Amp wants an agent for everything, not just code

OpenAI says startup Amp is pushing its frontier coding agent beyond code into writing, design and running other agents.

An AI agent is software that takes a goal and acts across tools to complete it. Amp built one to turn ideas into working software, and it now routes most of its token volume through OpenAI models, according to OpenAI. Amp CEO Quinn Slack, who leads the 20-person team, said OpenAI models have improved consistently and that safeguards are transparent to users.

That pitch comes with a cost problem.

Slack said some Amp users consume 10 billion tokens a week, and that watching every increment holds back experimentation. Amp launched Sign in with ChatGPT, which is an auth method that lets Amp users apply their ChatGPT plan to access more tokens and models inside Amp including GPT-6 Astra. Slack said the integration makes users more willing to give the agent a longer leash on larger tasks, in his words, to stop counting every token and see how much it can do as it helps with planning before building and work after shipping.

Writing and design are the test. Amp said its own team has no dedicated marketing hire but has an agent devoted to that work to improve new-user onboarding and discovery through search. For that, it uses GPT-6 Astra, which is OpenAI's model tuned for writing and design quality, to produce docs and pages in Amp's exacting house style. Slack said the team has been happy with the fit and pointed to a home experiment where he fed GPT-6 Astra his house plans to build a 3D model in Blender, then turned it into games with his children.

Puck is the coordination layer. It is a conversational assistant inside Amp that can start an agent to investigate a bug, orchestrate several agents, or retrieve a past technical conversation. Calling Puck, powered by GPT-Live-1, which is OpenAI's real-time voice model for talking to an agent, adds spoken control including from a watch, which Slack framed as a way to keep agents working while away from a computer, even on a run.

Speed is the other lever. Ultrafast mode is a service tier that trades cost for latency, and per OpenAI's developer documentation it is the fastest tier in the OpenAI API and broadly available for GPT-6 Astra and GPT-6.1 Sol. An August 13 preview noted it can run at up to 750 output tokens per second and up to 14x Standard processing speed on Cerebras infrastructure for GPT-5.6 Sol. Slack said Amp has been testing Ultrafast mode with GPT-6 Astra and that greater speed changes behavior, since users currently juggle several agents while waiting for completions and faster returns could let them run fewer in parallel or let agents self-check before replying. Slack said people have not seen a model this good that is this fast.

The skepticism is what is still gated. OpenAI disclosed no price for Sign in with ChatGPT inside Amp, no rate limits or quotas, and no benchmarks for GPT-6 Astra on writing or design to support the style claims. Ultrafast availability beyond the named models and its cost curve were also not detailed in the Amp post. For a 10-billion-token-a-week user, those missing numbers matter more than a demo of a lazy river in a house model.

Amp said the next step is to pull more pre-code and post-code work into the agent so builders can carry an idea further without handling every step. Whether users actually hand over that work will depend on verifiable quality, cost and control, not just speed and voice.