Claude vs ChatGPT (2026): Tested on Real Work Tasks

What did our tests actually measure?

Our tests measured whether models met pre-written criteria on professional tasks, not whether their outputs deserved professional sign-off. They also did not test the consumer apps.

Round 1 used 40 prompts, 30 English and 10 Spanish, across 10 professions. Each prompt had 5 rubric criteria and ran 3 times per model, producing 360 runs. Round 2 used 20 French and 20 Brazilian Portuguese prompts with the same design, producing another 360 runs.

We tested Claude Opus 5.5, Claude Sonnet 5, and GPT-6 Astra through a command-line API harness with tools, web, and memory off. Consumer-app results may differ. Sonnet 5 is an older version, not the current Claude Sonnet 5.5. On 2026-10-20 we added Claude Fable 5.1 on the same API harness and Gemini 4 Argon through Google’s Antigravity CLI with a tool-less agent, not the Gemini app, both on all 80 profession prompts with the same rubrics and 3 runs per prompt; consumer-app results may differ. We did not test Claude Sonnet 5.5 or Claude Haiku 5.5.

Each output was graded by another model family: Claude and Gemini outputs by gpt-6-astra, GPT outputs by claude-opus-5-5. These judges were not equally strict, so part of the published gap comes from the grader.

Test dates: 2026-10-19, with Claude Fable 5.1 and Gemini 4 Argon added 2026-10-20. API or CLI harness only; app results may differ, and unequal cross-family judges affect comparisons.

Calibration shows why the judge caveat matters. These were re-grades of API-harness outputs, not app tests:

• Round 1: 30 sampled Claude runs passed 93% of criteria under the published GPT judge and 98% under the alternate Claude judge. The 30 sampled GPT runs passed 97% under either judge.
• Round 2: 10 sampled Claude runs passed 74% under the published judge and 86% under the alternate judge. The 10 sampled GPT runs passed 98% under either judge.
• An editor agent, not a human, independently re-read verdicts before seeing the original judgment. It agreed with 18 of 20 in round 1 and 9 of 10 in round 2.
• Claude Fable 5.1 and Gemini 4 Argon (10 sampled runs each per round): round 1, 88% and 92% under the published gpt-6-astra judge versus 100% and 98% under claude-opus-5-5; round 2, 82% and 86% versus 90% and 88%.
• For the added models, the editor agent agreed with 8 of 12 Fable verdicts and 7 of 10 Gemini verdicts; in every disagreement the judge was stricter than the editor.

Opus, Sonnet, Fable, and Gemini shared the gpt-6-astra judge, making comparisons among them more direct. Calibration does not establish how much of the full cross-family gap reflects model performance.

Most lost criteria involved added content: a warning, date, benefit, or legal assumption absent from the input. The Claude models also lost criteria on word limits. Read the methodology alongside the first test-round report before using these results to shortlist tools.

Which one should you try for your profession?

GPT-6 Astra is a reasonable candidate for a local trial because it matched or exceeded the published results of Claude Opus 5.5 and Claude Sonnet 5 in every profession. That is an API-model recommendation, not proof that the ChatGPT app will outperform Claude for your work: the harness excluded tools, web, and memory, and unequal judges account for part of the gap.

Test date: 2026-10-19. Round 1 share of rubric criteria passed. API results may differ from app results; Claude and GPT outputs had different, unequally strict judges.

The recommendations below use this table with both caveats intact. The facts do not identify profession-specific causes of lost points.

For lawyers, Astra warrants a trial, while Opus led Sonnet under their shared judge. Test with a fictional or redacted document and check whether the draft adds obligations, deadlines, or legal conclusions. A perfect rubric result does not validate an authority or interpretation. Continue with our guidance for lawyers.

For accountants, Sonnet led Opus under the same judge, while Astra had the highest published result among the three models in the table. Do not assume the more expensive Claude model is necessarily the right fit for a particular accounting task. Check every figure against the source and review client-data permissions before choosing. Our accountants section is the profession-specific starting point.

For healthcare professionals, Astra led the published results and Sonnet led Opus. Neither ordering establishes clinical reliability. Begin with fictional administrative material, and resolve confidentiality and BAA requirements before using patient information. See our guidance for healthcare professionals.

For real estate agents, Astra led, followed by Opus and Sonnet. A useful trial is to ask each candidate to draft from the same fictional listing brief, then check every property claim against that brief. This is a suggested evaluation, not an additional tested workflow.

For sales managers, Astra’s published result supports trying it first, subject to the judge and harness limitations. Compare source fidelity and compliance with your requested length. Do not attribute the sales results to a particular failure type: our error summary covers the benchmark overall.

For engineers, Opus led Sonnet under their shared judge. Construction managers had the only three-way tie; Opus and Astra tied for event planners and interior designers. Astra led for travel agents. Where results tie, the table gives no basis for preferring one model.

For any profession, evaluate the actual app or API configuration you intend to use. Record unsupported additions, missed instructions, and the edits required before approval. Do not assume a result from this harness transfers to a workflow with search or other tools enabled.

What do Claude and ChatGPT cost in 2026?

Claude Pro and ChatGPT Plus both cost $20 per month with monthly billing. Beyond those individual plans, compare the exact billing terms and verified model access rather than treating similarly priced plans as equivalent.

