Claude vs ChatGPT vs Gemini, Tested: Which Fits Your Work

What do Claude, ChatGPT and Gemini cost per month?

Claude Pro, ChatGPT Plus and Google AI Pro have similar monthly prices, but ChatGPT Go and Google AI Plus are cheaper paid options. Compare the included models and billing terms before choosing.

Prices and plans checked 2026-10-19; USD list prices for US visitors, excluding tax.

For readers in the UK, Australia and India, local prices, taxes and rules differ: check them locally, including the vendor’s country-specific pricing page.

Claude’s annual Pro price is shown as $17 per month with $200 payable up front; use the stated upfront amount when budgeting. Team seat prices and Enterprise usage charges also need separate attention. A seat price alone does not establish the total cost.

Our guides to Claude plans for professionals and ChatGPT plans for professionals provide further context for those subscriptions.

Which models are in each lineup?

The current lineup differs from the models in our October 2026 round, and ChatGPT distinguishes Chat from Work and Codex. A subscription name alone does not establish which model you will use.

Model lineup and API prices checked 2026-10-19. API prices are USD per MTok, input/output; “not recorded” means no verified price here.

OpenAI’s 2026-10-07 system card says GPT-6 is launching across free and paid ChatGPT plans globally. Which plan gets which Chat model by default is unverified. GPT-5.5 was retired from ChatGPT-signed-in Codex and Work on 2026-10-14; its API was unaffected.

How did Claude and ChatGPT perform in our October 2026 round?

In our October 2026 round, added content was the main source of lost criteria; word limits also cost points for the Claude models. These findings concern the evaluated outputs, not every professional task.

Round 1 ran on 2026-10-19: 40 prompts, including 30 English and 10 Spanish, across 10 professions. Each prompt had 5 pre-written rubric criteria and 3 runs per model, totaling 360 runs.

Claude and GPT scores below came from a command-line API harness with tools, web and memory off, while Gemini 4 Argon ran through Google’s Antigravity CLI with a tool-less agent, not consumer apps; app results may differ. Claude Fable 5.1 and Gemini 4 Argon were added on 2026-10-20 on all 80 profession prompts with the same prompts and rubrics. Judges were asymmetric: GPT-6 Astra graded Claude outputs, while Claude Opus 5.5 graded GPT outputs. Their strictness differed, so part of the Claude-versus-GPT gap comes from the grader.

Results dated 2026-10-19 (Claude Fable 5.1 and Gemini 4 Argon: 2026-10-20), English prompts only; read with the harness and judge-asymmetry caveats above.

Calibration in the October 2026 API round illustrates the judge effect. On 30 sampled Claude runs, the published GPT judge passed 93% of criteria and the replacement Claude judge passed 98%. On 30 sampled GPT runs, both passed 97%. An editor agent, not a human, independently reread 20 verdicts before seeing the judge’s and agreed with 18.

Opus, Sonnet, Fable and Gemini 4 Argon share the GPT-6 Astra judge, making those comparisons like-for-like. These scores do not establish an overall winner; the older Sonnet 5 result must not be attributed to Sonnet 5.5.

The practical lesson from October 2026 is to check for unsupported warnings, dates, benefits and legal assumptions. A polished answer can still add something absent from the source.

Which one fits your work?

Claude Pro is a candidate when you specifically want Opus access; Gemini’s paid plans merit consideration when bundled storage matters; ChatGPT requires checking model availability in the surface you intend to use. Our command-line harness results cannot settle which consumer app fits your workflow.

For individual work, Claude Pro’s verified Opus access is a concrete purchasing distinction. ChatGPT Plus’s price does not establish its default GPT-6 variant. Google AI Pro lists Argon, Pro and Flash, but Argon’s app rollout matters when checking availability.

For heavier usage, compare the documented conditions rather than assuming a higher price means better answers. Max 5x includes Fable at 50% of weekly limits. ChatGPT reserves Ultrafast for Pro $500. Google AI Ultra distinguishes 5x and 20x AI Pro limits. We have not measured these usage ceilings.

For teams, obtain the actual seat quote, usage charges and contract terms. Claude has verified Team prices here; ChatGPT Business prices are unverified, and no Gemini team plan is recorded. Missing information is not evidence that an option is unavailable.

For storage needs, Google’s recorded bundles range from 15 GB on Free to 20 TB+ on Ultra. Treat storage as a budget consideration, not evidence of answer quality or permission to upload confidential files.

For API work, Claude Sonnet 5.5 and GPT-6.1 Sol list $2 / $10 per MTok. Gemini 4 Argon has that rate only through 2026-11-30, then moves to $4 / $20 on 2026-12-01. Fable 5.1 and Astra list $10 / $50. Subscription prices and API rates describe different charges.

Adapt your checks to your profession. Clinicians should compare every summary statement with the source; see AI for doctors and nurses. Engineers should verify units, assumptions and calculations; start with AI for engineers. Agents should check listing dates and claimed amenities; consult AI for real estate agents.

Use a nonconfidential sample to compare outputs before committing. Keep the input, instructions and grading criteria unchanged.

You are helping a [PROFESSION] in [COUNTRY/STATE].
Summarize the document in no more than [WORD LIMIT] words.

Use only facts stated in the document.
Do not add warnings, dates, benefits, prices or legal assumptions.
Mark missing information needed for the summary "NOT STATED".
Preserve all figures and units.
End with "Facts I was unsure about", listing uncertainties
or stating "None identified".

Document:
[NONCONFIDENTIAL SAMPLE WITH NO CLIENT, PATIENT OR EMPLOYEE DATA]

Use this review prompt, then verify its judgments yourself:

Compare the summary with the original document.
For each criterion, give PASS or FAIL and a brief reason:
- Every factual statement is supported by the document.
- The complete output stays within [WORD LIMIT] words.
- All figures and units match the document.
- Missing information is marked "NOT STATED", not invented.
- The "Facts I was unsure about" list is present.

Original:
[SAME NONCONFIDENTIAL SAMPLE]

Summary:
[MODEL OUTPUT]

Repeat the task and keep a tally. Our October 2026 round used 3 runs per model per prompt; repetition can expose variation, but does not establish reliability across your workload.

Limits: what must a licensed professional check?

A licensed professional in your jurisdiction must verify the applicable clinical, legal, tax or engineering conclusions before relying on them. This article is not legal, medical, tax or engineering advice.

We have not verified any vendor’s data terms for this page. We make no claim about retention, training on inputs, certifications or business protections. Read the specific plan’s data/privacy terms and your professional confidentiality rules before pasting client, patient or employee data.

For US healthcare, HIPAA Privacy Rule 45 CFR 164.502(e) permits a covered entity to disclose PHI to a business associate only with satisfactory assurance, a business associate agreement, that it will safeguard the information. An AI vendor receiving PHI is a business associate.

Confirm the applicable agreement and workflow with your compliance officer; a paid subscription alone establishes neither. See what you can paste into ChatGPT under HIPAA.

“Tested” here means at least 2 completed and graded runs, not professional sign-off. Rubric compliance is not a substitute for checking professional accuracy.

FAQ