OpenAI / News

GPT-6.1 Sol vs GPT-6 Astra in Codex: we gave both the same three coding jobs

We gave GPT-6.1 Sol and GPT-6 Astra the same three coding jobs in Codex, twice each. Both passed the same secret tests. Astra used about six times as many credits.

Bar chart of average credits per try: fixing a bug 1.2 for GPT-6.1 Sol and 8.0 for GPT-6 Astra, building from a description 1.6 and 8.3, adding a feature 1.9 and 9.5
Chart by Not an AI App from our own clean test of 11 October 2026. We estimated the credits from the token counts Codex showed and the credit rates OpenAI publishes.

OpenAI says its new model GPT-6.1 Sol gives "near-Astra performance" for less money. We tested that in Codex, OpenAI's AI helper for programmers. In Codex you pay with credits: the units a subscription's allowance is counted in. GPT-6 Astra uses five times as many credits per token. Tokens are the small pieces of text an AI reads and writes. We gave both models the same three coding jobs in Codex, twice each. Then we checked the work with tests the models never saw. Both passed exactly the same tests. But Astra used about six times as many credits.

Correction, 11 October 2026: our first runs also loaded the site owner's own Codex settings: personal instructions, add-ons and extra tools. That added text to every step. We ran all twelve tries again with a clean setup. Both models again passed exactly the same secret tests, and Astra still used about six times as many credits. But both used far fewer: 9 credits for Sol instead of 19, and 52 for Astra instead of 118. Both were also faster. The numbers in this article are the new ones.

What are Sol and Astra?

Screenshot of OpenAI's model page for GPT-6.1 Sol: near-Astra performance for complex work at a lower cost, $2 input and $10 output per million tokens
OpenAI's page for developers about GPT-6.1 Sol. Screenshot taken on 4 October 2026, cropped.

OpenAI showed GPT-6.1 Sol at its conference for developers on 29 September 2026. Its model page calls it "near-Astra performance for complex work at a lower cost". GPT-6 Astra is, in OpenAI's words, "our most capable model for the most demanding work".

On paper the two are close. Both can read 1,050,000 tokens at once. Both can write up to 128,000 tokens in one answer. And both know the world up to the same date: 30 April 2026. The difference is the price. Through the API, the paid connection for developers, Sol costs $2 per million tokens in and $10 per million tokens out. Astra costs $10 and $50.

In Codex with a ChatGPT subscription, you do not pay per token. Instead, each model uses up your plan's allowance at its own speed. OpenAI counts that in credits. That is what we wanted to measure. What do you get, and how much does it use, when you pick one or the other for normal coding work?

How we tested

Diagram of our test in three steps: three small projects, Codex does each twice per model with the same settings, and afterwards we check the work with secret tests
Diagram by Not an AI App: how we tested.

We wrote three small JavaScript projects. Each came with a short note about the job, like you might give to a co-worker:

  • A. Fix a bug. A piece of code turns text like "1h30m" into seconds. It added up the units wrongly. The job: make it do what its comments promise, for every valid input.
  • B. Build something from a description. A one-page description of how to work out a Dutch invoice. It has VAT (sales tax), discounts, credit lines, an EU tax rule called "reverse charge" and rounding rules. The job: build it and add tests.
  • C. Add a feature across several files. Due dates for a small to-do program you type commands into. It needed input checks, sorting, an "overdue" mark, a new command and exit codes.

Each model did each job twice, each time in a clean copy of the project. We used the Codex program you type commands into (version 0.162.0 in our clean re-run) on Windows 11. Both models ran on the middle setting for how hard the AI thinks. That is the normal setting for Sol. In our first run the models took turns. In the clean re-run, Sol's six tries ran first and Astra's straight after, on the same afternoon.

After each try we added secret tests: tests we wrote ourselves and only ran after the AI had finished. There were 24 for job A, 15 for B and 5 for C. The last ones ran the whole program. The models never saw these tests. Before the first try, we ran every secret test on our own answers. They all passed.

Want to see the test? One secret test checks that "1h banana" is refused, because there is junk after the time. Another checks that an invoice line of 1.005 kilos is allowed. Our test page shows all three jobs, the secret tests and the exact words we sent. You can also download them and run them on any AI yourself.

