OpenAI Astra 28 Days of Updates: The Day-by-Day Log

OpenAI has started a run of daily updates for GPT-6 Astra and GPT-6.1 Sol, and this page is the running log. Day 1 is a speed upgrade: the default generation speed is now roughly 50 percent faster, which the announcement puts at about 50 tokens per second instead of 30, across the ChatGPT subscription in every OpenAI product and in third-party tools that use Sign in With ChatGPT.

The pace matters as much as the change. On October 5, 2026, Codex lead Tibo Sottiaux promised 28 days of either a clear improvement or a full usage reset. This hub tracks the improvements, day by day, with the exact wording of what OpenAI said and a plain-language read on what it changes for people who build with these models. Days that have not happened yet are marked pending. We will not fill them with predictions.

The 28-day tracker

Dates assume Day 1 is October 5 and each day follows consecutively. If OpenAI skips a day or ships two items in one, we will correct the table rather than force the numbering.

Day 1: default speed is about 50 percent faster

Here is the announcement in full, as posted:

We have optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol through the subscription across all our products and partners using Sign in With ChatGPT (including OpenCode, Pi, Amp, Devin, ...). No changes needed on your end and this should be felt within the next two hours. Reaching 50 TPS instead of 30TPS, with also the most optimized tokenizer out there. And the models are quite the efficient ones in terms of number of tokens needed to get things done!

What changed, in one table

Is it 50 percent or 67 percent?

The numbers in the post describe the same improvement from two angles, and they do not match perfectly. Going from 30 to 50 tokens per second is a 67 percent increase in throughput. Measured the other way, the time to generate each token falls from about 33 milliseconds to 20, a 40 percent reduction. "Roughly 50 percent faster" sits between the two.

That is a normal marketing rounding, not an error worth arguing about, but it affects how you plan. If you run long agent loops and care about wall-clock time, think in terms of seconds saved on a long generation. A 4,000-token answer takes about 133 seconds at 30 TPS and about 80 seconds at 50 TPS. That is a real gain of nearly a minute, but it is not a halving of total task time, because tool calls, file reads and test runs do not speed up.

Treat TPS as a typical figure, not a guarantee. Speed varies with load, effort level, prompt length and whether the model is thinking before it answers. The way to verify the claim is to time a fixed prompt before and after on your own account.

Why the "default" part matters

On September 30, 2026, DevDay introduced Ultrafast as a paid speed tier, tied to Pro 500 and the top usage limits. That left an obvious worry: if the fast lane is paid, does the free lane get slower to make the upgrade attractive?

Day 1 points the opposite way. The standard path got faster for everyone, and Ultrafast keeps its own premium position above it. If you were on the standard tier and thought speed was the main reason to pay more, the gap just narrowed. Whether that changes your plan depends on how far Ultrafast pulls ahead of 50 TPS, and OpenAI has not published a comparable number for the paid tier in this announcement. We will add it here if it appears.

Why third-party tools are named

The most interesting line is the list of partners: OpenCode, Pi, Amp and Devin. These are not OpenAI products. They reach GPT-6 models because of Sign in With ChatGPT, the DevDay feature that lets Plus and Pro users spend included plan usage inside participating tools, with per-app caps.

By naming them, OpenAI is saying the speedup lives at the model-serving layer, not in its own apps. That is good news if you work in a non-OpenAI harness: you should not be second-class on speed. It also raises a quiet point about competition. A faster default on a flat subscription makes the subscription route more attractive than pay-per-token API keys for interactive coding, which is the use case where TPS is most visible.

What people are asking

Do I need to update my CLI or app?

No. The wording is explicit: no changes needed on your end. Speed was changed server-side. If you want to confirm, run the same prompt and note tokens per second on a long generation. Many harnesses show elapsed time and output tokens at the end of a turn.

Does faster mean the models got worse?

