Dissecting GPT-Live-1: OpenAI's full-duplex voice model at $0.05/min, decoded from its API
rdesh26 · x · 2026-09-12
OpenAI released GPT-Live-1, the API version of the model powering voice in ChatGPT, Codex and Work, with Hatch, Yelp and Speak already in production. With no official architecture details, the author infers from the API, pricing, docs and third-party evals:
- Full duplex: a single model reasons over incoming and outgoing audio together, with no turn/audio-done events — the model decides when to speak
- Frontend/backend split: GPT-Live only handles the conversation; reasoning and tool calls are delegated to a backend model (OpenAI's or your own), likely with an interleaved multi-stream frontend design
- Pricing: $0.05/min, no per-token price; roughly $28/M tokens assuming 12.5Hz audio tokens and chunked prefill
- Eval discrepancies: τ-Voice scores vary wildly — 86.2% on OpenAI's blog, 81.7% on the leaderboard, 67.9% on Artificial Analysis — with backend model choice mattering a lot
- OpenAI's CRAWL/WALK/RUN voice-agent eval harness is open and worth reading
More from coding & agent
- AXL launches remote MCP server exposing 1,284 sales and marketing APIs to Claude and Codex — EdTech_Padre · 2026-10-27
- Every's 'frontier team' tracks new models and audits AI spend per person and task — every · 2026-10-27
- How the Claude Code Team Reviews Code When Claude Writes 90% of It — inductionheads · 2026-10-27
- OmarchyTV hits GitHub: Omarchy-style TV launcher with AI TUI and MCP — alexcovo_eth · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Reddit Thread: Keeping AI Agents Grounded in Current Project Information — Few-Ad-1358 · 2026-10-27