Who did better? Neither: the same score twelve times out of twelve

Table of twelve tries: both models passed 23 of 24 secret tests on job A, 15 of 15 on job B and 5 of 5 on job C, with times between 76 and 131 seconds
All twelve tries from our clean test of 11 October 2026. We only added the secret tests after each try had finished.

Both models got the same score on every job, in both tries:

JobGPT-6.1 SolGPT-6 Astra
A. Fix a bug23 of 24, twice23 of 24, twice
B. Build from a description15 of 15, twice15 of 15, twice
C. Add a feature5 of 5, twice5 of 5, twice

One test in job A failed for both models, every time. That was our mistake. The test checked that "1h1h", the same unit twice, is refused. But our job note never said so. Both models sensibly read it as two hours. We count that test as unfair. So both models get full marks.

The answers themselves were very alike. In the invoice job, both models did not use normal decimal maths. Instead they worked with exact whole numbers. That is the safe way to handle money. Their code was built almost the same way, down to the small helper parts.

In our first run we saw two small differences. Astra added an extra test file in job A, which the job did not ask for. And the comment "into whole seconds" was read in two ways: some tries rounded "1.5s" up to 2 seconds, others dropped the half second. In the clean re-run neither happened. No model added an extra test file, and all four tries turned "1.5s" into 1 second. Our secret tests did not check this. Both readings fit the comment, so an unclear instruction can still get a different answer from one run to the next.

How much did it use? About six times the credits

Screenshot of OpenAI's Codex price page: GPT-6 Astra costs 250 credits per million input tokens and 1,250 per million output tokens, GPT-6.1 Sol 50 and 250
The credit rates on OpenAI's Codex price page. Screenshot taken on 4 October 2026, cropped.

Codex shows how many tokens each try used. We turned those counts into credits with the rates on OpenAI's Codex price page. Per million tokens, Astra costs 250 credits for text in and 1,250 for text out. Text it already read before and can reuse costs 25. For Sol the same rates are 50, 250 and 2.5.

JobSol, average per tryAstra, average per try
A. Fix a bug1.2 credits8.0 credits
B. Build from a description1.6 credits8.3 credits
C. Add a feature1.9 credits9.5 credits
All six tries9 credits52 credits

Most of the gap comes from the price per token. In the clean re-run, Astra read about 8% fewer tokens than Sol and wrote about 8% fewer. But more of what it read was new text, not reused text: about 62,000 new tokens against 53,000 for Sol. New text costs ten times as much as reused text. That turns five times the price into about six times the credits.

Time was not the difference either. Sol needed 610 seconds for its six tries, Astra 621. No single try took longer than 131 seconds.

In every try, most tokens in were text that Codex reads again at each step: its own instructions, the job and the files. Both models got exactly the same setup, so the comparison is fair. But your own numbers will depend on your setup. In our first run, our own Codex settings roughly doubled the credits. These are estimates. The usage page in Codex gives the official count.

What does that mean on Plus and Pro?

Screenshot of OpenAI's usage estimates: on Plus, 5 to 45 local messages per five hours with GPT-6 Astra and 15 to 160 with GPT-6.1 Sol; Pro has no five-hour limit
OpenAI's own estimates of how many local messages fit in five hours on Plus. Screenshot taken on 4 October 2026, cropped.

With a subscription, credits decide how long your allowance lasts. OpenAI's own estimates for the Plus plan show the gap clearly. In five hours you get about 5 to 45 local messages with Astra. With GPT-6.1 Sol you get 15 to 160. OpenAI says these are estimates, not fixed limits.

We ran our test on a ChatGPT Pro plan. The same page says "Pro plans currently have no five-hour limit". There may be weekly limits, though. Our twelve tries fitted without any limit message. On Plus, OpenAI estimates 5 to 45 Astra messages per five hours. So six tries like ours could already use a big part of that time window.

When the allowance runs out, Plus and Pro users can buy extra credits. Then Astra's five times higher price per token shows up directly in what you pay.

When could Astra still be worth it?

Screenshot of OpenAI's model page for GPT-6 Astra: our most capable model for the most demanding work, $10 input and $50 output per million tokens
OpenAI's page for developers about GPT-6 Astra. Screenshot taken on 4 October 2026, cropped.

