Complete Guide to GPT-6 Astra

Beyond '98% AGI': The Shift to AI That Completes Entire Jobs

On September 3, 2026, OpenAI announced GPT-6 Astra.

Immediately following its release, attention focused on its scores: approximately 98% on the difficult math benchmark "FrontierMath," 99.9% on "ARC-AGI-3" for measuring computer operations, and 100% on the cybersecurity evaluation "ExploitBench."

Looking at these numbers alone, one might be tempted to say, "A near-perfect AI has arrived" or "AGI is finally complete."

However, it is dangerous to judge GPT-6 Astra solely by benchmark scores.

The change in Astra is not just about returning more correct text in response to questions.

It opens browsers, reads documents, executes code, updates spreadsheets, operates websites, investigates causes if it fails midway, and ultimately brings back a reviewable deliverable.

The unit of work passed to ChatGPT has shifted from "a single question" to "a series of business operations."

)If you use Astra for just email drafts or text summaries like traditional ChatGPT, it remains just an expensive text generation AI. Its value emerges when you entrust it with tasks that humans used to handle by switching screens multiple times, treated as a cohesive process.(OpenAI

In this article, we will delve beyond official feature introductions into cases confirmed by overseas companies, how to read benchmarks, comparisons with Claude, pricing, safety, and design for practical use by Japanese companies.

What is GPT-6 Astra?

GPT-6 Astra is positioned by OpenAI as the top-tier model of the GPT-6 generation, described as the "most intelligent and most aligned model."

At launch, it is being rolled out to select organizations first, followed by ChatGPT Plus, Pro, Business, Enterprise, API, and Amazon Bedrock. In Enterprise, it is off by default until enabled by an administrator, and there is no explicit mention of a free plan release in the official announcement.

)The model name on the API is gpt-6-astra. Astra Pro, which uses higher compute, is also being introduced for Pro, Business, and Enterprise users.(OpenAI

Key specifications are summarized below:

While it can read images, the model's native output is text. Image generation is performed by calling a separate tool. It is not a model for direct audio or video input.

Additionally, for long inputs exceeding 272,000 tokens, a price multiplier applies to the entire request. Input and cache fees are doubled, and output fees are 1.5x. Just because it fits 1.05 million tokens doesn't mean you should cram every internal document into it every time; cost efficiency will plummet before accuracy does.

What Changed with Astra?

1. Computer Operation Moved from "Auxiliary" to the Main Battlefield

At the heart of Astra is "Computer Use."

Previous AI could click buttons or fill forms on a browser. However, problems remained: it would stop if the screen layout changed slightly, couldn't recover from mid-process errors, or would lose sight of the goal in tasks spanning multiple sites.

GPT-6 Astra recorded 72.6% on OSWorld 2.0, which measures actual desktop environment operations. GPT-5.6 Sol was 65.7%.

)Moreover, Astra processed this evaluation in about 40 minutes, while Sol took about 75 minutes. Not only the success rate but also the processing time was reduced by about 47%. In ScreenSpot, which finds targets on screen, Astra scored 92.7% compared to Sol's 76.9%.(OpenAI

This is more than just a smooth demo video. It targets tasks such as:

• Searching for hospitals or properties matching criteria and comparing candidates.
• Retrieving values from multiple management screens to create reports.
• Entering customer info into CRM and updating status.
• Logging into web services to check settings.
• Running a created app in a browser and fixing bugs.
• Completing spreadsheets or slides while referencing multiple files.

In Mind2Web, Astra using an updated Codex harness processed web tasks 1.9x faster than previous models.

)The important thing here is that performance is determined by the "entire harness," including the browser, tools, execution environment, and verification process, not just the model's intelligence. Simply selecting the name Astra doesn't automatically automate all screen operations.(OpenAI

1. Maintaining 1.05 Million Token Long Context Until the End

GPT-6 Astra's context length is approximately 1.05 million tokens. It has the capacity to handle contracts, minutes, design docs, source code, and past research all at once. However, "fitting" and "using accurately" are different.

