Inside Codex's Agent Loop: Lessons from Datadog and Sora's 28-Day Android Launch
The Codex agent loop is a simple control flow — prepare prompt, model inference, optional tool call, append result, repeat — that terminates when the model emits a normal assistant message. Three implementation details make it production-grade: prefix caching on the Responses API, tool-call results that live inside the conversation, and termination detected from output shape (not a turn counter). Datadog uses the loop for system-level code review on every PR across 1,000+ engineers; its incident-replay harness showed Codex catching ~22% of historical incident-related issues human reviewers had missed. OpenAI shipped Sora for Android in 28 days with four engineers, ~85% Codex-generated code, and a 99.9% crash-free rate — by scaffolding with exemplars and AGENTS.md, not by prompting 'go build it.' The same agent-loop architecture transfers to media-generation agents built on hiapi: stable system prompt + tool-call accumulation + critique-based termination, wrapping image and video endpoints.
hiapi
Guide · Jun 8, 2026 · 12
ChatGPT Image Generation: What's New and How to Build It Programmatically
ChatGPT's new image experience is powered by three models — gpt-image-1.5, gpt-image-2, and gpt-image-2-pro — all reachable through a standard OpenAI-compatible API. The visible improvements (text rendering, character consistency, 4K detail) come from the model layer, not the UI — so the same quality is available to your own application. gpt-image-2 is the new flagship at $0.03 per image (cheaper than gpt-image-1.5 at $0.05) and handles the majority of production workflows. Migration from ChatGPT to API is small: same prompts work, the request shape is plain chat/completions, and the image returns inline as base64.
GPT Image 2 Multi-Turn Editing and Style Consistency: A hiapi Capabilities Tour
gpt-image-2 (text-to-image) and gpt-image-2-image-to-image are two halves of the same workflow — first draft and surgical editor — both priced at $0.03 per call at 1K on hiapi. The two patterns that move the needle: a verbatim character bible reused across scenes for series consistency, and a first-render-then-edit two-call flow for any job that must look like an extension of an existing brand asset. Both variants speak the OpenAI-compatible Chat Completions payload at /v1/chat/completions — the only difference is whether the user message carries a string or a list of image_url plus text parts.
Does GPT Image 2 Really Nail Text? We Stress-Tested Signage, Posters and Labels
Across 24 text-heavy generations — including English headlines, CJK calligraphy, multi-string infographics, embedded brand blocks, currency, and dense UI mockups — every primary string rendered correctly. GPT Image 2 will sometimes add contextually appropriate text you didn't ask for (brand monograms, founding-year badges, scent names). Useful most of the time, undesired occasionally. The remaining ~1% error rate at production scale shows up on the smallest text — sub-12pt captions, dense legal copy, ingredient lists. Always proof at 100% zoom before shipping. If text matters, lock it in straight quotes. Smart quotes get rendered as smart quotes. Special characters (em-dashes, ampersands) work; single-character substitutions are the most common failure mode when they occur.
Building an E-commerce Product Image Workflow with GPT Image 2
A full six-step pipeline for producing an e-commerce product listing — main hero, lifestyle, macro detail, ingredient story, promotional banner, launch poster — using GPT Image 2 end-to-end. All six images for one product cost about $0.18 to generate. Wall-clock time: roughly ten minutes including text proofing. Use the same anchor description (the product) across every prompt to keep visual consistency between shots. Always proof rendered text by eye — GPT Image 2 is right ~99% of the time, but the 1% is what ships.
GPT Image 2 Review: A Hands-On Honest Take
Text rendering is the real upgrade — English headlines, CJK calligraphy, in-image labels and prices all render correctly the first time. Photorealism is genuinely strong for product photography and detailed interiors. Portraits are good but occasionally read as slightly 'too symmetric'. Speed is a step backward from GPT Image 1.5: ~107 seconds per image versus 18–36 seconds. Plan workflows around it. Hard limit: no transparent backgrounds. If you need PNG cutouts, you're still going to need a second model or a background remover.
GPT Image 2 Prompts That Worked: Templates With Real Outputs
Twelve prompt templates organized by job — product photography, posters with text, infographics, UI mockups, illustration — each shown with the actual image GPT Image 2 produced. Every prompt here is a real working template, not a 'sample prompt' that needs further refinement. Copy, swap the bracketed variables, generate. The most important pattern across all of these: name the job, lock the text in quotes, describe composition explicitly, state what must not appear. GPT Image 2 will often add contextually appropriate detail beyond what you asked for. Plan for it — leave room rather than over-specifying.
GPT Image 2 vs Nano Banana 2: Side-by-Side With the Same Prompts
GPT Image 2's marquee text-rendering advantage has narrowed — Nano Banana 2 renders English and CJK headlines just as accurately, and even adds contextually appropriate label text on its own. Nano Banana 2 produces more believable photorealism (portraits, product photography), while GPT Image 2 keeps a slight edge on editorial typography craft. Nano Banana 2 effectively ignores precise aspect-ratio requests — square or near-square is what you get. GPT Image 2 honors 3:2, 2:3, and 9:16 reliably. Price: GPT Image 2 is $0.03 per image, Nano Banana 2 is $0.085 — 2.8× the cost. Speed is comparable on average.
GPT Image 2 API Pricing: What You Actually Pay Per Image
GPT Image 2 starts at $0.03 per image at 1K resolution — about 40% cheaper than its predecessor Cost scales with output resolution: 1K at $0.03, 2K at ~$0.04, and 4K at $0.06 per image Most use cases — product mockups, social posts, quick prototypes — are well served by 1K or 2K Three practical levers can cut per-image cost: drop to a lower resolution tier, reuse similar prompts for cache hits, or batch generations of related images
Stay ahead of the curve.
Get the latest model launches, engineering deep-dives, and API tutorials delivered straight to your inbox.
Email subscription is coming soon
No spam. Quality content on AI image & video generation, unsubscribe anytime.