OpenAI kicks off Codex 28-day streak by speeding up GPT-6 models about 50%

The story began with a pledge from Codex lead Tibo on X: for the next 28 days, his team would either ship an improvement to Codex/Work every day or give users a full quota reset.

October 6 marks day one of that "crazy 28 days" run. Tibo's first deliverable was a speed optimization: default inference speed for GPT-6 Astra and GPT-6.1 Sol rose by about 50%, to 50 TPS, both through official subscriptions and through third-party tools that use Sign in with ChatGPT, such as OpenCode, Pi, Amp and Devin.

The focus on speed traces back to a recent stumble. When GPT-6.1 Sol launched on September 29, it ran noticeably slow even in Fast mode, prompting users to joke that it should be renamed "GPT-6.1 Turtle."

Tibo apologized, saying GPT-6.1 Sol had hit a surge in load that slowed it down, and promised to restore expected speeds and arrange quota resets for paid accounts. He then canvassed users and narrowed the roadmap to four areas: simplifying the product, improving efficiency for more usage, breakthrough features, and new models.

The first day did not land well. One user reported that speeds only reached about 35 TPS after four hours, below the advertised "50 TPS within two hours," while others argued the fix was merely catching up to a problem caused by the earlier load spike, and had not yet returned to pre-launch speeds.

OpenAI drew separate criticism the same day on two fronts: testing a new visual ad format inside image generation, and adding invisible statistical watermarks to text generated by ChatGPT and Codex in the EU over the coming weeks to meet AI Act requirements.

A report from independent research firm SemiAnalysis then put subscription economics under the spotlight. It argued that monthly subscriptions are heavily subsidized products, citing Anthropic as an example: subscriptions account for only about 10% of its total revenue but consume more than 40% of its inference compute.

Under cost pressure, the two labs diverged. Anthropic opted to quietly control costs by reducing the API-equivalent value of high-end models such as Fable 5.1, while OpenAI cut the API-equivalent value of its $200 Pro plan by roughly half and introduced a new $500 tier. Per SemiAnalysis, the new $500 plan's GPT-6 Astra allowance is only about 21% higher than the old $200 plan.

The result was a dramatic first day: an official speed upgrade met with a comment section full of complaints. Whether OpenAI's 28-day marathon can run its course will depend on balancing speed, quota and profitability.