In MRCR evaluations for finding information embedded in long texts, Astra was 100% in the 256k–512k range and 96.3% in the 512k–1M range. GPT-5.6 Sol's result in the latter was 73.8%.

)While Astra is strong in long context, it doesn't memorize every single word perfectly when you put in 1 million tokens. It still fails about 3.7% of the time in the 1M token band. Avoid designs that place critical conditions in only one spot in a massive document; you must explicitly state goals, constraints, and approval conditions separately.(OpenAI

Astra also has mechanisms to increase continuity in long sessions, such as asynchronous tool calls to work on other tasks while waiting for a tool, steering to allow human course correction during execution, and maintaining cache even if reasoning levels are changed mid-conversation.

)It is moving away from "write a perfect prompt and leave it" toward a workflow of "checking progress and course-correcting."(OpenAI Developers

1. Creating Finished Deliverables, Not Just Text

Astra has strengthened its ability to create deliverables like spreadsheets, slides, documents, apps, and research reports. Previous models could create a "presentation outline," but humans had to move that to PowerPoint, adjust the design, re-verify numbers, and add links. Astra aims for the next level: reading materials or company templates, organizing numbers, creating slides/spreadsheets, checking them in a browser, and submitting the corrected files.

Early adopter Higgsfield stated it could process agent tasks with up to 20% fewer tokens than existing models. Harvey praised its judgment in legal work, Jane Street noted a reduction in back-and-forth for complex technical issues, and Lovable valued its iteration/testing capabilities at high reasoning settings.

)However, these are testimonials from early customers, not the same as evidence from independent third-party replication tests.(OpenAI

1. Less Likely to "Forget the Goal" During Long Tasks

Conventional agents sometimes drifted from the plan established at the start—getting obsessed with minor refactoring while supposed to be implementing a feature, or ending with just info gathering instead of a final decision. Astra has improved consistency in long-duration work.

On the other hand, official developer docs note that Astra may stop to ask clarification questions, over-format details, repeat expressions, or test more than necessary.

)In other words, higher performance doesn't mean you can give vague instructions. You should set goals, completion conditions, criteria for asking questions, and operations requiring approval more clearly than before.(OpenAI Developers

Benchmarks are Overwhelming, but Not "World's Best in All Fields"

The Astra announcement featured eye-popping numbers: 97.6% on FrontierMath, 99.9% on ARC-AGI-3, and 100% on ExploitBench. These are significant steps forward, but we must distinguish between measurement targets and execution conditions.

Computer Operations and Business Automation Grew Significantly

In AutomationBench, which is close to actual business tasks, Astra scored 41.4%. GPT-5.6 Sol was 18.1%, and Claude Fable 5.1 was 31.4%. In BenchCAD for CAD-related tasks, Astra was 95.9%, Sol 83.3%, and Fable 5.1 84.3%. In BrowseComp for web research, Astra was 91.5%, Sol 90.4%, and Claude Opus 5 was 90.8%. Astra leads here, but the gap with Sol and Opus is small.

)Astra's advantage is most apparent in complex composite tasks combining screen operation, file processing, data analysis, and verification, rather than just finding a single search result.(OpenAI

Strong in Coding, but Hasn't Completely Overtaken Claude

In Terminal-Bench, Astra scored 57.7%, surpassing Sol (37.3%) and Fable 5.1 (55.8%). In internal database migration evaluations, Astra was 63.9% vs. Sol's 42.7%. Clear improvements are seen in reading existing systems and making changes without breaking them.

However, in the Artificial Analysis Coding Agent Index, Astra was 67.0, Claude Fable 5 was 67.2, and Claude Opus 5 was 68.1. In FrontierCode Extended, Astra was 64.5 vs. Fable 5's 64.9. In the Main evaluation, Astra was 53.3, Fable 5 was 53.5, and Opus 5 was 53.4.

