A day after promising 28 days of improvements, OpenAI shipped what looks like the first one. On October 5, 2026, Codex and ChatGPT lead Tibo Sottiaux wrote that OpenAI had "optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol through the subscription across all our products." Press coverage labelled it day one of the sprint we covered in OpenAI's 28-day pledge.
This post covers what the number means, what it does not cover, and a quick way to check whether it improves your real work. We keep a skeptical eye on round numbers, so there is also a short section on the arithmetic.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| What changed? | Default generation speed for GPT-6 Astra and GPT-6.1 Sol is about 50% faster. |
| Numbers? | Reported roughly 30 to 50 tokens per second. |
| Who gets it? | Subscription products and apps using Sign in with ChatGPT. |
| API? | No. Not an API change. |
| Price change? | None announced. |
| Rollout? | Reported within about two hours of the post. |
| Is it the sprint's day one? | Appears so, per press coverage. |
The numbers, and a note on arithmetic
The reported shift is from about 30 tokens per second to 50. That is a 67 percent increase in speed, not 50. There are two charitable readings. "About 50 percent faster" may be a rounded, conservative description of an improvement that varies by load. Or it may be framed in time: 50 tokens per second finishes a fixed output in 60 percent of the time, a 40 percent reduction. Neither is wrong, but they are different numbers, and the 30 and 50 figures come from press reporting rather than a table published by OpenAI. Treat both as approximate.
What matters is that streaming at 30 tokens per second feels slow for long outputs, and 50 feels noticeably better. A 2,000-token answer takes about 67 seconds at 30 and 40 seconds at 50, a difference you can feel on every long response.
Who is affected
The change applies "through the subscription across all our products," and to third-party apps that use Sign in with ChatGPT. Coverage names OpenCode, Pi, Amp and Devin as examples. In those apps you authenticate with your ChatGPT account instead of an API key, so the speed change reaches you without any setting.
It is not an API change. If you call GPT-6 Astra or Sol with an API key, this announcement does not say your speed changed. OpenAI also did not announce a price change. For API pricing and the Sol versus Astra trade-off, see our GPT-6.1 Sol launch coverage.
This is separate from Ultrafast, the paid speed tier announced at DevDay with far higher throughput on top-priced plans, covered in our Ultrafast and Pro 500 post. The default-speed boost raises the floor for everyone, while Ultrafast is the premium ceiling.
Why OpenAI would do this now
Three pressures line up, all documented in our recent coverage.
- Launch-week load. After GPT-6.1 Sol's launch, Tibo apologized for slow speeds and said capacity was being added, as described in our Dots, Pro 200 and reset breakdown.
- Quota anxiety. A faster default is a cheap way to make plans feel better without changing limits, in a month when Pro 200 limits are being halved.
- Competition. Developers have been comparing Codex with Claude Code, and speed is one of the easiest dimensions to improve and to demonstrate.
None of that is unusual. What is notable is that the pledge was specific about daily deliverables, and the first one was an infrastructure improvement rather than a new feature.
Does faster streaming mean faster work?
Only partly. For agent tasks, wall-clock time is the sum of several parts:
| Component | Helped by faster streaming? |
|---|---|
| Model writes code and explanations | Yes |
| Tool calls and shell commands | No |
| Running tests and builds | No |
| Network and API waits | No |
| Queueing and rate limits | No |
| Long hidden reasoning before output | Depends on whether reasoning tokens stream at the same speed |
An agent that spends most of its time running a slow test suite will barely notice a faster stream. An agent that writes large files or long reviews will notice more. This is why the RuntimeWire write-up questioned whether streaming speed translates into faster task completion. That is the right question, and only your own workload answers it.
How to measure it in ten minutes
- Pick three representative tasks, for example a refactor, a bug fix with tests, and a long explanation.
- Time each from prompt to done, and separately note the time the first token appears.
- Record output length so you can compute tokens per second yourself. Many apps show usage after each turn.
- Repeat at a different time of day. Speed varies with load, and a single measurement misleads.
- Compare with your notes from last week, if you have them, and with another model on the same tasks.
- Keep a small log across the 28 days. It lets you separate real gains from placebo.
What this tells us about the sprint
Day one delivered something measurable, which is a good sign for the commitment. It also illustrates the pledge's structure. An "improvement relevant for most users" can be a quiet infrastructure change like this, while a "reset" is a quota refill. Both count. A speed bump does not raise your quota, change your plan or fix quality problems, so it should be weighed alongside the other things OpenAI promised to work on: simplifications, efficiency for more usage, new features and new models.
If you want to track the whole sprint, the useful approach is a simple table of date, what shipped, whether it reset usage, and whether it changed your week. We included a template in the pledge post. A single speed increase is nice. Whether the next 27 days add up to something is the real test, and the Oct 30 Pro 200 cut lands inside the window.
What to watch in the next 27 days
A speed bump is a good first deliverable, because it is measurable and affects everyone. The more informative items will be the ones that change what you can do. Here is a short watch list, in the order we would care about it.
- A durable usage gain. Anything that raises how much work a fixed plan completes, not just a refill. Efficiency improvements are the only sprint item that survives day 28.
- Reliability. Fewer failed runs and fewer stalls during launch weeks. Speed is wasted if sessions error out, as they did during the September outage.
- Simplification. Tibo said the team is focused on simplifying. Fewer overlapping modes, clearer meters and fewer choices between Astra, Sol and speed tiers would be real progress.
- A model or tier milestone. The "6.1 coming soon" post, which we analyzed earlier, pointed toward a Sol speed tier. If a faster tier arrives, compare it against the new default before paying for it.
- Clarity on the Oct 30 change. Whether any of the daily improvements soften the Pro 200 reduction, or only cushion it.
Keep your own log. A simple table with the date, what shipped, a speed measurement and a note on whether it changed your week will make the summary at day 28 an evidence-based one rather than a mood.
A fair way to compare Codex speed with other tools
If you are comparing Codex with Claude Code or another harness, speed comparisons are easy to get wrong. Use the same task, the same repository state and the same starting conditions. Measure time to first token separately from total time. Run each at least three times at different hours. And measure time to a correct result, not time to a finished-looking one, because a fast wrong answer costs a retry. Streaming speed is one input to that, and for many tasks not the largest.
What this means for what you build or pay
If you use Codex or a Sign in with ChatGPT harness for long outputs, you probably benefit today at no cost. If you build on the API, nothing changed, so do not assume the same speed there. If you were considering Ultrafast mainly for responsiveness, test the new default first, because it may be good enough for your use. And if you are comparing vendors, add streaming speed as one column among cost per solved task, quality and limits, rather than letting a single number decide.
Related reading
- OpenAI's 28-day pledge: daily improvement or reset
- Dots shipped, Pro 200 halved, Sol got a reset
- GPT-6.1 Sol launch, pricing and benchmarks
- Ultrafast and Pro 500 at DevDay
- What Tibo meant by "6.1 coming soon"
- Codex Auto-review is free
- OpenCode: open-source coding agent guide
Primary: Tibo Sottiaux's post on X, October 5, 2026 · RuntimeWire and other press coverage of the speed change
Details are accurate as of October 6, 2026. The 30 and 50 tokens-per-second figures are from press reporting, not an OpenAI benchmark, and speeds vary with load. Verify on your own tasks.