The iteration speed of AI models from various companies recently has been anything but ordinary. Following its launch at the end of June, OpenAI GPT-5.6 Sol Shortly afterward, the new-generation “GPT-6 Astra” model was officially unveiled ahead of schedule! Although less than three months had passed, GPT-6 Astra represents a comprehensive upgrade, with stronger computer operation, browsing, and coding capabilities. It even scored 99.9% on the ARC-AGI-3 test. OpenAI President Greg Brockman also stated that he personally believes OpenAI has achieved AGI, and the media briefing even concluded with the phrase, “Welcome to the AGI era.”
Moreover, its cybersecurity capability has also crossed into the Critical threshold of the OpenAI Preparedness Framework for the first time, meaning that with the right tools and permissions, GPT-6 Astra can independently identify unknown vulnerabilities, develop exploit methods, and even execute complex end-to-end cyberattacks without step-by-step human guidance. This also makes it the first OpenAI model to be rated at this level.
GPT-6 Astra officially debuts: OSWorld task time reduced by 47%, ARC-AGI-3 reaches 99.9%, and security capabilities cross into Critical for the first time.
GPT-6 Astra It is currently OpenAI’s highest-tier general model, with core capabilities covering computer operation, web browsing, software development, professional work, scientific research, and cybersecurity. Unlike previous models, Astra is being developed to “take over complete workflows”—it can fill out online forms, update customer data in a CRM, manage calendars, and research information in a browser before drafting summaries in an email or document editor.
Simply put, once GPT-6 Astra opens the tool, it can handle everything from start to finish.
According to Axios, Brockman described it as a “generational leap” in capability during the GPT-6 Astra media briefing, and said he personally believes OpenAI has achieved AGI. When a reporter asked whether, looking back in the future, Astra might be seen as the turning point marking the arrival of AGI, he replied, “I think it might be about this model,” and closed the briefing with “Welcome to the AGI era.”
In the ARC-AGI-3 test, GPT-6 Astra achieved 99.9%, GPT-5.6 Sol scored only 7.8%, and Claude Opus 5 scored 30.2%:
GPT-6 Astra is now also better at knowing when to make decisions on its own and when to stop and ask you.
If a blank left in the instructions doesn’t affect the main outcome, it fills in reasonable details based on context. It only asks more focused questions when the answer could change the direction of the task. More importantly, in Codex, while GPT-6 Astra waits for the user’s response, it can continue working on parts unaffected by the question, so the entire task doesn’t stall.
OpenAI also acknowledged that when earlier models received supplementary instructions mid-task, they sometimes treated the new message as a separate objective and ended up forgetting what they were originally supposed to do. GPT-6 Astra has been trained specifically for this kind of “interrupted halfway through” scenario, allowing it to incorporate new requests into the existing workflow and answer side questions before returning to the main task.
When using Codex for extended periods, GPT-6 Astra also adds a new cross-content window memory approach.
Once past conversations fill up the context window, the system has to compress the earlier process into a summary, and each time it compresses, it may lose track of why a particular fix failed or what results a certain test produced.
The newly added mechanism allows GPT-6 Astra to save notes and search earlier conversations and tool outputs, so even if a detail wasn’t written into a summary, there’s still a chance to retrieve it. However, this experimental feature initially needs to be manually enabled in Codex’s `config.toml`, and OpenAI plans to gradually roll it out as a default Astra feature over the coming weeks.
The Most Noticeable Improvement: Computer Operation
Computer operation is where GPT-6 Astra delivers the most noticeable improvement for everyday users. According to OpenAI’s results on the OSWorld 2.0 benchmark, Astra scored 72.6%, higher than GPT-5.6 Sol’s 65.7% and Claude Opus 5’s 70.2%. OpenAI also simulated task execution times, finding that Astra takes about 40 minutes on average to complete a task, while Sol takes roughly 75 minutes—a reduction of about 47% in time.
ScreenSpot-Pro, which tests a model’s ability to recognize screens and identify the correct operation positions, Astra reached 92.7%, while GPT-5.6 Sol scored 76.9%; on Agents’ Last Exam, which evaluates workflows across different occupations, the two scored 59.3% and 53.6% respectively. Combined with OpenAI’s concurrently updated Codex operation framework, Astra’s completion speed on the Mind2Web test reached 1.9 times that of the current Sol experience:
GPT-6 Astra’s programming capabilities have also improved significantly, though it doesn’t take first place in every category.
In the Terminal-Bench 4.0 test, Astra scored 57.9%, the previous generation GPT-5.6 Sol was 37.3%, Claude Fable 5.1 was 55.8%, and Claude Opus 5 was 52.3%. For DeepSWE 1.1, Astra’s 74.1% was also slightly higher than Gemini 3.8 Flash’s 73.8% and Claude Opus 5’s 73.7%.