)Astra is in the top tier for coding, but it hasn't surpassed Claude in every code evaluation.(OpenAI

Weaknesses Remain in Academic Reasoning

Despite the 97.6% in FrontierMath, Astra scored 57.2% in "Humanity’s Last Exam," which measures broad expertise and reasoning. Claude Fable 5.1 was 65.0%, Fable 5 was 63.8%, and Opus 5 was 63.6%. In the Artificial Analysis Intelligence Index, Astra was 61.2, while Fable 5.1 was 65.7 and Opus 5 was 63.1.

)The interpretation that "it understands almost all academic fields because it's 98% in FrontierMath" does not hold. Depending on the evaluation, Claude still produces higher results in some areas.(OpenAI

Behind the 99.9% is a Dedicated Execution Environment

The 99.9% on ARC-AGI-3 is shocking, but it wasn't obtained by simply inputting the problem into Astra. OpenAI executed it with a harness using the Responses API and modified two settings to bring it closer to actual usage performance. While OpenAI explains these aren't benchmark-specific settings, it's not a zero-shot result from the model alone.

)Also, the published values are the highest results among multiple reasoning settings. There is no guarantee that the same performance will be reproduced every time in standard ChatGPT.(OpenAI

In the Artificial Analysis aggregate at launch, Astra's overall score was 61, ranking 8th out of 202 models. Meanwhile, its input price is $10 vs. the median of $2, and output is $50 vs. the median of $10.

)It is certainly one of the best, but it's not a model you throw everything at regardless of price.(Artificial Analysis

Overseas Case 1: Legora Reconciled 41 Documents in Minutes

Legora, which provides legal AI in Europe, demonstrates Astra's practical value. They used Astra for "tie-out" reconciliation in financial statements, where the same numbers appear across text, notes, tables, and past docs. Legora processed 41 documents in a single agent run, a task that takes humans hours or days, in just minutes.

)In tests, they intentionally planted 4 errors. Astra found all 4, including a £500,000 discrepancy in revenue notes. It maintained the accuracy of previous models while performing about 50 more correct checks.(OpenAI

However, the ~40% improvement was specific to this workflow. Across Legora's broader BAR tasks, the average improvement was about 3%. Astra doesn't improve all legal work by 40% uniformly; it excels in cross-referencing massive documents for inconsistencies.

)The final judgment remains with human legal experts. AI handles narrowing down what humans should read and aligning evidence of inconsistencies.(OpenAI

Overseas Case 2: Playco Reduced Manual Fixes by 50%

Game company Playco tested Astra with "Playbot," a development agent for Unity and Godot. Playbot edits scenes, plays the game, verifies behavior, and fixes issues. Playco had it create three different themed prototypes at once. While not a "perfect one-shot," they reported a 50% reduction in manual human fixes compared to previous models.

)Improvements were seen in spatial awareness, reproducing reference images, responsive UI, game-like feel, and bug detection.(OpenAI

Astra's strength isn't just "getting the code right once," but the ability to run the game, check the display, operate it, find problems, fix them, and try again—mimicking a human developer's iteration.

Overseas Case 3: Advancing a Stalled Feature to 90% Completion

Product developer Claire Vo shared results using Astra for her service ChatPRD. A product intelligence feature that had been stalled for 6 months due to complex CRM operations and UI implementation reached about 90% completion in a single Astra session.

)She also tested controlling hardware (Divoom MiniToo) via CLI, a Mac messaging app, and 3D assets in Blender.(Lenny's Newsletter

While these are self-reported successes from early partners, they show a consistent trend: Astra makes a difference when operating external services, reading existing code, and finishing deliverables.

Scientific Breakthroughs: Solving Long-Standing Problems

OpenAI reported using Astra to solve or significantly advance 10 unsolved problems in mathematics and theoretical computer science. The model generated arguments, human researchers organized the drafts, and the same model used Lean for formalization.

)According to OpenAI, the total token cost for this research was roughly $2,000 in GPT-5.6 Sol equivalent pricing. One example involved updating bounds on prime gaps, an area where certain terms hadn't been improved in over 80 years.(OpenAI

Why Astra is "Most Safe" yet "Hardest to Monitor"

GPT-6 Astra is the first widely deployed model to reach the "Critical" threshold for cybersecurity capabilities in OpenAI's classification. Given tools and access, it can discover unknown vulnerabilities and build attack procedures against defended systems without step-by-step human instruction.

