Merge Uses GPT-6 Astra and Codex to Reshape Product and Business Tools

From Code Review to Executive Dinners: A Full Picture of Merge's AI Adoption

While most companies are still debating whether to use large models at all, Merge has already embedded the OpenAI technology stack into product development, marketing support, sales operations, and even the preparation of executive events. According to an official OpenAI case study, Merge is using the trio of GPT-6 Astra, ChatGPT, and Codex, covering multiple business lines from product code review to marketing content production.

This combination isn't mysterious, but the way it's used is worth unpacking: GPT-6 Astra handles more complex reasoning and generation tasks, ChatGPT serves as the team's daily interaction entry point, and Codex specializes in code-level understanding and review. The clear division of labor shows that Merge isn't treating AI as a universal hammer, but rather selecting tools by scenario.

Product Side: Codex Enters the Code Review Process

Merge's core business is providing enterprises with integration and API management capabilities, and the quality of its product code directly affects customer trust. The source material shows that Merge uses Codex to review product code. This means AI is no longer停留在 the level of "completing a few lines of code," but participating in understanding and vetting existing codebases.

Code review has always been a time-consuming process that depends on the experience of senior engineers. Letting models participate directly shortens feedback cycles and screens out obvious issues before human review. But this also places new demands on the team: review standards must be clear and context must be sufficient, otherwise the model's suggestions are hard to act on.

Codex is a series of code-specialized models introduced by OpenAI, initially known for powering GitHub Copilot's code completion capabilities, with later versions gradually expanding to code generation, explanation, refactoring, and review tasks. Compared with general conversational models, Codex places more emphasis on code corpora and programming language structures during training, so it performs more consistently in understanding function call relationships, identifying potential defects, and following coding standards. In code review scenarios, it typically intervenes by "screening first": checking naming, boundary conditions, exception handling, and obvious logic flaws against team standards, and front-loading low-value manual verification work. It should be noted that model review cannot replace senior engineers' judgment on architectural trade-offs and business semantics, and its effectiveness depends heavily on the context scope provided in prompts and a clear review checklist.

Business Side: ChatGPT Supports Marketing and Sales Operations

Beyond R&D, Merge's team uses ChatGPT to support marketing work and prepare sales update materials. What these two types of tasks have in common is: high information density, clear format requirements, and strong repetitiveness—exactly the areas where current large models excel.

One detail: the source material also mentions that the team uses AI to plan executive dinners. Such tasks may seem "soft," but actually involve multiple constraints such as agenda design, guest matching, and topic arrangement. The fact that they can be handed to AI shows the team has considerable confidence in the model's performance on open-ended planning tasks.

Sales updates typically refer to periodic reports for customers or internal management, covering customer progress, opportunity status, changes in key metrics, and next steps. The difficulty of such materials lies not in the writing itself, but in consolidating information scattered across CRM, meeting notes, and emails into a clearly structured and consistent narrative. The value of large models here is mainly format conversion and information compression: organizing scattered notes into standard templates, unifying wording style, and adjusting detail levels for the audience. Marketing work is similar—from channel copy and campaign briefs to customer case rewrites, these are all "sufficient input, clear format, batch-processable" tasks, suitable for models to produce first drafts, followed by human fact-checking and brand tone vetting.

A Signal Worth Noticing

Merge's case reflects a trend: the use of AI tools is moving from "single-point trials" to "cross-departmental embedding." The same OpenAI technology stack serves the engineering team's code quality, the output efficiency of marketing and sales, and even extends to event planning at the executive level.

The prerequisite for this breadth is a clear understanding of tool boundaries—code review goes to Codex, daily generation and communication go to ChatGPT, and complex reasoning goes to GPT-6 Astra. The finer the granularity of tool selection, the more measurable AI's actual business gains become.

Takeaways for Similar Teams

If your team is preparing to expand its use of AI, Merge's approach offers several ideas worth borrowing:

• Match models to task types, rather than having the whole company use a single entry point;
• Prioritize areas with clear acceptance criteria, such as code review and sales material preparation;
• Don't underestimate planning tasks—event planning and agenda design can also be effectively assisted by AI;
• Reuse the same technology stack across departments to reduce integration and training costs.

Merge's practice shows that AI's value in the enterprise doesn't depend on how new the model is, but on whether it has found a position that meshes with business processes.