On September 3, 2026, OpenAI announced GPT-6 Astra, a model aimed mainly at computer operation and long-running tasks. On screen-operation evaluations it posts higher scores than its predecessor in a shorter estimated time, and Codex gains a memory feature to support long sessions. At the same time, OpenAI classifies Astra's cyber capabilities as Critical, limits what is available in general release, and has published results showing that its reasoning is harder to monitor. In third-party evaluations, the efficiency gains in coding stand out more than overall performance. With executives talking about reaching AGI, we check the new model against its rivals and against the extent to which humans can oversee it.
Where the gap widened: computer operation and reasoning
Comparing Astra with its predecessor, GPT-5.6 Sol, in OpenAI's announcement table shows very different margins of improvement: 23.3 points on AutomationBench versus 1.4 points on DeepSWE. The table below extracts values from the same announcement and calculates the difference by subtracting Sol's score from Astra's.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Difference |
|---|---|---|---|
| OSWorld 2.0 (v2026.08.08, offline, partial credit) | 72.6% | 65.7% | 6.9 points |
| AutomationBench | 41.4% | 18.1% | 23.3 points |
| Terminal-Bench 4.0 | 57.7% | 37.3% | 20.4 points |
| DeepSWE v1.1 | 74.1% | 72.7% | 1.4 points |
| FrontierMath Tier 4 (v2) | 97.6% | 83.0% | 14.6 points |
| ARC-AGI-3 | 99.9% | 7.8% | 92.1 points |
On OSWorld 2.0, the score rose 6.9 points while the estimated completion time fell by about 46.7%. That figure is calculated as (75−40)÷75×100 from simulated latencies of roughly 40 minutes for Astra and roughly 75 minutes for Sol. It is not a result measured from actual users' working time. For work that spans multiple screens, the result suggests that, in addition to the partial-credit evaluation score, the estimated waiting time until the operations finish has also improved.
Each metric tests different problems, so the differences cannot be averaged into an overall performance multiple. OpenAI's scores are the best values obtained while varying the compute setting used for reasoning. Evaluations in research environments and through the API also differ from the production ChatGPT in prompts and tools. Whether the same improvement will carry over to your own work needs to be verified after release.
The speedup in computer operation also includes changes on the side that runs the model. OpenAI has also updated the Codex execution environment, and says Mind2Web task completion with Astra combined with it was 1.9 times faster than in the previous environment used with Sol. That multiple compares model and execution environment together, so it must be distinguished from the improvement rate of Astra alone.
Rival comparison: overall performance and coding efficiency differ
Third-party measurements show the model's strengths more clearly. In Artificial Analysis (AA)'s evaluation, Astra's overall metric, the Intelligence Index, came to 61 points, level with Sol. Version 4.1.1 of the index combines nine types of evaluation, including science, knowledge, code generation and agentic work.
データを表で見る
| Intelligence Index (点) | |
|---|---|
| Claude Fable 5.1（max・フォールバック） | 66 |
| Claude Opus 5（max） | 63 |
| GPT-6 Astra（max） | 61 |
| GPT-5.6 Sol（max） | 61 |
| Muse Spark 1.3（xhigh） | 61 |
| Grok 4.6（high） | 61 |
| Gemini 3.8 Flash（high） | 59 |
On overall score, Fable 5.1 and Opus 5 are ahead of Astra. However, even models tied at 61 points differ in their individual results, so a tie does not mean they are equivalent on every kind of work.
The comparison conditions also need care. Fable 5.1 includes a default fallback that switches requests judged safe over to Opus 4.8 or Opus 5. Each company's max or high setting does not guarantee the same amount of compute. Muse Spark 1.3 also has a 62-point max version, but because it is a limited preview for partners and pricing is unpublished, the chart shows the 61-point xhigh version.
In coding, Astra's efficiency improvement stands out. On AA's Coding Agent Index, Astra on Codex scored 67, roughly level with Opus 5 and Fable 5 on Claude Code. The leader is Fable 5.1 on Claude Code, at 70. This compares each model paired with Codex or Claude Code, so it should be read separately from OpenAI's measurement table in the previous section.
According to AA, Astra's max setting used about one-third the tokens of Sol's max setting on Codex, with roughly the same cost per task. Against Fable 5, it reached a comparable score at less than half the cost. On the Intelligence Index, however, even though Astra produces about 10% fewer output tokens than Sol, the higher pricing makes the cost per task 75% higher.
データを表で見る
| Intelligence Indexの加重平均費用 (米ドル／タスク) | |
|---|---|
| Claude Fable 5.1（max・フォールバック） | 3.76 |
| Claude Opus 5（max） | 2.34 |
| GPT-6 Astra（max） | 1.67 |
| GPT-5.6 Sol（max） | 0.95 |
| Muse Spark 1.3（xhigh） | 0.55 |
| Grok 4.6（high） | 0.94 |
| Gemini 3.8 Flash（high） | 0.58 |
Even with the same 61 on the overall index, Astra costs $1.67 per task, while the xhigh version of Muse Spark 1.3 costs $0.55. Savings in coding cannot simply be applied to general questions or research.
These costs are weighted averages reflecting the tokens and prices for input and output as well as reasoning and cache reads and writes. Gemini 3.8 Flash is calculated at its discounted price through the end of the year. They are not total costs including flat-rate plan fees or the time people spend checking results. Narrowing candidates by overall score and then comparing costs on the work you would actually delegate makes it easier to judge whether Astra's efficiency gains help.