The announcement says "optimized the default speed" and does not mention a smaller model, a lower reasoning effort, or a quantized variant. That is an absence, not a denial. After the September launch, OpenAI published a post-mortem on an Astra quality problem, so it is reasonable to run your own quick regression set the first day or two. If you have evals, run them. If you do not, pick three tasks you know well and compare.

What is the "most optimized tokenizer" claim?

A tokenizer splits text into the units the model reads and writes. A more efficient tokenizer needs fewer tokens to represent the same text, which cuts both cost and time. That claim is stated without a benchmark in the post, so we cannot verify the "most optimized out there" superlative. What we can say is that token efficiency has been a recurring theme: Artificial Analysis found that GPT-6.1 Sol costs about 78 percent less per Intelligence Index task than Astra even when it uses somewhat more output tokens than its predecessor, because the quality gain offsets the count. Tokens needed per task and tokens per second multiply together, so speed and efficiency compound.

Does this change my usage limits?

The announcement says nothing about quotas. Faster generation means you burn through output tokens in less time, so if your limit is metered in tokens and you were rate-limited by time, you may hit the ceiling sooner within a session. Watch your usage meter. The wider context is that the Pro 200 plan drops to 10x Plus on October 30, and earlier this month limits were cut roughly fourfold for Astra.

Which should I use now, Astra or Sol?

Day 1 does not change the basic trade. Astra is the flagship on raw capability at $10 and $50 per million input and output tokens, while Sol is $2 and $10. The Astra versus Sol comparison and the GPT-6.1 Sol launch file cover when each makes sense. With both at the new default speed, the practical question is again quality per dollar, not which one is quicker.

How to measure the speedup yourself

You do not need special tooling. This is a simple method that works in any harness:

1. Pick a prompt that produces a long, deterministic-ish answer, such as "write a 1,500-word explanation of how a B-tree works."
2. Run it once and record the total time and the output token count if your tool shows it.
3. Divide tokens by seconds. Ignore the time before the first token, since that is thinking and queueing, not generation speed.
4. Repeat three times at different hours, because load changes the result.
5. Record the figures in a notebook with the date. Comparing against the same prompt on October 4 is ideal if you have the number.

output_tokens: 3,820
generation_seconds: 76
tps = 3820 / 76 = 50.3

If you consistently see roughly 30, the rollout may not have reached your account yet, or your harness may be using a different route. Check back after the two-hour window.

What this means for what you build

For anyone running agent loops, speed is not a vanity metric. Interactive coding sessions feel different above roughly 40 TPS: reviewing a diff as it streams becomes practical, and a wait after each instruction stops breaking concentration. Tools that wrap these models, including OpenCode and Devin, inherit that without shipping an update.

For anyone choosing a stack, the more important signal is direction. OpenAI is spending its daily slots on things users feel immediately, and Day 1 is exactly that. If the next 27 days follow this pattern, the quota-and-speed story that has dominated the last month may get quieter. If instead the days turn into full resets, as the pledge allows, the plan story will dominate. We will log which it is.

How this log works

Each new day gets an entry with the same structure: the verbatim announcement, a table of facts, one paragraph on what is verified versus claimed, and a note on whether action is needed. Resets are logged as resets, with whatever detail OpenAI gives. Corrections get an update line with the date rather than a silent edit.

If you want the daily entries as they land, the newsletter above is the lowest-friction way. You can also follow @explainx_ai on X.

Related reading on explainx.ai

• OpenAI promises Codex 28 days of daily improvements or full resets
• GPT-6 Astra vs GPT-6 Sol: price, benchmarks and when to use each
• GPT-6.1 Sol launch: pricing and benchmarks
• GPT-6.1 Sol cost efficiency versus Astra
• Sign in With ChatGPT at DevDay 2026
• Ultrafast and Pro 500 at DevDay 2026
• Pro 200 usage halved on October 30
• What is TPS? Tokens per second explained
• Claude Code vs Codex rate limits

Specs, speeds and plan details are accurate as of October 5, 2026 and are based on OpenAI's own announcement. Speeds vary with load and effort level. Check the official OpenAI documentation for current figures.