Startup Playbook: success through failure

GPT-6.2 Sol for Small Business: What OpenAI's Cheaper Model Costs, What Testers Said and How to Test It Before You Switch

TL;DR: Should you switch to GPT-6.2 Sol?

OpenAI released GPT-6.2 Sol, its cheaper everyday model, together with GPT-6.2 Astra on 17 November 2026. Sol costs $2 per million input tokens and $10 per million output tokens in the API and runs in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Early testers were harsh: Matt Shumer called the new models "essentially unusable", and Sam Altman admitted OpenAI has "more work ahead of us on making the models nice to work with". Our advice: test Sol on five tasks you repeat every week before you move a single workflow.
GPT-6.2 Sol is the half of OpenAI's new release that most small teams will meet first, and the early reviews of GPT-6.2 Astra and Sol have been rough. This guide is for women founders, freelancers and small teams on ChatGPT Plus or the OpenAI API: what Sol is, what testers said, and how to test it on your own work.

What is GPT-6.2 Sol?

OpenAI's announcement calls Sol "our efficient, versatile model for everyday work", as quoted by Simon Willison, and Astra its "new frontier model for your hardest projects". In its launch post on the developer forum, OpenAI adds that "Many OpenAI employees use GPT-6.2 Sol as their go-to model across technical and non-technical work."
The price is the headline. In the API, Sol costs $2 per million input tokens and $10 per million output tokens, a quarter of Astra's $8 and $40. Our MeanCEO report on the new prices notes that Sol stays at $2 and $10 while Astra got 20% cheaper. Pro 200 subscribers also get Ultrafast mode with Sol, which 9to5Mac says "can generate tokens at up to 4.5 times the standard rate".

Where can you use GPT-6.2 Sol?

The forum post says: "Starting today, GPT-6.2 Astra and Sol are available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users."
  • Plus, Pro, Enterprise and Edu: Sol in ChatGPT Work and Codex; Pro 200 and Pro 500 also get Ultrafast.
  • OpenAI API: gpt-6.2-sol, with Ultrafast available.
  • Regular Chat: a forum user quotes OpenAI's Help Center: "This rollout does not add either model to regular Chat." We could not open that page to confirm it.
One forum user wrote that the picker now "lists 6.1 as a legacy model"; we could not confirm it, and the API deprecations page lists no retirement tied to this launch. If you set up GPT-6.1 Terra for admin work, keep those prompts: they become your test set.

What does OpenAI claim Sol does better?

  • Work tasks: +5.3 points over its predecessor on GDPval v2, an internal OpenAI benchmark, per OpenAI's numbers in VentureBeat's table. VentureBeat also counts Sol ahead of GPT-6.1 Sol on eight of nine launch charts.
  • Code: "On FrontierCode 2.0, Sol nearly matches Astra on merge-ready changes to real codebases at about one-fifth the cost."
  • Refusals: the GPT-6.2 System Card says both models "refused fewer harmless requests than their predecessors" and were less resistant to jailbreaks; OpenAI then adjusted the refusal boundaries.

What did testers say about GPT-6.2 Sol and Astra?

The first reviews were harsh, and VentureBeat put it in the headline: "'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users".
  • Matt Shumer (investor and former HyperWrite AI CEO, per VentureBeat), on Astra and Sol: "I had a terrible experience testing them. They were essentially unusable. I do not recommend these models."
  • Theo Browne (t3.gg), on Astra: "I think this is the worst model OpenAI has ever released." Brave New Coin quotes his longer letter: "It runs in loops and goes nowhere".
  • Ethan Mollick, on Astra: "I found it very uneven & weird." Harder to steer, he wrote, yet strong on an ambitious Blender and Unity project.
  • Dan Shipper (Every): "Overall a disappointing release compared to Fable 5.2."
Most detailed complaints target Astra, and the remarks about Sol lean positive. Peter Gostev: "GPT-6.2 Sol on Ultrafast is my favourite release. Sol is super capable, and for day to day the speed is genuinely incredible." Derya Unutmaz answered Shumer: "My early testing experience is exactly opposite of yours! They are amazing models!"

How does Sol score in independent tests?

  • Artificial Analysis Intelligence Index: Sol 57 at max effort; Astra 62, Claude Fable 5.2 65, Gemini 4 Hydrogen 54.
  • Coding Agent Index: "GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol", says Artificial Analysis.
  • Vals AI: Astra and Sol rank #7 and #9, 1.6 points apart, and Sol costs less than a third as much per task.
  • Made-up answers: The Decoder reports Artificial Analysis hallucination rates of 55% for Sol and 59% for Astra.

How did Sam Altman respond?

Altman answered on X the same day: "Although the new models are very smart, we have more work ahead of us on making the models nice to work with, and will prioritize that for upcoming releases. But even if it takes a little extra work, the intelligence and capability are extraordinary." Earlier he singled out Sol: "6.2 Sol ultrafast is something I think people will really enjoy." Small teams feel that kind of problem first, because nobody has time to babysit a model.
The counter-evidence deserves space too. ARC Prize says "GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3." Artificial Analysis calls Astra the "#2 model" on its index, and Epoch AI gave Astra and Claude Fable 5.2 the same score of 174, according to VentureBeat and The Decoder. For Sol, the strongest argument is cost per task.