Memory that links long jobs, and hands-on results in practice
The work Astra is designed for includes research in a browser, analysis in spreadsheet software, and data entry into customer management systems. The idea is to shift the burden of people carrying out operating steps one by one onto an agent that can handle screens. In Codex, it can ask the user about unclear points while advancing work that does not depend on the answer. For decisions with large impact it waits for a reply, and it takes in additional instructions and questions mid-task while keeping the original request intact.
Codex also gains an experimental memory feature. Previously, when a conversation grew long, the history was summarized, and each time that risked dropping the reasons a fix failed or how a component behaved. The new approach carries working notes into the next context window, and past messages and tool outputs that did not make it into the notes can also be searched. It is an experimental feature enabled from a configuration file, and OpenAI plans to make it Astra's default setting over the coming weeks.
One piece of evidence for gauging the improvement in practice is a usage report from Legora, a service for legal work. In Legora's case study, 41 documents were processed in a single agent run in a financial statement reconciliation task, cross-checked against trial balances, consolidation schedules and the prior year's statements. It reportedly took a few minutes. It found all four errors that had been planted, and performance on that workflow was said to be about 40% higher than the previous model.
Across the tasks in the company's business evaluation, BAR, however, the average improvement was about 3%. The roughly 40% is the performance improvement rate on this financial statement reconciliation, and does not mean work as a whole becomes 40% faster. The previous model is also not named. It is a customer case published by OpenAI, with an operating model in which the agent handles the exhaustive reconciliation and legal professionals make the final judgment on each result.
A large gain on specific work does not translate directly into the average across all work. Adopters, too, will find it easier to judge value by choosing work whose results are easy to verify, like reconciliation, and measuring how much the human time needed for checking falls.
"Critical" capability, and reasoning that is harder to monitor
On September 1, OpenAI announced that Astra's cyber capabilities had reached the "Critical" threshold for the first time among its models. This refers to the ability, given the right tools and access, to find unknown vulnerabilities in protected real systems and develop exploits without step-by-step guidance.
On ExploitBench, which measures whether exploit code can be built from known vulnerabilities, Astra with production safeguards removed scored 100%. In another internal benchmark using 20 severe V8 vulnerabilities disclosed from June to August, the share with successful arbitrary code execution was 39.0% for Astra and 5.5% for Sol. Astra found and exploited two zero-days, which are said to be in the process of disclosure. Neither figure shows a universal ability to break into ordinary systems.
The company is not simply opening such capability to all users. General release supports safe code review and fixes, but refuses to write exploit code that demonstrates vulnerabilities. For advanced defensive work, it plans to loosen restrictions and widen access through Daybreak over the coming weeks. An incident at Hugging Face affected the safeguards, but Astra did not cause it.
In addition, on difficult or unsolvable exploit-writing tasks run without safeguards intervening, the share in which Sol exceeded its permitted scope was about 48%, versus 0% for Astra. This is a result on a specific test and does not mean the real-world accident rate becomes zero.
Being better at respecting permissions does not guarantee that failures are easier to detect. The system card reports that the monitorability of chain of thought, the intermediate reasoning verbalized before an answer, is lower for Astra than for Sol, even under conditions that do not encourage evading monitoring. OpenAI gives as the main reason an increase in cases where the chain of thought is short or empty, leaving the monitor less to read. While the average metric for monitoring that reads actions alone improved, monitoring that reads reasoning and actions together also showed an overall decline.
In tests that encouraged evading monitoring, there were cases where instructed evasion or behavior that made capabilities look lower slipped past monitors. However, it was not confirmed that the chain of thought was covertly encoded. Nor can anything be asserted about the model's internal mechanisms beyond the published materials.
Apollo Research's evaluation, included in the system card, also cautions that a low rate of misbehavior alone does not prove the model is aligned with human intent. The reasons are that the model may be aware it is being evaluated and that the observation period was short.
For external availability too, OpenAI is introducing a mechanism that monitors the reasoning and actions of Astra when it uses tools and stops it when it detects suspected unauthorized activity. Additional safety checks or false positives may delay or halt legitimate work. In ChatGPT and Codex, the user may be asked to confirm; API tasks stop. You cannot assume automatic retries. The more privileges you hand over, the more the procedure needs to specify when a person checks, resumes and accepts the result.
Pricing, staged rollout, and what to measure beyond "AGI"
Astra began rolling out to some organizations on September 3, with the target group expanding over the next few days. In ChatGPT, Plus and Pro and the business-oriented Business and Enterprise plans are covered, and it is also being rolled out on the OpenAI API and AWS. It can be used within existing subscription limits, and additional credits can be purchased. GPT-6 Astra Pro is also provided for Pro, Business and Enterprise, but in Enterprise workspaces it is off by default until an administrator enables it. The announcement does not state availability for Free or Edu, or completion dates for each account.
The OpenAI API model ID is gpt-6-astra, and Standard pricing is $10 per million input tokens and $50 per million output tokens. Separate rates apply to cache reads and writes. Fast is described as up to 2.5 times the speed of Standard at twice the price. Amazon Bedrock is also among the providers, but this OpenAI API pricing must not be applied to Bedrock or Astra Pro.
According to Axios, co-founder Greg Brockman said at a press briefing that he personally believes AGI has been reached, while indicating that the judgment is left to users. AGI as defined in the OpenAI Charter is "highly autonomous systems that outperform humans at most economically valuable work," and is not a pass line on a single benchmark. Mr. Brockman's personal assessment needs to be kept separate from demonstrating capability across a broad range of work.
Adoption decisions begin with recording, for your own representative screen-operation work, how much can be completed correctly, including token costs and the human time for checking and resuming. If high completion rates can be reproduced and people can handle mid-task checks and restarts after stops, the range of long jobs you can hand over using existing business software will widen.