GPT-6 Astra, Image 2.5 & Codex Updates: Hands-on Review & Quota Tips

GPT-6 Astra Computer Use

The author identifies Computer Use as GPT-6 Astra's standout capability. While many demonstrations focus on Blender 3D modeling for visual impact, the author tested general computer operation and found it highly capable, succeeding at tasks that "Work"-class applications could not handle.

For Blender testing, the author completed installation then delegated the rest to GPT-6 Astra with a prompt (shown in an accompanying image). The task took approximately 1.5 hours with thinking depth set to high, consuming roughly 10% of the weekly quota. The one-shot result exceeded expectations, though close inspection reveals numerous minor issues.

Image 2.5 Features

Image 2.5 has been rolled out to ChatGPT, ChatGPT Work, and Codex across desktop, mobile, and web. The author notes that desktop output quality lags behind web for identical prompts, cause unknown.

Sketch-to-Image (@Sketch)

In ChatGPT, users can invoke @Sketch to draw a rough outline; the model then generates a high-quality image from the sketch.

Templates

Built-in templates cover posters, product images, multi-panel brochures, and other ready-made layouts. Selecting a template auto-fills a prompt (in English; the author translated it for the example).

Precise Inpainting

Generated images support precise editing: users can circle an area or specify text changes to modify only the selected region while preserving other details. The official claim states it handles complex subjects and backgrounds with ease.

The author demonstrates by merging their own avatar with a full-body Steve Jobs image ("Zhang Bu Si"), then changing clothes, gestures, and adding an iPhone Duo. A noted limitation: the model struggles with real-world physical dimensions — despite reference images showing correct phone-to-hand scale, it often produces oversized or undersized phones.

Generation Progress Indicator

A small UX improvement: the interface now shows a percentage progress indicator in the bottom-right corner during image generation.

ChatGPT/Codex Client Updates

Computer History

Computer History records all user operations on the computer, automatically identifying sequences that can be turned into Skills. The author considers this a critical capability, though it consumes quota. It logs every action and suggests skill creation.

Example output from the history (captured in a code block):

You refined a Doubao Seed model evaluation document, during which you moved materials between Obsidian, ChatGPT, Doubao, Finder, and Feishu/Lark reporting documents. You cleaned up case notes, removed invalid image references, copied summary content, and apparently pasted the prepared case evaluations into an online collection form.

The system then drafts the prompt for generating the corresponding skill.

Record & Replay Skill Recording

An official OpenAI plugin adds a "Record Skill" function. Users open Record & Replay, perform a workflow once (up to 30 minutes per recording). ChatGPT observes clicks, inputs, and window content, then auto-generates a Skill. Subsequent similar tasks can be executed by invoking that Skill.

The author highlights the paradigm shift: skill creation moves from "writing prompts" to "demonstrating to the AI," akin to an apprenticeship model.

Analytics Page

Settings now include an Analytics page showing weekly or monthly consumption for desktop and web, model usage breakdown, and tool/skill invocation statistics.

Quota Management Tips

Avoid GPT-6 Ultra: The author recommends keeping effort at "Medium" or "High" rather than Ultra. Citing Tibo, GPT-6 Astra at low effort outperforms GPT-5.6 Sol at high effort.

Separate ChatGPT and Codex roles: Use ChatGPT (as product manager) to clarify requirements, then hand off to Codex (as programmer) for implementation. Avoid using Codex mode for non-coding tasks.

The author concludes that ChatGPT Pro remains the only indispensable AI subscription, having briefly downgraded to Plus and found it insufficient, then returning to Pro×5. The time saved far outweighs the cost.