How do you test GPT-6.2 Sol on your recurring tasks?

Benchmarks measure someone else's work. Your client emails, proposals and spreadsheets are the test that counts, and a week of side-by-side runs costs very little.
Your taskTry Sol?What to check
Client emails and repliesYes, a good first testTone, names, dates and any promise you did not make
Proposals and quotesYes, side by side with your current modelPrices, scope and terms against your own rate card
Spreadsheets and formulasYes, on a copy of the fileRecalculate three cells by hand and check formula ranges
Social posts and blog draftsYesYour style rules, facts and every link
Code fixes in CodexOn a separate branchRun your tests: Sol lost 3 points on a coding agent index
Long multi-step jobsCompare with your current model firstEarly stops, loops and skipped obstacles
Legal, tax or medical factsOnly with sources you checkEvery claim, against the official source
Score each answer on two things: how many corrections you made and how long they took. OpenAI quotes early testers who saw "fewer rounds of revision" in code; your count shows whether that holds for your work.

How do you keep your prompts portable?

A portable prompt works in Sol, your previous model and Claude with no rewrite, so a disappointing release costs you minutes.
  • Write the context in full: who you are, who reads the output, the format and one sample of good work.
  • Keep prompts in one shared doc, noting which model gave the best answer for each, and skip magic phrases one model happened to like.
  • Record which model is approved for which task in your AI usage policy.
  • Start from a prompt library: our 30 days of ChatGPT prompts for founders make a ready test set.

When should you stay on the previous model or use Claude instead?

  • A workflow already works. If your current model handles client work with few corrections, Sol has to prove itself first.
  • Coding agents. Keep your current setup until your own tests pass, given Sol's 3-point drop on the Coding Agent Index.
  • Long unsupervised jobs. Shipper said Astra "seemed to stop too early frequently", and the system card says Astra "omitted a known obstacle to completing its work in 22.28 percent of trials". OpenAI's fix is a Codex prompt that "is not applied to direct API requests", so in your own API calls ask the model to list anything that blocked it.
  • Tasks where Claude wins your test. Claude Fable 5.2 scored 65 on the Artificial Analysis index at $8 and $40 per million tokens, four times Sol's price.
Testing through the API? Keep keys out of shared prompt docs and browser code: stolen AI API keys are a target for automated agents.

What this means for you

  1. Wait a week before you switch and give Sol five tasks you repeat every week.
  2. Save a baseline of your current model's answers so the comparison is fair.
  3. Check every output for names, numbers and prices.
  4. Count the edits. The model that needs fewer corrections is the cheaper one, whatever the token price says.
  5. Keep prompts portable so you can move between Sol, the previous model and Claude in minutes.
  6. Switch one workflow at a time, starting with low-stakes drafts.
Build your first startup with an AI co-founder

The Fe/male Switch startup game takes you from idea to first paying customers in weekly quests, with an AI co-founder and a community of founders.

👉 Start the game

FAQ: GPT-6.2 Sol

What is GPT-6.2 Sol?

OpenAI's cheaper everyday model, released on 17 November 2026 with GPT-6.2 Astra. It costs $2 per million input tokens and $10 per million output tokens in the API.

Is GPT-6.2 Sol available in regular ChatGPT?

OpenAI says Sol is in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. A forum user quotes the Help Center saying the rollout does not add either model to regular Chat.

How much does GPT-6.2 Sol cost?

In the API, $2 and $10 per million input and output tokens, a quarter of Astra's $8 and $40. In ChatGPT it comes with Plus, Pro, Enterprise and Edu plans.

Is GPT-6.2 Sol better than GPT-6.1 Sol?

On OpenAI's numbers, yes: +5.3 points on GDPval v2. Artificial Analysis found Sol 3 points lower on its Coding Agent Index, so test coding work separately.

Why are the GPT-6.2 reviews so negative?

Testers such as Matt Shumer, Theo Browne, Ethan Mollick and Dan Shipper found the models hard to steer or disappointing next to Claude Fable 5.2. Most complaints target Astra, and Sam Altman promised to make the models nicer to work with.

Should I use Claude instead of GPT-6.2 Sol?

Use Claude where it wins your own side-by-side test. Claude Fable 5.2 scored 65 on the Artificial Analysis Intelligence Index against 57 for Sol, at four times Sol's API price.

More on this story

About the Author

Violetta Bonenkamp, also known as MeanCEO, is an experienced startup founder with an impressive educational background including an MBA and four other higher education degrees. She has over 20 years of work experience across multiple countries, including 5 years as a solopreneur and serial entrepreneur. Throughout her startup experience she has applied for multiple startup grants at the EU level, in the Netherlands and Malta, and her startups received quite a few of those. She’s been living, studying and working in many countries around the globe and her extensive multicultural experience has influenced her immensely.
2026-11-18 09:30 news