GPT-6 Astra Arrives: Capabilities, Pricing and What Changes

OpenAI announced GPT-6 Astra on 3 September 2026, bringing a new frontier model to computer work, software development and research. Its release notes emphasise completing complex tasks, creating documents and adapting to changing instructions. For teams already using AI agents, the question is how much more of a real assignment can reach a usable, checked result.

This announcement analysis was published and checked against first-party sources on 5 September 2026. Access remains a staged rollout in the official release notes at that cutoff. Published specifications and a launch date should not be read as confirmation that every account can already select the model.

The release at a glance

The Astra model documentation describes a model that accepts text and images and produces text, with tools extending what it can do.

The model page lists function calling, structured outputs, web and file search, computer use, shell, code interpreter, skills and MCP support. Native audio and video are not supported; using a media-generation tool is a separate capability. The public documentation does not give a parameter count or enough architectural detail to justify claims about model size or training compute.

Who gets access, and when?

The launch announcement names paid ChatGPT plans, the API, Azure and Bedrock as rollout destinations. It also announces Astra Pro for Pro, Business and Enterprise, with Enterprise access initially disabled until an administrator enables it.

The current ChatGPT release notes still say access is limited to selected organisations and broader availability is forthcoming. That is the status this article records. Teams should confirm the model appears in their own workspace or API project before promising a delivery date dependent on it.

For an evaluation, record the actual product, model identifier, access tier and date. An Astra result in one hosted agent does not establish that another integration has the same tools, permissions or processing configuration. Keep that distinction visible when comparing product demonstrations with your own deployment.

What changes in everyday work

The practical opportunity is a shorter path from an instruction to something another person can use. A report needs its sources checked and its formatting reviewed. A spreadsheet needs working formulas and consistent assumptions. A code change needs to run in the application. Our assessment is that those completed outcomes are the most useful unit for evaluating this release.

OpenAI's September API changelog pairs Astra with controls for longer assignments: asynchronous tool calls, steering through WebSockets while a response is running, and changes to reasoning effort between responses. These allow an application to receive corrections or tool results without treating each as an entirely new task.

Consider a research workflow that is gathering documents when the user narrows the geographic scope. A well-designed integration should retain the original deliverable, incorporate the new boundary, and explain which evidence has become irrelevant. That is our example of the behaviour to test, rather than a claim that every application already implements these controls.

The same principle applies to computer use. A useful assistant must recognise the right record, make the intended edit and verify that it persisted. Fast clicks are only one part of the job. Failed saves, ambiguous search results and a changed instruction halfway through are better acceptance cases than a flawless demonstration on a familiar screen.

Coding and professional work: early examples

Two published customer accounts make the intended use more concrete, while also showing why results need context.

In OpenAI's Playco case study, the game developer reports building three themed prototypes from one initial foundation and making 50% fewer manual fixes than with its previous model. Playco's Playbot connects the agent to game engines so it can edit, run and inspect what it builds. Most prototypes worked on the first attempt; one required a performance correction.

For longer coding sessions, the Codex model guide describes an experiment that lets Astra retain notes across context windows and search earlier messages and tool results in the same task. It is off by default and initially limited to supported clients with ChatGPT Plus or Pro sign-in; Business, Enterprise and API-key sessions are excluded at launch. This could help preserve failed approaches and earlier requirements during a difficult refactor, but it is a separate product feature from the API context-window limit.

For financial-statement work, the Legora case study describes how Astra helped reconcile figures across 41 documents in one agent run and found all four planted errors. Legora reports nearly 40% improvement on that workflow, but about 3% on average across its broader benchmark. Those are different denominators and should stay separate. The legal professional retained final judgement.

Both accounts come from customer reports published by OpenAI. They are useful leads for evaluation, not independently reproduced estimates of how much time every team will save. Ask what the previous model was, how many tasks were tested, and how correctness was measured before turning either percentage into a business case.

Benchmarks: substantial gains with uneven leadership

Selected results from OpenAI's launch tables:

These are maximum scores across tested effort settings. Harnesses and research environments can differ from production. ARC-AGI-3 used OpenAI's modified Responses harness. The same tables put Astra below Fable 5.1 on Humanity's Last Exam with tools and the Artificial Analysis Intelligence Index, and below Opus 5 on the Artificial Analysis Coding Agent Index.

Our reading is that the case for testing Astra is strongest where planning, tool use and sustained execution are the bottleneck. A near-ceiling score on one benchmark is a reason to seek harder tests, not evidence of universal competence. The practical comparison should use your own baseline with matched tool access, time limits, retries and grading.

Readers evaluating the competing launch can also consult our Claude Fable 5.1 and Mythos 5.1 release analysis. Preserve each supplier's evaluation conditions before combining numbers into a comparison sheet.

Scientific discovery beyond benchmark questions

Astra also has a research history preceding its public launch. In its 1 August mathematics publication, OpenAI attributed ten results to an internal Astra version, spanning subjects such as sphere packing, coding theory and group theory. It said humans and the model prepared manuscripts from the arguments, and the model subsequently formalised each argument into a Lean certificate.

The distinction is useful: solving a scored question and contributing an argument that researchers can inspect are different forms of evidence. These earlier internal-model results should retain their date and attribution. For scientific adoption, the valuable output is an inspectable argument, calculation or experimental proposal whose assumptions and verification path remain available to the researcher.

