GPT-6 Astra and the GPT-6 Family: Pricing, Access, and Tests

👁 122 reads

GPT-6 Astra is OpenAI’s model for reasoning, coding, and tool use, introduced on September 3, 2026. You can use it across OpenAI products and via the API, though availability and billing vary depending on the chosen service and account tier.

Related reading: GPT-6 Astra vs Opus 5.5: 9 tests, results, and costs

Latest update: OpenAI pauses rollout of GPT-6.1 Astra - model exceeded permissions in evaluations

What is this model officially called?

The formal name is GPT-6 Astra. Search queries frequently include variations such as ChatGPT Astra, Astra GPT, chat gpt astra 6, chat gpt 6 astra, chatgpt 6 astra, gpt6 astra, OpenAI Astra, plain GPT-6, or simply “new ChatGPT”. Typographical errors like “gtp 6 astra” also circulate. All refer to the same underlying system. Spacing differences such as “gpt6 astra” without a space or hyphen refer to the same model.

Two key distinctions matter in practice: ChatGPT is the consumer application, while GPT-6 Astra is the model operating behind certain interfaces within it. Second, Astra running in Work or Codex is not identical to GPT-6 Pro in standard chat. Seeing the model name in a press release does not guarantee immediate access inside your specific workspace or account.

Regarding deployment schedules, see our tracker on when GPT-6 Astra becomes available to all users.

In our benchmark inside Codex across 24 runs on six custom logic puzzles, the High and Ultra configurations achieved identical score totals. They differed in execution latency and measured token usage. When selecting an operational setting, start with the specific task requirements rather than defaulting to the highest menu option.

GPT-6 Astra API Pricing and Long Context Thresholds

According to the OpenAI pricing documentation, standard API access charges $10 per million input tokens, $50 per million output tokens, $1 per million cached read tokens, and $12.50 per million cache write tokens. Once input size exceeds 272,000 tokens, those rates jump to $20, $75, $2, and $25 respectively. That higher rate applies to the entirety of the request, not solely the volume beyond the cutoff threshold.

Here is an illustrative calculation under standard rates without caching, tool overhead, or regional fees: an input payload of 270,000 tokens with 10,000 output tokens costs $3.20. Increasing input volume to 280,000 tokens with the same 10,000 output tokens brings the bill to $6.35. That additional 10,000 input tokens adds $3.15 to the run cost. This reflects the published price structure, not an empirical inference benchmark.

The Fast processing tier doubles these base rates. Batch and Flex access models run at half standard rates. These billing modes should be evaluated separately. In-app ChatGPT subscriptions operate under separate quotas detailed in our ChatGPT pricing breakdown. Usage costs inside Codex do not translate 1:1 to direct API request expenses.

Checking GPT-6 Astra Access on Your Account

In its launch announcement, OpenAI outlined a phased release across ChatGPT Plus, Pro, Business, and Enterprise plans, as well as via the API, Microsoft Azure, and AWS Bedrock. Enterprise workspace administrators may need to manually enable the model. The public release notice does not mean immediate activation on every account.

Our published evaluation confirms availability through Codex in the testing environment used. It does not verify every regional broker, subscription tier, or deployment zone. If you search for “chat gpt 6” wanting an immediate prompt session, open the official ChatGPT interface and inspect the model selector directly. Third-party applications using “6” in their marketing do not necessarily route to Astra.

Large Context Limits vs Document Grounding

The model system card lists a context window of 1,050,000 tokens, a maximum output capacity of 128,000 tokens, and a training knowledge cutoff date of April 30, 2026. Astra accepts text and image inputs and produces text. Integrated image rendering utilities remain distinct from native generative output.

A massive context window accommodates large document bundles, yet responses still require factual verification against the provided text versions. When processing extensive corpora, prompt the system to return direct quotations, source file references, and section locators for every deduction. If source support is absent, instruct the model to report that directly. Guidance on structuring these directives is detailed in our guide on how to craft effective AI prompts.

OpenAI Official Benchmark Comparisons

The following figures reflect benchmarks published by OpenAI, not our independent lab measurements. Metric designations and model scores match the original release records. Empty fields indicate omitted results in the vendor baseline data. Because this table reflects Astra’s launch on September 3, 2026, later systems like Claude Opus 5.5 and Gemini 4 Argon are not listed here. Their comparative scores appear in our overview of Gemini 4 Argon vs GPT-6 Astra vs Opus 5.5.

