ChatGPT for Professionals (2026): Tested, Plans, Limits

What is ChatGPT, and which models does it offer?

ChatGPT is OpenAI’s AI assistant, but its Chat, Work, Codex and API model listings are not interchangeable. The official pages distinguish the October Chat models from earlier versions used elsewhere.

The verified lineup is:

• GPT-6 Sol (October, Chat) and GPT-6 Luna (October, Chat): released 2026-10-07. The system card announces GPT-6 across free and paid ChatGPT plans globally. We have not verified which model each plan gets by default.
• GPT-6.1 Sol: the balanced API model, also rolling out in ChatGPT Work and Codex to Plus, Pro, Business, Enterprise and Edu. It is not in Chat.
• GPT-6 Sol: the September version, listed for Work and Codex rather than the October Chat experience.
• GPT-6 Luna: the fast API model and September version. Free and Go get it in the desktop app’s Work/Codex.
• GPT-5.6 Terra: listed for Work, Codex and API; it remains available during the rollout.
• GPT-6 Astra: the API flagship and the model evaluated in our test round.

GPT-5.5 was retired from ChatGPT-signed-in Codex/Work on 2026-10-14; its API availability was unaffected. GPT-5.6 Sol and GPT-5.6 Luna are older versions, not the current Chat lineup.

The practical point: check which part of the product a model listing refers to before using it to choose a subscription.

Which ChatGPT plan fits your work?

Start with Free to assess usefulness on tasks that contain no confidential information. Our verified facts do not establish enough differences in features or usage limits to recommend a paid tier simply because you work alone, use it daily or manage a team.

Prices checked 2026-10-19 on the official pricing page.

The Chat rollout does not establish a default-model split between plans. These are USD list prices shown to a US visitor, excluding tax. Readers in the UK, Australia and India should check the vendor’s pricing page for their country because local prices, currency and tax differ.

For a solo practice, compare Go and Plus against a concrete task rather than assuming the more expensive plan is necessary. For Pro, establish why you need the selected tier before committing. For a firm considering Business or Enterprise, confirm the seat details, price and data terms; we have not verified the differences between the Business seats.

How did ChatGPT do in our tests?

GPT-6 Astra performed strongly on our rubric, but this was an API test, not an evaluation of a ChatGPT subscription. The other models listed above are not in our test round yet.

On 2026-10-19, GPT-6 Astra passed 99% of rubric criteria on average; 114 of 120 runs passed every criterion. Judge-asymmetry caveat: claude-opus-5-5 graded these runs, and the judges are not equally strict. On the same 30 sampled GPT runs, claude-opus-5-5 and gpt-6-astra each passed 97% of criteria. They agree closely on GPT runs, but the Claude-versus-GPT gap is partly the grader; these scores do not prove that GPT-6 Astra is better overall than Claude or Gemini, which were graded by a different judge than GPT.

The round used 40 prompts, comprising 30 English and 10 Spanish prompts across 10 professions, with 5 rubric criteria per prompt and 3 runs per model per prompt. All 120 GPT-6 Astra runs completed and were graded. An editor re-read 20 verdicts and agreed with 18.

In a separate Excel / Microsoft 365 task round on 2026-10-20 (10 new prompts, 3 runs each, never pooled with the round above), GPT-6 Astra passed 88% of rubric criteria, and 15 of 30 runs passed every criterion. In that round’s judge swap on 6 sampled runs, claude-opus-5-5 passed 87% of criteria and gpt-6-astra passed 90%. These were text-only answers: formulas were judged by their logic and the stated result on the given data, not executed in Excel, and the models were chat models answering a pasted task, not Copilot in Excel or Microsoft 365.

Spanish round 2, dated 2026-10-21, used a separate set of 10 new Spanish prompts for engineers, travel agents, construction managers, event planners and interior designers; its results are not pooled with rounds 1 and 2 or the Excel round. GPT-6 Astra passed 98% of rubric criteria, and 27 of 30 runs passed every criterion. Because claude-opus-5-5 graded GPT-6 Astra while gpt-6-astra graded the other models, this figure is not like-for-like with theirs; in that round’s judge swap on 6 sampled GPT runs, claude-opus-5-5 and gpt-6-astra each passed 93% of criteria. An editor re-read 12 verdicts and agreed with 8.

We used a command-line API harness with tools, web and memory switched off. Consumer-app results may differ, and the findings should not be treated as scores for the Chat models or any subscription tier.

Most lost points came from safety criteria involving a warning, date, benefit or legal assumption added without support in the input. When reviewing a draft, look for those additions rather than checking only grammar and tone. See our methodology for the full test design and grading limitations.

What should you check before pasting client or patient data?

We have not verified the vendor’s data terms for this page. We make no claims about retention, training on your data or Business and Enterprise data protections.

Read the data and privacy terms for the exact plan you intend to use, alongside your professional confidentiality rules. Check any employer requirements and client agreements before submitting client, patient or employee information.

While those questions remain unresolved, use fictional examples or material you are authorized to share. Do not assume that a paid plan, a business-oriented name or removing a person’s name settles the confidentiality question.

Confidentiality duties also depend on your profession: our pages for lawyers, healthcare professionals and accountants list the rules that apply in your country.

When should you not rely on ChatGPT?

Do not treat its output as the final authority when an error could affect a person’s rights, finances, health or safety. Tax, legal, health and property work must be checked by a licensed professional in your jurisdiction.

Avoid relying on an answer when you cannot verify its factual claims or assumptions. Our test findings make that review particularly important: a plausible addition can still be unsupported by the source material.

For current facts, check a primary source yourself. Our tests had web access off and do not establish how the app handles live research.

For drafting, keep the source material available during review. Compare names, dates, obligations and recommendations against it before sending the result to a client or colleague.

For how the models did on real professional tasks, read our test-round report; ready-made task instructions for Claude are in the Claude Skills library.

FAQ