on the reviewed model page.

October 1 context: GPT-6.1 Astra release reporting is separate

Reported release change, not a specification update: SecurityWeek reported on September 29 that the planned GPT-6.1 Astra launch was called off, citing the Wall Street Journal and comments attributed to OpenAI safety-systems head Saachi Jain. Its account concerns scope, authorization and how the model reports its work. The Journal original was not fully accessible in this review. This does not establish permanent cancellation, a replacement date or withdrawal of GPT-6 Astra. The facts above describe gpt-6-astra only; do not copy them to the reported successor. No GPT-6.1 Astra capability or price is verified here. SecurityWeek: attributed GPT-6.1 Astra release report.

October 2 update: OpenAI response on future Astra models

Carly Page’s September 29 report for The Register says OpenAI confirmed the decision to shelve the planned October GPT-6.1 Astra release. The same report attributes a separate statement directly to the company: more Astra models are planned, alongside other forthcoming models that have cleared its safety requirements. This adds a company response to the earlier release reporting; it is not a dated launch commitment. The report does not name the future Astra version, establish public access or announce a replacement API identifier. Do not interpret its general timing language for other models as a deadline for Astra. For an evaluation or procurement decision, keep the existing gpt-6-astra contract separate from an unnamed future model; this report supplies no successor pricing, context limit or measured performance. The specification review date above remains September 27. The Register: OpenAI response reported September 29.

October 4 rumor watch: Bel is not a verified Astra specification

Unverified rumor, single-origin claim: Rohail Saleem’s August 25 Wccftech report attributes to the pseudonymous X account @synthwavedd a claim that OpenAI had completed a pretraining run called Bel with more than 10 trillion total parameters. The embedded post calls it a successor to Doug; the article associates the rumor with future Astra/GPT-6 models. We could read the report and embedded text, but the linked X original returned 403. No underlying training evidence was verified. This establishes the existence of the public claim, not its accuracy or the poster’s access. Wccftech: August 25 report and embedded origin post.

Why the reported codename should stay separate

Nitish Garg’s October 4 CellCog article revisits the same account’s claim and describes Bel as a possible post-GPT-6 base. That framing differs from the older Wccftech interpretation. CellCog also reports a September 23 statement that public access remained months away; this is another claim from the same origin, not independent corroboration or a release commitment. We read CellCog’s indexed article text; direct retrieval failed. The record does not establish Bel’s relationship to the released GPT-6 Astra model, its parameter count, a successor product name, or a usable endpoint. CellCog: October 4 rumor review (indexed text).

What the rumor would change if confirmed

The potential impact is high because a new base model could change which future language-model releases users should evaluate. Confidence is low because the central claim remains unverified and the product mapping is disputed in public reporting. Keep that possible impact separate from any prediction of a launch. For a current Astra evaluation, do not substitute the rumored size for documented limits or infer stronger answers from it. The next useful evidence would identify the model and connect it explicitly to an available product, with access terms and comparable evaluations. No such connection is established by these reports. The specification review date above is preserved. Attribution and uncertainty record.

The practical decision

Image input is not the same capability as native image generation. For a visual workflow, assess whether you need a model to explain an image or an endpoint that creates pixels. Tool-generated images also need their own model and cost recorded. For text and code evaluation, record reasoning settings alongside the model name so that a rerun can be interpreted.

What to record in your own comparison

Limitations and unknowns

This profile does not establish latency, accuracy or superiority over another model. Tool fees, fast modes and volume discounts are outside the simple standard-price fields. Confirm billing against your selected API mode before estimating a workload.

Compare your saved results

Use the DualView editor to inspect files you already have. The accepted-output calculator uses your own rates and counts. Neither link runs a model or purchases generation.

Related reporting

Related model profiles

All model profiles and dated changes · Subscribe via RSS