A Barely-There "Improvement"—From 30 to 50 Tokens Per Second. Can You Really Call That an "Improvement"? (updated October 7, 2026)

Here's today's rumor!

OpenAI says their Day 1 "improvement" is simply that they made things faster.

Faster than they were slow, that is.

Shortly after 2:00 AM Japan Time on October 6, Tibo (Thibault Sottiaux) posted and announced: The default speed for GPT-6 Astra and GPT-6.1 Sol has been increased by approximately 50%. This applies to all products and partners used via ChatGPT sign-in (OpenCode, Pi, Amp, Devin, etc.). No changes are required on the user side, and the change should be noticeable within two hours (summary). Tibo added to this shortly after: The speed has gone from 30 tokens per second (TPS) to 50 tokens, noting that the tokenizer is the most optimized in the industry and the models require fewer tokens for the same tasks.

Okay, getting serious for a sec. The basis for this Day 1 shipment is just these two posts on X. As of the morning of October 6, the ChatGPT & Codex changelog for October 5 only shows a bug fix for Codex CLI 0.160.1; there is no mention of speed. As for resets, the latest check on the third-party Codex Reset Monitor was still on October 2 (21:18 UTC).

Okay, serious time's over.

4 Highlights

• "About 50%" and "from 30 to 50" don't match up. Going from 30 to 50 is an increase of about 67% in terms of amount per second. In terms of time per token, it's a 40% reduction. It's neither a "50% shorter" story nor a "50% increase" story. That word "about" is doing a lot of heavy lifting here, isn't it?
• Users don't have to do anything. While it reads like they changed the settings, OpenAI hasn't given a single word explaining why it was slow in the first place. On October 2, a few days ago, they apologized for the slowdown during the startup of GPT-6.1 Sol and stated it had returned to expected speeds (summary). Whether that "expected speed" was 30 TPS or this new 50 TPS isn't written in either post.
• The blank space in the promise finally got an answer. The "majority of the population" that was left blank in the first table turned out to be all products via ChatGPT sign-in in this Day 1 shipment. The time it arrived was about 21 hours after the promise.
• The reaction was fast. Lon Lundgren, whose profile claims an Amazon Aurora award, quoted it and wrote this about an hour later. Admitting that they had been running a subscription account at half speed and revealing that all it took to fix it was flipping a single switch—well, you'd have to be a pretty nasty person to do that (summary). This is Lon's interpretation; OpenAI hasn't admitted to it. Just in case.

Sorami's Take

"Day 1 counts as a 'clear improvement affecting the majority of users' as promised": 65% Confidence
Reasons: The scope is broad, there's a specific number attached, and it's in a form that can be felt within 2 hours. Weaknesses: The numbers are OpenAI's self-reported figures, I haven't found any independent measurements within the scope I could verify, and the numerical discrepancy mentioned above.

Theory: "They were intentionally throttling it before": 10% Confidence
Evidence: Only Lon's sarcasm, and OpenAI hasn't explained the reason. However, it's true that no one has explained "why it was 30."

"Day 2 will be handled as a shipment without issuing a reset": 60% Confidence
Reasons: Precedent of Day 1 being handled as a shipment, and the forecast from the Codex Reset Monitor. The site estimates a 38% chance of a reset within 24 hours and 64% within 48 hours (third-party estimate).

"Day 2 will call it an improvement if they say 'We made it 1% faster'": 3% Confidence
This is just a joke. If I get this right, I'll brag about it with everything I've got. If I'm wrong, I'll insist it was a joke from the start.

The current tally, now that Day 1 is over: 1 day shipment, 0 days reset, 0 unconfirmed. The only confirmation of the shipment is the primary source of Tibo's post; it hasn't appeared in the official change logs yet. I'll check the results here every time.