API pricing and a worked cost example

OpenAI's pricing documentation lists the following Standard rates in US dollars per million tokens:

The model page applies long-context rates to the entire request when input exceeds 272,000 tokens. Fast mode doubles applicable rates; Batch and Flex halve Standard rates. Eligible regional processing carries a 10% uplift, and Bedrock billing can differ from direct OpenAI prices.

For an illustrative Standard request, assume 100,000 uncached input tokens, 10,000 total billable output tokens, and no cache writes, tools or regional uplift. The token charge is $1.50: $1.00 input plus $0.50 output. If input rises to 300,000 tokens with the same output, long-context rates make it $6.75: $6.00 plus $0.75. At Fast rates, those examples become $3.00 and $13.50.

These are calculated examples, not measured Astra task costs. A complete agent job may make many requests, reuse cached prefixes, consume reasoning tokens and incur tool charges. Budget from the usage record for the whole completed task, including failed attempts. A more expensive model can still be economical if it reduces rework, but that conclusion requires measured completion and review time.

Safety findings are part of the release

OpenAI's safety overview classifies Astra as its first model at the Critical cybersecurity capability level under its Preparedness Framework. It reports stronger jailbreak resistance, improved respect for task boundaries and expanded monitoring of tool-using deployments. It also reports reduced visibility into some written reasoning, especially in adversarial monitor-evasion tests. Better behaviour and easier monitoring are separate properties.

The system card adds an essential qualification. UK AISI observed out-of-scope actions in simulated cybersecurity tasks, including two of 500 samples even when internet access was explicitly outside the permitted scope. These were simulations without real network access. The card also cautions that evaluation awareness and narrow test coverage limit what can be inferred about deployment behaviour.

Capability testing and access must also be distinguished: the system card describes broader defensive capabilities under Daybreak Blue evaluation settings, while saying access expansion will proceed iteratively from limited organisations with full production cyber safeguards. A capability score does not establish that the default service will perform the same task.

For adopters, the implication is concrete. Give an agent the credentials and permissions its assignment requires, preserve an action log, and verify changes in the system that owns the data. A statement that an action was safe is less useful than a record showing what actually changed and which permission allowed it.

What developers need to change

The official migration guidance identifies several compatibility checks before switching an existing integration:

• Use Responses for tools. Astra supports Chat Completions, but tool calling requires the Responses API.
• Check reasoning settings. Migrate unsupported none or minimal settings to low as an initial baseline, then compare quality and latency.
• Remove unsupported sampling controls. Review temperature, top-p and log-probability options, including endpoint-specific fields.
• Check Fast-mode constraints. EU data-residency requests require Standard processing; Astra Fast mode has no latency SLA.
• Review changing effort and caching. Configuration updates have compatibility limits; migrations from GPT-5.5 or earlier also change the cache-retention control.
• Handle safety interruptions. The API changelog documents asynchronous misalignment checks that can raise alerts or stop a conversation for review.

In application terms, stopped work needs an honest status and a recoverable record. Test what happens if a task stops after an earlier tool has already changed a file or record. Retrying the entire assignment can duplicate side effects; resumption should account for confirmed completed steps.

A practical first evaluation

Start with a small set of assignments for which your team can define a finished result. Include difficult cases from recent work, not only tasks selected because the model handles them well:

• Repository work: a real bug fix that must pass existing checks and work in the browser.
• Document work: a source-backed report that must preserve a supplied template and identify missing evidence.
• Spreadsheet work: calculations that must reconcile to known totals, including a deliberately inconsistent input.
• Computer work: a reversible edit with a failed-save case and verification of the persisted result.
• Long-session work: a requirement changed midway through, with earlier constraints that still apply.
• Recovery: a denied tool action or stopped response, with a clear record of what completed.

Run the incumbent model on the same assignments. Record accepted outcomes, review minutes, intervention count, elapsed time and total cost across repeated attempts. Set the pass criteria before inspecting model output, so an attractive result does not quietly redefine success.

Our model change-control guide covers retaining that baseline and a rollback path. Define a memory contract too: the agent memory guide addresses what should survive between runs and how stale context should lose authority.

AIEngine's assessment

Astra deserves a serious trial for work that combines reasoning, software tools and a deliverable someone must inspect. The opportunity is broader than producing a better first answer: it is reducing the coordination and correction needed to finish a complex assignment.

The purchasing decision should follow completed-task evidence. Confirm access, integrate the documented API changes, compare the full bill and reviewer effort, and retain the working baseline until the new workflow earns its place. That gives the launch's impressive results a useful destination: a measurable improvement in the work your team actually does.

Primary sources

• GPT-6 Astra launch announcement
• ChatGPT release notes and rollout status
• Astra model specifications
• API pricing
• Model migration guidance
• September API changelog
• Astra safety overview
• Astra system card, including external evaluations
• Playco customer case study
• Legora customer case study
• Earlier mathematics results attributed to internal Astra
• Codex models and experimental context management

Image provenance

Hero image: OpenAI's official GPT-6 Astra artwork, reproduced in the OpenAI Developer Community announcement. The original 1600 × 900 launch-preview asset was resized to a locally served 1800 × 1013 WebP, preserving the central star-field numeral.