OpenAI opens its 28-day Codex pledge by speeding GPT-6 Astra and Sol to about 50 tokens a second

Tokens per second is a streaming rate. A coding agent also plans, calls tools, runs code and checks its work, and output speed does not measure that full cycle, as RuntimeWire pointed out ^[10]. At the quoted rates, a fixed block of output now streams in about 60% of the time it used to ^[22]. That cut applies only to the generation share of each loop. A step that waits on a slow test suite gains little.

For the 50 tokens a second to transfer to a given team, the agent's loop has to be bound by generation, and the rate has to hold on that team's product, plan and task mix. Sottiaux called both rates approximate ^[5]. He also said the models use a highly optimized tokenizer and need fewer tokens per task ^[6]. OpenAI did not say how it measured the rates, whether they hold across products, plans and tasks, how many tokens the tokenizer saves, or whether it changed usage limits, the models or per-request compute ^[5][6][7]. The announcement undersells its own figures. The "about 50%" OpenAI put on the change ^[1] is smaller than the roughly 67% that 30 to 50 implies ^[21].

The delivery path is good engineering. Sign in with ChatGPT lets an eligible user run a participating app on their ChatGPT plan, with usage counted against the plan and no API key ^[11]. The speedup shipped as a default, so it reached those apps with no setting to change ^[4]. Sottiaux said users should feel it within two hours of his October 5 thread on X ^[9]. The allowance is shared the same way. Users can set weekly limits per app, but an app's limit does not create a separate pool of usage ^[12].

Resets work as the penalty because OpenAI has made them routine. Sottiaux reset Codex limits for all paid plans in April. He reset usage twice in one day in July after multi-agent regressions, and again when Codex and ChatGPT Work passed 8 million active users. He and Sam Altman handed out a banked reset on stage at DevDay on September 29 ^[17]. An unofficial tracker had a 28-day scoreboard up within hours of the pledge ^[18]. One developer quoted by The New Stack suggested burning the allowance every day and hoping OpenAI misses a ship ^[19]. Mark Kretschmann, a software engineer at ASYS Group, read the pledge as 28 Codex resets in 28 days ^[20]. "Let the improvements begin," Sottiaux wrote ^[16].

OpenAI announced the Pro cut alongside a new $500 plan. That plan offers 25 times the Plus allowance and an Ultrafast tier for GPT-6 Astra at up to eight times standard speed in Codex ^[15]. I'd expect that tier to matter more to latency-sensitive teams than day one did, if the eight times holds on their workloads. For a team choosing among agents on a ChatGPT plan before the window closes, I'd measure task completion time on its own repositories, and how fast each tool drains the shared allowance.