Astra does not lead across every evaluation category. Your own testing on copy editing, inference latency, or correction overhead may lead to a different model preference than synthetic mathematical benchmarks suggest. You can explore ongoing evaluations in our AI benchmarks archive.

High vs Ultra Setting Evaluation: A Tie in Accuracy

In our GPT-6 Astra High vs Ultra benchmark, we ran six challenging puzzle tasks twice across each configuration mode. Both High and Ultra scored 12 out of 12 total points. Median completion latency reached 26.52 seconds on High versus 47.97 seconds on Ultra. Ultra required an 80.92% longer median turnaround time without yielding an accuracy improvement on this test set.

Recorded input and output token consumption totaled 285,181 for High and at least 1,007,209 for Ultra. This represents a lower-bound figure for Ultra, as internal telemetry captures were incomplete. Ultra spun up auxiliary sub-agents, whereas High ran without delegating sub-agents. Consequently, this difference does not represent pure reasoning compute or a direct cost multiplier.

You can review the raw run telemetry in the JSON test results file alongside the test bundle and protocol published with the original benchmark. This small tool-augmented puzzle evaluation does not predict which setting performs better during hours-long codebase refactoring or complex software debugging.

Limitations and Production Caveats

Flawless completion across a controlled, six-task logic sample does not imply universal infallibility. This test did not evaluate medical diagnostic safety, jurisdictional legal accuracy, or broad codebase architectures. Correct final answers can still stem from redundant steps or excessive execution cycles. In production setups, evaluate both the final diff and the reasoning chain used to justify changes.

OpenAI designated Astra as “Critical” tier in cybersecurity risk, implementing the protections described in its GPT-6 Astra safety overview. Standard consumer accounts do not receive unconstrained access to research-grade capabilities. A session paused by automated safety controls does not indicate platform downtime.

How to Choose the Right Reasoning Setting

Begin with a workflow whose accuracy you can verify objectively: isolating a reproducible software bug, extracting metrics from a target financial report, or synthesizing cited references. Document success criteria before execution. Supply identical prompts and source assets across tested configurations. Track latency, total score, required follow-up iterations, and token utilization.

Opting for a higher reasoning setting makes sense when the resulting performance justifies the additional latency and resource burn. If both configurations satisfy project specifications, deploy the more cost-effective option within your infrastructure budget. If both fail, restructure source data chunks or decompose the problem. Pushing a reasoning slider to maximum cannot replace strict validation.

How Astra Compares Across Our Broader Testing

Vendor metrics offer only an initial picture. In an independent evaluation covering nine challenges across coding, translation, and instruction following, GPT-6 Astra was the only model out of eleven to achieve a perfect score, finishing one point ahead of the most expensive competing model tested.

In a separate language generation evaluation conducted in Polish, four models achieved maximum scores, with usage costs varying by up to a factor of thirty-two.

Latest comparison: Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5 - a breakdown across 19 benchmarks and the cost of three production tasks (September 30, 2026).

If you use more than one model, my service Jantaris puts GPT-6 Sol, Claude Opus 5.5, Gemini 3.6 Flash, DeepSeek V4, Veo 3.1 and Nano Banana Pro on one prepaid USD balance with one API key - no subscription, the balance never expires and failed generations cost $0. Everything in one place: the Jantaris guide.

Chat Access and GPT-6.1 Terra

OpenAI’s release notes say GPT-6 Sol is used in the Chat tab on Plus, Pro, Business, and Enterprise, while Free and Go use GPT-6 Luna. The Pro reasoning option continues to use Astra and does not support Intelligent UI. Enterprise access depends on workspace settings.

OpenAI is also rolling out GPT-6.1 Terra in Work and Codex across Plus, Pro, Business, Enterprise, and Edu. Its availability on an account depends on the plan, region, and workspace settings. It is not included in the historical four-model writing table above. For a separate overview, see the GPT-6.1 Terra guide.

Does Free ChatGPT use Astra?

No. OpenAI says Free and Go use GPT-6 Luna in chat.

Where can I select GPT-6.1 Terra?

When it reaches an eligible account, select it from the model list in Work or Codex.

Sources for this update

• https://help.openai.com/en/articles/6825453-chatgpt-release-notes

// read moreGPT-6 Astra Pro: Same Price & Context as Astra. What’s Different?NewsGPT-6 Cyber: Specs, Benchmarks, Pricing and AccessNewsWho should use GPT-6 Astra vs cheaper alternativesAI Tools

Kinetyra