Evaluating AI Chat Knowledge Breadth: Cross-Disciplinary Coverage Test
Vendor documentation does not establish knowledge-breadth winners, and this hands-on plan compares coverage, errors, and citation support across fixed physics, biology, and interdisciplinary prompts.
Sources checked 2 Oct 2026
The ChatGPT pricing page, Claude pricing page, and Google AI plans page document access, prices, features, and limits, but they do not establish which assistant has broader cross-disciplinary knowledge. Chat Picker has not tested ChatGPT, Claude, or Gemini for this use, so this page provides a test plan rather than scores or a winner. This rewrite removes the older page’s statistics and test results that had no sources.
Know these limits before you start
Accuracy comes first. OpenAI says ChatGPT may be incorrect or misleading, sound confident when wrong, and lack later knowledge unless tools are used. Anthropic says Claude can make convincing but ungrounded statements and advises checking original sources. Google says Gemini’s outputs reflect training-data limits and may include overgeneralizations. Treat every output as a claim to verify, not evidence.
Make research mode explicit. OpenAI says completed deep research outputs include citations or source links and follow normal data controls. Claude Research uses multiple searches, counts against standard usage limits, and may consume them faster; Gemini Deep Research uses Google Search by default, creates an editable plan, and requires a signed-in user aged 18 or older.
As read on October 1, 2026, OpenAI’s File Uploads FAQ sets a 512 MB hard limit per ChatGPT file, with three uploads a day for free users. Anthropic’s upload page lists 500 MB per chat file and 30 MB per project file. Google’s upload page lists up to 10 files per prompt, 100 MB per non-video file, and 2 GB per video, subject to availability.
Before connecting sources, record your data settings. ChatGPT’s pricing page lists an opt-out on Free, Go, Plus, and Pro; Claude’s pricing page lists an opt-out on Free, Pro, and Max and says Team is not trained on by default. Google’s plans page stated no training setting we could verify. Keep this exercise general rather than medical, legal, or financial advice: OpenAI’s usage policies prohibit tailored licensed advice without appropriate professional involvement, Anthropic’s Usage Policy treats legal guidance as high risk and calls for qualified review, and Google’s guidelines identify harmful medical and physical-safety inaccuracies.
The test plan
Run the test in one sitting and record each assistant’s plan and country; do not assume similarly named tiers are equivalent. For every run, save the prompt, full response, elapsed time, visible search or app use, and any limit message.
-
Build the inputs and answer key. Before testing, list the mechanisms, assumptions, likely errors, and one reliable source you expect to need for each item. Prepare four prompts: “A 2 kg block starts from rest and slides down a frictionless ramp with a 3 m vertical drop. Using g = 9.8 m/s², derive its exit speed and state every assumption.” “Explain how gene-expression changes can alter a protein’s abundance and location without changing its amino acid sequence, then explain why natural selection changes populations rather than an individual’s genes during its lifetime.” “A biofilm forms on a catheter. Explain how fluid flow, surface adhesion, quorum behavior, and gene regulation can interact to increase antibiotic tolerance.” “In dry soil, warming initially raises root respiration while water becomes scarcer. Trace the feedback among plant carbon use, microbial activity, and drought, and identify two uncertainties.” Save your key before asking any assistant to answer.
-
Run the plain-chat pass. Append this instruction to every prompt: “Do not use web search, connected apps, or deep research. Give a concise definition, a step-by-step mechanism, assumptions, one likely misconception, and two uncertainties. Mark any claim you cannot support.” Send the same wording to all three assistants. Record visible source or tool use and mark each expected mechanism as covered, missing, or contradicted.
-
Check physics and biology separately. Examine every equation, definition, and causal link against your key and reliable source. Record pass, partial, or miss for each expected element, keeping contradictions and unsupported statements separate.
-
Trace interdisciplinary handoffs. For the two mixed prompts, mark each transition between fields. Record the first missing, weak, or contradictory link, and whether the answer identifies uncertainty instead of presenting inference as established fact.
-
Audit every citation. Open each cited page. Label it “supports,” “partial,” “no support,” or “unverifiable”; record the source type, publication date, and whether the cited passage contains the claim. Keep uncited key claims separate, and do not equate citation count with citation quality.
-
Repeat with research tools. Append: “Use web research. Cite every key claim, distinguish sourced facts from inference, and keep the same five-part structure.” Record the research plan shown or edited, citations opened, inaccessible sources, completion, and any usage interruption. If a plan does not expose the research feature, log that rather than silently changing the test conditions.
How to read your results
Build one row per prompt and mode. Mark coverage, correctness, citation support, omissions, and task fit separately; keep elapsed time in its own column. Coverage means the expected mechanisms were addressed. Correctness means they match your key and reliable sources. Citation support means the cited source actually backs the claim.
Do not collapse these dimensions into a universal ranking. If you need one decision score, set the weights before testing and show the component marks beside it. Treat unsupported claims as errors in that run, not proof that an assistant “lies” more often. Choose the workflow that performs consistently on the domains, source requirements, and usage limits you actually need.
Where the plans differ
In the Claude vs ChatGPT matrix, inspect the rows for Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Team plan, Context window in the app, How usage limits are described, Training on your chats, and Ads. The vendor figures below were read on October 1, 2026.
OpenAI’s Plus page lists Plus at $20 per month. Anthropic’s pricing page lists Pro at $20 monthly or $17 per month with annual billing and $200 up front. Google’s U.S. plans page lists AI Pro at $19.99 per month.
For app context, ChatGPT’s pricing page lists 27K on Free, 54K on Go and Plus, and 128K on Pro for instant models; its reasoning-model figures are 256K on Go and Plus and 400K on Pro. Claude’s pricing page says context can reach 1M on every plan and varies by model. Google’s plans page gives no app token figure, so that cell remains unverified.
For usage, OpenAI and Anthropic provide no exact message counts on their pricing pages. Google describes compute-based limits that refresh every five hours up to a weekly limit. For ads, ChatGPT Go may include ads, while the pages read did not establish ad terms for the other plans; Claude and Gemini remain unverified. Confirm current prices and terms on the vendor page before subscribing.