However, on FrontierCode 1.1 Main, Astra’s 53.3% is slightly lower than Claude Fable 5’s 53.5%; its Artificial Analysis Coding Agent Index is also only 67.0, below Claude Fable 5’s 67.2 and Claude Opus 5’s 68.1. Therefore, OpenAI’s claim of being the “current best software engineering model” is better viewed as an official statement of overall capability rather than a lead across every benchmark.
Other test results and cybersecurity
In the results for mathematics, science, and abstract reasoning, Astra scored 97.6% on FrontierMath Tier 4, GPT-5.6 Sol scored 83.0%, and Claude Fable 5.1 scored 87.8%; on Terminal-Bench Science 0.1, the score rose from Sol’s 22.4% to 64.6%.
OpenAI also stated that Astra helped researchers reduce the known upper bound on the distance between infinitely many pairs of primes from the previous 240 down to 186, and improved a result on large prime gaps that had remained unchanged for more than 80 years.
In terms of cybersecurity risk, Astra scored 100% on ExploitBench, while GPT-5.6 Sol scored 78.5%. On ExploitGym, the success rate was 42.4%, compared to Sol’s 30.3%.
In the SRE-Bench test, which evaluates models’ ability to reverse-engineer software executables without access to source code, Astra achieved a single-attempt solve rate of 88.0%, improving to 99.2% within four attempts, while Sol scored only 55.9% and 68.7%, respectively.
OpenAI stated that GPT-6 Astra even discovered and exploited two previously unknown zero-day vulnerabilities during testing, and it is currently reporting these to the relevant maintainers. It also became OpenAI’s first widely deployed model whose security capabilities reached the “Critical” threshold under the Preparedness Framework.
In other words, it not only analyzes known vulnerabilities—given the right tools and permissions, it may also uncover undisclosed issues on its own and develop new methods of exploitation.
To reduce risk, the standard version of Astra will refuse to create more advanced proof-of-concept exploits.
OpenAI plans to gradually provide less restricted access through the Daybreak program in the coming weeks, for use in vetted defensive work, including verifying vulnerabilities, analyzing malware, and creating detection rules. This also means that even if security researchers obtain Astra, the tasks they can actually perform will still be limited by program terms, eligibility, and security mechanisms.
Although OpenAI also lists “not acting beyond boundaries without authorization” as another improvement of GPT-6 Astra, this does not mean there are no security concerns at all.
OpenAI acknowledged in a safety report that Astra’s reasoning process is harder to monitor than GPT-5.6 Sol. When researchers deliberately asked the model to evade monitoring, it was better able to control the reasoning content it wrote, and it may also avoid internal monitoring when strategically lowering performance or partially sabotaging tasks.
GPT-6 Astra Deployment and Usage
GPT-6 Astra is currently rolling out in phases. The first wave is available to a small number of organizations, and over the next few days it will gradually open up to ChatGPT Plus, Pro, Business, and Enterprise users. OpenAI’s product lead Tibo Sottiaux also explained on X that the team hopes to bring Astra to all Plus users, not just Pro, Business, and Enterprise, but the full rollout will take a few days.
To compensate paying users waiting for Astra to become available, Tibo subsequently announced compensation: starting from the day of the announcement, for each day Astra is unavailable to paid ChatGPT plans, users will receive a storable reset credit (banked reset), with the first one expected to be issued about three hours after the post:
Pro, Business, and Enterprise plans will also get GPT-6 Astra Pro.
However, note that Enterprise is disabled by default at launch and must be manually enabled by an administrator. GPT-6 Astra also counts toward the normal usage quota included in each plan, and users can use 100% of their quota on Astra. Of course, once it runs out, additional credits can be purchased.
Developers can also use `gpt-6-astra` via the OpenAI API. The official documentation lists a context window of 1,050,000 tokens, a maximum single output of 128,000 tokens, and a knowledge cutoff of April 30, 2026. The model supports text and image inputs, but does not directly support audio or video inputs; reasoning intensity options are low, medium, high, xhigh, and max, with no none.
| GPT-6 Astra Standard API | Price per million tokens |
|---|---|
| Please provide the Traditional Chinese text you would like me to translate. | 10 dollars |
| Cache input | 1 US dollar |
| Cache write | 12.5 US dollars |
| Output | 50 US dollars |
If a single prompt exceeds 272K tokens, the input and cache rates for the entire request become 2x, and the output rate becomes 1.5x.
Batch and Flex are 50% of the standard price; Fast mode can reach up to twice the speed of Standard, and the rate is also double.
