OpenAI Releases GPT-6 Astra: A Review of the New Flagship Model

What is GPT-6 Astra  

Astra continues OpenAI’s “named” tradition: over the summer, the company released the GPT-5.6 family of three models — Sol, Terra, and Luna — and in July launched the ChatGPT Work agent based on them. GPT-6 Astra is the first model designed specifically for this new shell: OpenAI recommends trying Astra via the ChatGPT desktop app.

The name Astra first appeared not in the release, but in a safety report. On September 2, OpenAI published a piece titled “Path to Astra,” where it said the model was the first to cross the “critical” cybersecurity threshold under the Preparedness Framework — the company’s internal risk assessment system. The full release followed the next day.

Astra is not “just another version of GPT-5.” OpenAI describes it as a next-generation model that combines advanced reasoning, computer use, and “stronger judgment” for complex work in code, apps, and research. The key change is that the model was built for long agentic sessions: it can keep notes across context windows without compressing accumulated details into a single summary, and ask clarifying questions without stopping the work in progress.

Separately, OpenAI emphasizes alignment. In its internal tests, Astra never once went beyond the bounds of the allowed task (0.0% versus 48.2% for Sol in the ExploitGym trap test), did not attempt to bypass the AutoReview code-checking system, and was three times less likely than Sol to misrepresent its own capabilities. The hallucination rate on the internal benchmark fell to 4.2% from 12.2% in its predecessor. 

Capabilities and Benchmarks  

OpenAI compares its new model with GPT-5.6 Sol, the previous generation’s flagship, as well as Anthropic’s Claude Fable 5.1. 

Computer use. On OSWorld 2.0 — 72.6% versus 65.7% for Sol, with the average time per task cut from 75 to 40 minutes. On ScreenSpot-Pro, a test of accuracy in identifying interface elements — 92.7% versus 76.9%. On Agents’ Last Exam — 59.3% versus 53.6% for Sol and 55.5% for Opus 5.

Professional work. On AutomationBench, a business process test, — 41.4% versus 18.1% for Sol, more than double. On BenchCAD, where the model writes code for 3D CAD models, — 95.9% versus 83.3%. On BrowseComp — 91.5%.

Programming. Terminal-Bench 4.0 — 57.9% versus 37.3% for Sol and 55.8% for Fable 5.1. DeepSWE v1.1 — 74.1% versus 72.7%. FrontierCode 1.1 Extended — 64.5% versus 60.6%. On database migration tasks — 63.9% versus 42.7%.

Science and reasoning. Terminal-Bench Science 0.1 — 64.6% versus 52.6% for Fable 5.1. FrontierMath Tier 4 — 97.6% versus 83.0% for Sol, which is close to saturation. GPQA Diamond — 96.0%. Humanity’s Last Exam with tools — 57.2%: here Astra trails Fable 5.1 with its 65.0%.

Cybersecurity. ExploitBench — 100% versus 78.5% for Sol, the benchmark is saturated. On ExploitGym — 42.4% versus 30.3%, and on SRE-Bench for binary reverse engineering — 88.0% on the first try and 99.2% in four.

The most talked-about number is 99.9% on ARC-AGI-3, an interactive test of abstract reasoning, where the previous Sol scored 17.8%. An important caveat: in the standard ARC Prize harness, the model scores 62.7%. ARC Prize Foundation President Greg Kamradt noted that on 96% of levels, Astra beat the human baseline on action efficiency.

Prices, Context, and Effort Levels  

According to the OpenAI documentation, Astra’s context window is 1,050,000 tokens, the maximum output is 128,000 tokens, and the knowledge cutoff date is April 30, 2026. The model accepts text and images, and responds with text. There are five reasoning effort levels: low, medium, high, xhigh, and max; a “no reasoning” mode is not supported.

The base price in API is $10 per million input tokens and $50 per million output tokens — exactly the same as Claude Fable 5.1, which OpenAI compares the model to in its announcement. Recall that Anthropic released its Mythos-class flagship in the summer, and the updated 5.1 version — just two days before Astra. Cache reads cost $1 per million, and cache writes cost $12.5 per million. There are three modifiers: requests longer than 272,000 input tokens are billed at double the input rate and 1.5× the output rate; Fast mode with double speed costs twice as much; Batch and Flex provide a 50% discount.

According to Artificial Analysis estimates cited by Simon Willison here, Astra scores 61 points on the composite index versus 66 for Fable 5.1, but for coding tasks it is roughly twice as cheap with comparable results. 

