New GPT-3 capabilities: Edit & insert

For as long as GPT-3 has been public, it's done one thing: take your prompt and keep going. You type a sentence, it finishes the thought. Useful, sure, but also limiting if what you actually wanted was to fix a paragraph in the middle of a document or patch a broken function in the middle of a script. OpenAI's latest update changes that math entirely by adding edit and insert modes to both GPT-3 and Codex.

The practical difference is bigger than it sounds. Instead of feeding a model text and hoping the continuation happens to fix your problem, you can now hand it existing content plus an instruction — rewrite this for clarity, add error handling here, swap the tone from formal to casual — and it will operate directly on that text. Insertion works the same way in reverse: give it a start and an end, and the model fills the gap intelligently instead of just tacking new material onto the tail end.

For coders this is arguably the bigger deal. Codex has been decent at generating fresh functions from scratch, but real engineering work is mostly editing existing code — refactoring, adding a missing feature, tightening a loop. An edit-capable Codex means the tool finally matches how programmers actually spend their day, rather than how a demo video makes it look.

OpenAI frames this as a step toward models that behave more like collaborators than autocomplete engines. That's marketing language, but the underlying shift is legitimate: completion-only models forced every editing task into an awkward workaround where you'd delete the parts you wanted changed and pray the regenerated text improved things. Direct editing removes that guesswork and, more importantly, removes a whole layer of prompt-engineering gymnastics that users had built up just to compensate for a missing feature.