claude code and openai codex cli are gimmicks
claude code and openai codex cli feel like gimmick offerings if viewed from a medium-long term perspective.
i am not denying they are bad per se but they feel like a means to an end - given the potential of innovation that can be done by the talented people at both these places. the end result i am envisioning is - just a code editor - some sort of next-gen code editor - cursor++ lol
why did they release terminal based code editors?
anthropic was probably trying to squeeze the last bit of their codegen and sonnet branding. get some monies from token burning. it's also a way to showcase the possibility of end to end task completion - sonnet 3.7 eager for task completion perfectly fits into that mould. maybe they plan to release exclusive model support or innovation like mega-speculative-decoding for claude code
openai released codex cli recently as an open-source project. if you look at the implementation, it's not extraordinaire. they define functionality in via bash and then pass it to handleFunctionCall kinda functions... similar to swe-agent. the functionality in both of these have been there in open source since more than a year now. i agree that swe-agent is not user friendly but Aider chat (from polyglot benchmark people) is already a functional product with good ux. [all this said i am excited to see what all they improve in it]
as people have speculated on X, this was not really openai's priority, it was likely an attempt to dissolve anthropic's moat and steal some spotlight (commoditizing the complement). a code editor hasn't been their first priority but it's somewhere in the todo list given they have tried to acquire cursor twice and now in talks to acquire windsurf. why build your own when you can buy a heavily polished successful product and innovate on top of that...
own the infra, own the model, own the interface, own the experience
coming to the main point, OAI and Anthropic - if they spin up a code editor - they will own both the "model" layer and the "application" layer. this opens up a lot of space for pushing exclusive features and model support. model level UX unlocks feel way better - almost like magic and the dual control accelerates this. this is how cursor + sonnet got popular (2nd wave popularity).
think better codegen model support, autocompletion models, any sort of new innovation that's not been released in any other product. maybe some kind of rl finetuning for your codebase. i am pretty sure the lab people have lots of better and novel ideas than what i presented (overfitted to my exposure)
some examples from the past - cursor's fast apply cursor blog link fireworks post link - where they used some variant of speculative decoding. to make the inference super fast, they had partnered with fireworks AI for a custom model that had a high throughput. OAI/anthropic have these things natively... imagine what sort of autocomplete models oai people can cook.
but sankalp, they can provide all these features in terminal based code editors too? yes i guess. tbh i would need to spend more time looking at the implementations to answer this but i would argue that it would be easier to provide them in an application outside terminal i.e a code editor.
for mass adoption you need GUI. better multi-tasking, better visual affordances, feature discoverability, ux bandwidth etc. there's one area i can think of where claude code shines is the - give the agent a task and it will perform the tasks in the background like a toddler (the async autonomous agent paradigm with some human in the loop) - analogous to daemon processes; but this can be easily accommodated in terms of tabs in editors like cursor...
closing note: cursor people have gotten more capital recently and hired some super researcher-product people so i am excited to see what they offer in the future. they are basically a small codegen R & D lab atp. also excited to see what oai/anthropic cook.