Where it is Available and on Which Plans  

Astra usage counts toward existing subscription limits, and you can buy additional credits beyond them. According to OpenAI’s help page, Pro users paying $100 and $200 get their full limit under Astra, while Plus and standard Business seats get a limited number of requests with optional credits.

Pro, Business, and Enterprise subscribers also get access to GPT-6 Astra Pro for the most complex tasks. There is no free plan on the list.

Cybersecurity: the First “Critical” Level Model  

Astra is OpenAI’s first model that meets the Critical threshold for cybersecurity under the Preparedness Framework. In practice, this means the model can independently find zero-day vulnerabilities and build exploit chains in hardened systems. During evaluation, it discovered two previously unknown vulnerabilities, escaped the browser sandbox, and gained root access by chaining multiple OS flaws.

OpenAI’s response is layered defense. Regular users will not get access to offensive workflows: exploit development, malware analysis, and detection engineering have been moved into a separate OpenAI Daybreak program with looser restrictions and participant vetting. In tests, Astra rejected 91.5% of attempts to bypass cyber-related safeguards, versus 59% for Sol.

The second layer is monitoring. Classifiers review the model’s reasoning and actions and automatically halt unauthorized activity. OpenAI warns that, because of this, legitimate work in ChatGPT may sometimes be paused pending review, and in the API it may be stopped. The company also acknowledges that Astra’s reasoning has become harder to control than Sol’s, and calls monitorability a research priority.

The release comes amid mounting pressure on the industry. According to Al Jazeera, after the Hugging Face incident, Senator Bernie Sanders and Congressman Greg Casar introduced a bill calling for a pause in advanced AI development until federal safety rules are adopted.

How to Test the Model in Practice: Building a 3D Aircraft Model From a Blueprint in Blender  

We’ll test Astra’s key promise — real work in desktop apps — by modeling in Blender. We asked the neural net to build a fighter jet from a three-view blueprint: side, front, and top.

With Blender, desktop ChatGPT has two options. The first is the Computer Use plugin: the model sees the screen, clicks through menus, and drags vertices like a human. It’s universal, but every step goes through screenshots, and access to each app has to be approved separately. The second is an MCP server: the model controls Blender via its Python API, and objects appear instantly in the open editor window. For blueprint-based geometry, the second approach is noticeably more accurate — so that’s what we’ll use. We covered a similar setup with Claude back in the spring, but there the connector could be installed from a catalog in a few clicks. ChatGPT doesn’t have a ready-made Blender connector, so you can delegate the entire setup to Astra itself.

You’ll need the new ChatGPT desktop app for macOS or Windows, plus a paid plan with Astra access. You don’t have to install Blender in advance: the model will do it itself.

Step 1: Open the App and Select the Model

In the model selector, choose GPT-6 Astra. The app will ask for permission to execute commands. There will be several such prompts.

Step 2: Ask Astra to Set Up the Stack

The first prompt is about infrastructure. The model should check Blender, install anything missing, install the official MCP Server extension from the Blender Foundation, add the server to the app config, and launch the editor.

Prompt

Set up a connection between you and Blender on this computer, then launch Blender. 

Then Astra takes over: it checks the Blender version, installs uv, downloads the extension archive and installs it via a bpy script, and updates the config. On Windows, the Blender installer may trigger a system permission prompt — you will need to approve it manually, and the model will warn you about this on its own. In the end, the Blender window opens on screen with an empty scene.

Step 3: Build the Model 

The second prompt is the actual modeling. Attach the blueprint to your message: Astra accepts images as input and will “read” the proportions itself, and based on the file name, it will load the same image into Blender as a reference. The model will lay out the workflow and checkpoints on its own. 

Prompt

Build a 3D model of the fighter from the attached three-view drawing (side, front and top views). 

Astra calls the MCP server tools: it models and textures the object. All the user has to do is check the result.

As a bonus, the neural net rendered the finished plane from a nice angle and dropped the image into the chat.

What the Astra Launch Changes for Users

GPT-6 Astra is OpenAI’s first flagship where the main stage is not chat, but the desktop: Blender, Excel, and game engines are shown in the announcement not as experiments, but as standard use cases. Given that the ChatGPT desktop app has already merged chat and Work, OpenAI is building a unified environment where the model works equally well with documents, code, and desktop software.

The price of the release is the strictest security posture in the company’s history. The “critical” level of cyber capabilities means that regular users will not see some of Astra’s features, and monitoring will sometimes intervene in its operation. Whether businesses will accept this trade-off will become clear in the coming weeks — as will how OpenAI handles regulators’ complaints, which have noticeably increased since the Hugging Face incident.