Official shares complete test data table.
| Category | Test item | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash |
|---|---|---|---|---|---|---|---|
| Computer operation | Agents’ Last Exam | 59.3% | 53.6% | — | 48.7% | 55.5% | — |
| OSWorld 2.0 | 72.6% | 65.7% | — | — | 70.2% | — | |
| ScreenSpot-Pro | 92.7% | 76.9% | — | 87.3% | — | — | |
| Professional work | AutomationBench | 41.4% | 18.1% | 31.4% | 17.4% | 26.9% | — |
| BenchCAD | 95.9% | 83.3% | 84.3% | 67.5% | 82.1% | — | |
| BrowseComp | 91.5% | 90.4% | — | 87.4% | 90.8% | — | |
| OpenScore String Quartets（1 – OMR-NED） | 0.84 | 0.19 | — | — | — | — | |
| Internal Design Tasks | 50.0% | 47.4% | — | 35.8% | — | — | |
| Internal Data Science Tasks | 40.9% | 30.5% | — | 34.7% | — | — | |
| Artificial Analysis Intelligence Index v4.1.1 | 61.2 | 60.9 | 65.7 | 62.1 | 63.1 | 58.7 | |
| Programming ability | Terminal-Bench 4.0 | 57.9% | 37.3% | 55.8% | 42.0% | 52.3% | 19.1% |
| DeepSWE v1.1 | 74.1% | 72.7% | 67.4% | 69.9% | 73.7% | 73.8% | |
| FrontierCode 1.1 Extended | 64.5% | 60.6% | 63.6% | 64.9% | 63.6% | 56.3% | |
| FrontierCode 1.1 Main | 53.3% | 47.5% | 50.9% | 53.5% | 53.4% | 43.6% | |
| Internal Database Migration Tasks | 63.9% | 42.7% | 57.8% | 50.3% | — | — | |
| Artificial Analysis Coding Agent Index v1.4 | 67.0 | 65.1 | — | 67.2 | 68.1 | 61.2 | |
| Academic ability | Terminal-Bench Science 0.1 | 64.6% | 22.4% | 52.6% | 21.4% | 30.0% | — |
| FrontierMath Tier 4（v2） | 97.6% | 83.0% | 87.8% | 87.8% | 73.2% | — | |
| GPQA Diamond | 96.0% | 94.6% | 93.7% | 92.6% | 93.7% | 95.3% | |
| Humanity’s Last Exam（w/ tools） | 57.2% | — | 65.0% | 63.8% | 63.6% | — | |
| Science and Health | GeneBench Pro | 37.8% | 28.7% | — | — | — | — |
| MedChemBench（Internal） | 49.3% | 47.4% | — | — | — | — | |
| LifeSciBench | 60.3% | 59.9% | — | — | — | — | |
| HealthBench Professional（length-adjusted） | 63.4% | 60.5% | 58.1% | 60.9% | 56.4% | 52.1% | |
| Cybersecurity capability | ExploitBench | 100.0% | 78.5% | — | — | 70.0% | — |
| ExploitGym | 42.4% | 30.3% | 30.4% | 28.4% | 22.0% | — | |
| ExploitBench (June–August 2026) | 39.0% | 11.5% | — | — | — | — | |
| SRE-Bench | 88.0% | 55.9% | — | — | 12.5% | — | |
| SEC-Bench Pro | 85.4% | 79.1% | — | — | — | — | |
| Alignment and Safety | Internal computer use safety benchmark (lower is better) | 2.4% | 22.0% | 9.5% | 18.3% | 11.5% | — |
| Internal computer use safety benchmark + AutoReview (the lower the better) | 1.8% | 4.3% | — | — | — | — | |
| Internal circumvention benchmark (lower is better) | 0.00% | 0.29% | — | — | — | — | |
| ExploitGym honeypot (lower is better) | 0.0% | 48.2% | — | — | — | — | |
| Impossible ExploitGym | 100.0% | — | — | — | — | — | |
| Internal hallucination benchmark (lower is better) | 4.2% | 12.2% | — | — | — | — | |
| Long content | OpenAI MRCR v2 8-needle 256K–512K | 100.0% | 91.5% | — | — | — | — |
| OpenAI MRCR v2 8-needle 512K–1M | 96.3% | 73.8% | — | — | — | — | |
| Abstract reasoning | ARC-AGI-3 | 99.9% | 7.8% | — | — | 30.2% | — |
| ARC-AGI-2 | 95.0% | 92.5% | 90.0% | 89.2% | 90.4% | — | |
| ARC-AGI-1 | 98.5% | 97.5% | 97.5% | 98.5% | 97.5% | — |
Source: KOCPC Chinese