Prices and plans checked 2026-10-19 on the vendors’ official pages. USD list prices for US visitors, excluding tax.

Readers in the UK, Australia, and India should check locally: local prices, taxes, and rules differ, and we have not verified local-currency pricing.

Model access needs a separate check. OpenAI’s October system card says GPT-6 launched across free and paid ChatGPT plans, replacing GPT-5.6 Sol and GPT-5.6 Luna. The October Chat models are GPT-6 Sol and GPT-6 Luna; we have not verified which variant each plan uses by default.

GPT-6.1 Sol is in Work and Codex, not Chat, with rollout to Plus, Pro, Business, Enterprise, and Edu. GPT-5.5 was retired from ChatGPT-signed-in Codex and Work on 2026-10-14; its API was unaffected.

Claude Pro includes Opus but not Fable. Max 5x has verified Fable access under the limits shown above; do not infer access on other plans from their price. Claude Mythos 5.1 is restricted to vetted cyber-defense and life-science organizations, not a general consumer option.

For API budgeting, verified input/output prices per million tokens were:

• Claude Fable 5.1: $10 / $50.
• Claude Opus 5.5: $4 / $20.
• Claude Sonnet 5.5: $2 / $10.
• Claude Haiku 5.5: from $0.10 / $0.50.
• GPT-6 Astra: $10 / $50.
• GPT-6.1 Sol: $2 / $10.
• GPT-6 Luna: $0.10 / $0.50.

These API prices were checked 2026-10-19. The GPT-6 Luna API entry is the September version, distinct from the October Chat model. API rates are not consumer subscription prices. Our Claude overview and ChatGPT overview provide separate product references.

What must you check before sharing confidential data?

Read the specific plan’s data and privacy terms and your professional confidentiality rules before pasting client, patient, or employee information. We have not verified either vendor’s retention, training, or business-protection terms for this page.

We therefore make no claim about whether either plan uses inputs for training, provides particular security protections, or carries a certification. A business or enterprise label does not establish suitability.

HIPAA: Under 45 CFR 164.502(e), a covered entity may disclose protected health information to a business associate only after obtaining satisfactory assurance, through a business associate agreement, that it will safeguard the information. An AI vendor receiving PHI is a business associate. Confirm the required BAA and applicable plan terms before disclosure; we have not verified either vendor’s BAA availability.

Legal ethics: The State Bar of California lists its Practical Guidance for the Use of Generative AI in the Practice of Law and ABA Formal Opinion 512. California’s guidance connects AI use with competence and diligence under Rules 1.1 and 1.3, and confidentiality under Bus. & Prof. Code §6068(e) and Rule 1.6. It warns that AI products may use inputs to train or refine models. Attorneys elsewhere should check their own jurisdiction’s guidance.

Financial and state privacy questions: Ask your compliance lead or counsel whether GLBA or state privacy obligations apply to the proposed workflow. We have not analyzed those requirements here and cannot determine what your firm may disclose.

For an initial trial, use fictional material and placeholders. If proposing redacted material, have an appropriate reviewer assess it before upload; removing names alone should not be treated as permission to disclose. Document which tool and plan your firm approves for each data category, and assign a person to review outputs.

How should you prompt Claude or ChatGPT?

Specify the source, allowed content, output format, and length, then review the result against those instructions. These suggested templates address the benchmark’s general failure patterns; we have not measured their effectiveness or tested them as interventions.

Use fictional material or placeholders, never real client or patient data, in these examples.

Draft for a [PROFESSION] in [JURISDICTION].
Use only the facts in the input.
Do not add dates, warnings, benefits, legal assumptions, or figures.
If necessary information is missing, write [MISSING: what is needed].
Maximum length: [WORD LIMIT] words.
Format: [OUTPUT FORMAT].

Task:
[TASK]

Input:
[FICTIONAL SOURCE TEXT]

For document comparisons, make the requested evidence explicit:

Compare these fictional documents for a [PROFESSION].
Identify differences affecting obligations, figures, dates, or parties.
Quote the exact wording from each document.
Do not infer intent or supply missing legal rules.
Separate textual differences from questions requiring professional review.

Output columns:
Topic | Document A wording | Document B wording | Review question

Document A:
[FICTIONAL TEXT]

Document B:
[FICTIONAL TEXT]

For correspondence, keep the review visible:

Draft a [LETTER / EMAIL / NOTICE] for [AUDIENCE] about [TOPIC].
Use only the supplied fictional facts.
Tone: [TONE].
Maximum body length: [WORD LIMIT] words.
After the draft, list missing information under "Needs review".
Do not fill gaps with assumptions.

Input:
[FICTIONAL FACTS]

Check the output yourself rather than relying on the model’s self-review. Trace factual claims to the input, verify calculations and quotations, and check length. If you revise a prompt, compare outputs before assuming the change helped. Use our prompt library for further starting points.

Limits: what requires professional review

This page is not legal, medical, tax, or financial advice. “Tested” means at least 2 runs completed and graded, not professional approval.

An appropriately licensed professional in your jurisdiction must review legal conclusions, tax or accounting treatment, and clinical statements before reliance. Your compliance or contracting lead should confirm data permissions and required agreements before confidential information is involved. Recheck vendor prices and plan terms before buying.

FAQ