We Tested OpenAI's GPT 5.6 for a Month

Dan Chipper rates GPT 5.6 as the most usable, powerful, and fast model currently available for knowledge work and coding, placing it at roughly 60 to 70 percent of Claude Fable's capability. GPT 5.5 was insufficient for his use cases, but 5.6 produces email drafts that read as if he wrote them himself, with approximately 90 percent requiring no changes or only minor edits. The model makes fewer mistakes than prior versions, is more careful about not sending emails the user would not want sent, and now offers a recommended choice among product options rather than presenting options without a point of view.

Chipper describes 5.6 and Codex together as his operating system for work, covering email, writing, and model training. Codex is OpenAI's answer to Claude Code and Claude Desktop, built after observing the paradigm those tools established, and is considered a cleaner and more powerful implementation than the Codex CLI. It functions like an operating system because it combines a powerful model with an in-app browser and persistent context about the user, giving the model access to the whole computer and the whole internet. Dan Shipper clarifies that Codex is not purely a coding tool despite its name, and notes two notable features: Chronicle automatically takes screenshots and converts them locally into a contextual feed about the user's activity, with on-device processing being the reason Shipper keeps it enabled, and a Record and Replay plugin allows users to demonstrate a task on screen and have Codex convert it into a reusable skill.

Chipper built an open source email app called Ten that sweeps the inbox, creates summary cards per email, and drafts replies automatically, learning over time what the user cares about and revising its prompts based on archived or replied-to items. A separate system called Mailroom gives a Codex agent a dedicated email address using the plus-format to route tasks automatically, and a Slack internal agent is configured to email the Codex agent rather than Chipper directly for operational tasks. Codex threads can message other Codex threads, enabling orchestration of multiple agents. A Codex Goal feature allows the model to pursue a high-level long-term objective until verifiable completion criteria are met, with one model training task running four hours continuously and weekend tasks running 20 hours each. Router threads and heartbeat threads are described as key architectural patterns for using Codex effectively.

Fable is significantly more powerful than 5.6 but so slow and expensive that it is only practical for machine learning work or orchestrating fleets of agents, a skill most people do not yet have. Fine-tuning a model is now practical for non-machine learning engineers using 5.6 and Fable, covering data collection, synthetic data generation, and running experiments. OpenAI and Anthropic no longer meaningfully offer fine-tuning, making open source models necessary for that use case. One speaker notes their organization has accumulated tens of thousands to hundreds of thousands of copy edits over five to six years as potential fine-tuning training data, but current frontier models cannot perform copy editing to a human copy editor's standard even at a 70 percent threshold.

On design, a checkpoint feature introduced with 5.6 brings design quality from a C or D rating to roughly a B, though the model defaults to warm paper backgrounds without clear reason. Claude, including Opus and version 4.8, is considered better than GPT-based tools for visual design work, particularly for landing pages.

The SaaS Apocalypse narrative is described as overstated because 99 percent of the world will want software maintained by someone else rather than one-shotting everything. Building and maintaining software over time remains expensive and hard, making that the rare and valuable capability. One speaker switched from vibe-coding a CRM to using a CRM vendor with a CLI because the vendor handled corner cases and ongoing maintenance. Codex-native SaaS providers bear no token costs because they host software rather than run AI, restoring normal software business margins, contrasted with a company making 80 million dollars per year in revenue but spending 60 million dollars per year on tokens. A prediction is made that within a year OpenAI, Anthropic, and possibly Cursor will have app stores specifically for agent-native apps, with the mobile-era App Store distribution model cited as the likely playbook.

Shipper describes the core mindset shift required to use Codex effectively as moving from doing individual tasks yourself to managing the system that does the tasks. His recommended onboarding sequence is to download the tool, grant it computer access, give it context about who you are, and ask it to configure itself. He advises new users to start with one specific problem rather than attempting to build a comprehensive system from the beginning, cautions that a setup working well for one person may not transfer to another, and argues that genuine curiosity about a concrete use case is a more reliable entry point than FOMO.