Our jobs were small and clearly described. A lot of everyday coding work is like that. But OpenAI does not say that is where Astra is strongest. OpenAI's guide to models in Codex suggests Astra "for the hardest end-to-end work". It adds that Astra "is better at asking focused questions and incorporating your guidance". For long work that you repeat often, "when cost matters", it points to Sol.

Our test does not prove that wrong. But it cannot prove it right either. Three jobs with two tries each is a small test. Bigger projects, unclear jobs or work that mixes code and documents could show differences that ours did not.

What our result does suggest: for clear coding jobs of this size, start with Sol. Switch to Astra only when Sol gets stuck. That is a sensible default.

Who can use GPT-6.1 Sol in Codex?

Screenshot of OpenAI's Codex models page: the GPT-6.1 Sol rollout includes Plus, Pro, Business, Enterprise and Edu in Codex, not Free and Go
Who gets GPT-6.1 Sol in Codex, according to OpenAI. Screenshot taken on 4 October 2026, cropped.

OpenAI's Codex model page says the GPT-6.1 Sol rollout "includes Plus, Pro, Business, Enterprise, and Edu in Codex in the desktop app and CLI". The CLI is the Codex program you type commands into. In Enterprise and Edu accounts, an administrator has to switch it on first. Free and Go plans do not get it at the start.

In the Codex app and the command program, you pick the model for each job. In the command program you use the -m option, for example -m gpt-6.1-sol. That is also how we ran the test.

This follows our test of Claude Code mods. There, too, we only wrote about what we could run and measure ourselves.

Questions people ask

List of questions people ask Google about this topic, with the 4 questions we answer marked
"People also ask" questions from Google (US), 9 October 2026, via DataForSEO. Chart by Not an AI App.

Can I use GPT-6 Astra?

Yes, if you pay. Developers can use it through the API, the paid connection for developers. In Codex, it uses up your ChatGPT plan's allowance. For the Plus plan, OpenAI estimates about 5 to 45 local messages with Astra per five hours. We used it on a ChatGPT Pro plan.

How much is GPT-6 Astra?

Through the API, GPT-6 Astra costs $10 per million tokens in and $50 out. Tokens are the small pieces of text an AI reads and writes. That is five times the price of GPT-6.1 Sol. In Codex, Astra costs 250 credits per million tokens in and 1,250 out, says OpenAI's Codex price page.

Is GPT-6 Sol better than Astra?

Not better, but just as good in our test. We tested GPT-6.1 Sol against GPT-6 Astra on three coding jobs in Codex, and again with a clean setup on 11 October 2026. Both got the same score on every job, but Astra used about six times as many credits. OpenAI still suggests Astra for the hardest, longest work, which our small test does not measure.

Which GPT model is the strongest?

OpenAI calls GPT-6 Astra "our most capable model for the most demanding work". But in our test of three small coding jobs, GPT-6.1 Sol got exactly the same score for about a sixth of the credits. For clear jobs like these, start with Sol.

How we checked

Timeline: a first run on 4 October 2026 that also loaded the owner's Codex settings, and a clean re-run of all twelve tries on 11 October 2026.
The numbers in this article come from the clean run of 11 October 2026. Diagram by Not an AI App.

We first ran the test on 4 October 2026. That run also loaded the site owner's own Codex settings. On 11 October 2026 we ran all twelve tries again with a clean setup: the Codex command program 0.162.0 with only a ChatGPT login (the owner's Pro account) in an empty settings folder, and the add-ons that come with the account switched off. The numbers in this article come from that clean run. We used Windows 11 and the middle setting for how hard the AI thinks. The AI could only write files inside the test folder. Each of the twelve tries started from a clean copy of the project. We timed each try, took the token counts from Codex itself, and checked the work with the secret tests afterwards. We kept the full results, the job notes and the secret tests in our test folder.

The same day we read OpenAI's model pages for GPT-6.1 Sol and GPT-6 Astra, the API price list, and the Codex price and models pages. The screenshots are our own captures of those pages. We only cropped them. We made the charts and diagrams ourselves from our results file. None of the images is AI-generated.

This was a small test of two models on three jobs. It tells you something about clear, everyday coding work, not about every kind of project. The credits are our estimate from the published rates. OpenAI's usage page gives the official count.

Update, 19 October 2026: we added our own diagram to a section that did not have an image yet. It shows what this article already says. No facts changed.