OpenAI opens GPT-4 API to all paying developers

If you've been stuck on OpenAI's GPT-4 waitlist since March, your inbox can finally get a rest.

OpenAI just officially opened up the GPT-4 API to general availability. Any developer with an active account and at least one successful billing history can now call the 8K context GPT-4 model directly. New developers will get access by the end of July, with rate limits climbing higher as compute capacity allows.

Alongside the flagship rollout, APIs for GPT-3.5 Turbo, DALL·E, and Whisper are now officially declared production-ready. Fine-tuning support for both GPT-4 and GPT-3.5 Turbo is also in the pipeline, targeted for release later this year.

The death of freeform prompts

The bigger story for engineers is a major platform cleanup: OpenAI is officially retiring the legacy Completions API.

When OpenAI dropped the Chat Completions API back in March, critics wondered if formatting everything into messages was overkill. Turns out it won out completely. OpenAI revealed that Chat Completions already handles 97% of its total GPT API traffic.

The structured format didn't just unlock multi-turn dialogue and function calling—it built a defensive wall around prompt injection by cleanly separating system instructions from messy user input. Starting today, the old Completions API is stamped as legacy, and no new models will ever ship for it.

Mark January 4, 2024 on your calendar

Developers have six months to rip out old endpoints before the lights go out on January 4, 2024.

Classic GPT-3 base models like davinci and curie will automatically jump to davinci-002 and babbage-002. Anyone relying on text-davinci-003 will need to repoint their calls to a drop-in replacement named gpt-3.5-turbo-instruct. Both test beds will drop in the coming weeks.

Legacy embedding models are also getting swept into text-embedding-ada-002, which already represents 99.9% of embedding usage. Because re-indexing entire databases costs real money, OpenAI says it will cover the financial cost for affected teams re-embedding their data.

Why it matters

Lifting the waitlist turns GPT-4 from an exclusive research prototype into reliable utility infrastructure. Up until now, thousands of startups had to patch together unstable workflows on GPT-3.5 because they couldn't get stable API keys. With direct access unlocked, we're about to see software shipped with dramatically higher reasoning capabilities across the board.

Phasing out raw text completion is equally telling. It signals that structured, steerable APIs with built-in guardrails are the only architecture OpenAI believes in moving forward. Cleaning house six months out is a sharp, disciplined operational move.

Covering customer costs for re-embedding content is a rare class act in SaaS infrastructure that buys tremendous developer goodwill. But don't sleep on capacity: compute remains tight, and teams scaling up in production should still prepare for rate limit friction during peak traffic spikes.