AI Assistants in Marine Science Research: Data Analysis and Model Building
A practical afternoon test plan compares ChatGPT, Claude, and Gemini on marine data cleaning, model building, source checks, reproducibility, cost, and speed.
Sources checked 2 Oct 2026
[OpenAI’s data-analysis page], [Anthropic’s upload guide], and [Google’s file guide] document ways to work with uploaded material, but the vendor pages we read do not establish performance on marine-science data or model building. Chat Picker has not tested the assistants for this use, and this rewrite replaces the older page’s unsourced statistics and results with an afternoon test plan.
Know these limits before you start
The vendor facts and figures below are as read on their linked pages in October 2026.
[OpenAI’s file-uploads FAQ] gives every uploaded file a 512 MB hard limit, caps text and document files at 2M tokens, and limits CSV files and spreadsheets to approximately 50 MB. Its [data-analysis page] says a stateful Jupyter environment runs Python for some tasks but cannot make external web requests or API calls. [Anthropic’s upload guide] lists 500 MB per chat file and 30 MB per project file; XLSX requires code execution and file creation. [Google’s file guide] allows up to 10 supported files per prompt, subject to availability, and 100 MB per non-video file; it also says uploaded spreadsheets can produce customizable charts.
Before uploading sensitive material, check settings. [ChatGPT’s pricing page] says an opt-out is available on Free, Go, Plus and Pro; [Claude’s pricing page] says it is available on Free, Pro and Max, while Team is not trained on by default. [Google’s US plans page] states no training setting, so this article marks that point as Not verified. [OpenAI’s accuracy guidance] says ChatGPT may sound confident while wrong and urges verification of important quotes, data and references; [Anthropic’s accuracy guidance] says not to rely on Claude as the sole source of truth and to inspect original sources; [Gemini’s safety guidelines] describe its outputs as probabilistic and subject to training-data limits.
For elevated-risk work, [Anthropic’s Usage Policy] calls for relevant human expertise, while [OpenAI’s Usage policies] require human review for sensitive automated high-stakes decisions.
The test plan
Use the same files, prompts, account settings and network conditions in fresh chats, and keep a read-only master copy. Do not upload restricted data just to make the comparison even.
-
Clean a heterogeneous ocean dataset. Prepare
ocean_samples.csvwithstation_id,sample_time, latitude, longitude, salinity, temperature, nitrate andquality_flag; add duplicates, mixed units, missing cells and planted outliers. Prompt: “Inspect the attached file before changing it. Identify duplicate keys, unit conflicts, missing-value patterns and outliers; propose reversible cleaning steps; return a validation table and Python; do not overwrite the source.” Record detected issues, false alarms, unsupported assumptions and whether the code runs. -
Parameterize a biogeochemical model. Attach
biogeochemical_model_spec.pdfwith equations and a parameter table, plusforcing.csv. Prompt: “Map each parameter to its unit, stated range and forcing variable. Cite a page or table for every mapping, list undefined parameters and build validation checks; do not invent values.” Record missing mappings, citation mismatches and assumptions needing domain review. -
Test a downscaling workflow. Attach
coarse_grid.csv,station_observations.csvand metadata defining coordinates, time and units. Prompt: “Design a downscaling workflow that preserves coordinates, units, timestamps and missing-data flags. Separate observations from modeled values, return code and validation steps, and mark every expert choice.” Record leakage risks, transformations, unsupported claims and unresolved metadata. -
Check reproducibility. Reuse the model files. Prompt: “Create a README, data dictionary, configuration file, dependency file, seeded script and validation report. State every assumption and do not invent dependency versions.” Look for complete artifacts and inspectable code; record missing dependencies, ambiguous steps and whether a clean rerun reproduces the tables and plots.
-
Audit domain claims. Attach trusted marine-science sources. Prompt: “Using only these sources, explain parameter identifiability and bias in a biogeochemical model. Cite a page or section after every factual claim; say when evidence is insufficient.” Record unsupported claims, citation mismatches and places where inference is presented as evidence.
-
Measure cost and speed yourself. Run the same downscaling prompt in a fresh chat. Record start and finish time, retries, limit messages, plan, billing tier, model label shown, tools used, any usage counter and the amount billed. Compare completed work, not first-answer latency; do not estimate tokens or cost when the product does not display them.
How to read your results
Use a log row for each assistant-task pair. Record exact file hashes, prompt, plan, model label, checks passed, checks needing revision, unsupported claims, elapsed time, billed cost, retries and limit notices.
Before testing, choose hard gates: source traceability, no silent data changes, reproducible artifacts and domain review. Rank the remaining needs by project importance, but keep accuracy evidence separate from speed and cost. Compare only runs with the same inputs and instructions. When a tool, plan or source condition changes, rerun under that condition and record it. These are your observations, not Chat Picker tests or scores.
Where the plans differ
Use Chat Picker’s Claude vs ChatGPT matrix to check the rows 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. Confirm vendor pages before subscribing because plans change.
- All three list free access: [OpenAI’s pricing page], [Anthropic’s pricing page] and [Google’s US plans page]. [Anthropic] lists no low-cost tier between Free and Pro. [Google AI Plus] is $4.99/month with 2x Free limits and 400 GB storage. [OpenAI Go] appeared at A$13/month on the Australian page; its US-dollar price was not verified, and the page says the plan may include ads.
- Main paid plans are [ChatGPT Plus] at $20/month, [Claude Pro] at $20 monthly or $17/month with annual billing and $200 upfront, and [Google AI Pro] at $19.99/month.
- Heavy-use prices are [$100, $200 and $500 per month] for ChatGPT’s Pro tiers; [Claude Max] starts at $100 per month with 5x or 20x Pro usage choices; [Gemini AI Ultra] is $99.99 or $199.99 per month.
- Context and usage differ. [ChatGPT] lists 128K on Pro for instant models and 400K on Pro for reasoning; [Claude] says up to 1M on every plan, varying by model; [Gemini’s app page] gives no token figure, so the matrix leaves it blank. ChatGPT and Claude publish no exact message counts; [Gemini] says compute-based limits refresh every 5 hours up to a weekly limit.