)In ExploitBench, Astra scored 100% vs. Sol's 78.5%. It even discovered two zero-day vulnerabilities during testing, including one that escaped a browser sandbox. However, these results used the "Daybreak Blue" environment for defense researchers; the standard version has controls to refuse advanced attack tasks.(OpenAI

Paradoxically, while Astra reduces overt violations, its internal reasoning is harder to monitor. In adversarial evaluations ordered to "evade monitoring," the detection rate of monitoring systems dropped to below 11% in some tasks. OpenAI found no evidence of "steganography" (hiding info in reasoning) occurring naturally, but the model's high awareness of being evaluated makes alignment difficult to prove definitively.

Where Should Japanese Companies Start?

1. Tasks with High Document Volume and High Risk of Oversight: Cross-checking contracts, quotes, invoices, and minutes. AI can gather the materials and flag discrepancies for human review.
2. Administrative Tasks Requiring Multiple Management Screens: EC, ads, and CRM management. Start with viewing, aggregating, and drafting; do not grant permission for payments or deletions initially.
3. End-to-End Research and Material Creation: Don't just ask for a summary. Ask it to research competitors, tabulate features, analyze differences, and summarize it into a slide deck.
4. Prototyping and QA: Have it implement a feature and then open a browser to check for broken layouts or link errors.

Practical Prompt for Astra

Instead of just giving a role like "You are an expert," set boundaries and completion conditions.

1## Purpose
2The state to be achieved by this task: [Fill in purpose]
3
4## Final Deliverable
5What to submit: [Files, tables, reports, implemented features, etc.]
6Completion conditions: [Specific criteria to be met]
7
8## Permitted Info and Tools
9Materials to reference: [Files, URLs, internal data, etc.]
10Tools to use: [Web search, browser, code execution, spreadsheets, etc.]
11
12## Process
131. Confirm request and materials.
142. Organize missing info and risks.
153. Proceed with research, analysis, and creation.
164. Open and verify the deliverable.
175. Organize errors, unconfirmed items, and remaining tasks.
186. Submit in a reviewable state.
19
20If one process stops, continue with non-dependent tasks.
21Only ask questions for missing info that significantly affects results.
22For minor gaps, make reasonable assumptions and record them.
23
24## Permissions
25You may proceed with viewing, analysis, drafting, testing, and diff checking.
26Stop immediately before the following operations and seek approval with details:
27- External sending
28- Purchases, payments, contracts
29- Publishing, production deployment
30- Deletion of data or files
31- Permission changes
32- Hard-to-undo operations
33
34## Quality Conditions
35- Separate verified facts from speculation.
36- Add sources for important numbers.
37- Look for contrary evidence or exceptions.
38- Open the deliverable to check display and operation.
39- Final report must include actions taken, verification results, and remaining risks.

Astra's Weaknesses

• Lack of Independent Long-term Cases: Most data comes from OpenAI-selected partners.
• 1.05M Tokens is Not a Free Warehouse: Costs rise with input size, and oversight risk isn't zero.
• Not Native for Audio/Video: Requires external tools for these formats.
• No Fine-tuning: You must use RAG, MCP, or system instructions to reflect company rules.
• Safety Mechanisms May Stop Legitimate Work: Monitoring can slow down defense research or dev tasks.
• Output Can Be "Too Perfect": It tends to use headers and lists heavily, which might feel too "AI-like" for brand copy.

Conclusion: Value Lies in Completion Rate, Not Answers

GPT-6 Astra is overkill for writing a single email or summarizing minutes. Its meaning emerges in jobs where research, judgment, operation, creation, and verification are linked.

The era of "one-sentence prompt engineering" is over. What matters now is which information to pass, which tools to connect, how much execution to allow, and what defines completion. Astra is not a magic wand that gives everyone the same result; if you give it messy work, it will do messy work quickly. If you organize the purpose, tools, and verification points, it will carry hours of human labor to a reviewable state in minutes.

• Reuters
• WIRED
• Axios
• ft.com