Choosing Legal AI Tools: Statute Retrieval and Case Analysis
Vendor pages document different prices, file limits, research modes, and data controls for ChatGPT, Claude, and Gemini, so legal teams need a controlled citation check rather than an assumed winner.
Sources checked 2 Oct 2026
For statute retrieval and case analysis, vendor pages document tools, limits, and controls, but do not establish which assistant is more accurate or suitable for a particular matter. Chat Picker has not tested ChatGPT, Claude, or Gemini for legal work; this rewrite removes the former page’s unsourced statistics and test results.
Know these limits before you start
As read on October 1, 2026, input caps are not interchangeable. OpenAI’s File Uploads FAQ says Free users have three uploads per day and each file uploaded to a conversation has a 512 MB hard limit. Anthropic’s file-upload help page lists 500 MB per chat file, 30 MB per project file, and text plus visual analysis for PDFs of 100 pages or fewer. Google’s Gemini file-upload help page says one prompt can contain up to 10 supported files, subject to availability, with each nonvideo file up to 100 MB. Use identical public PDFs.
A source link is not proof of support. OpenAI’s accuracy guidance says ChatGPT can sound confident when wrong. Anthropic’s incorrect-response guidance says Claude can produce convincing but ungrounded quotations. Google’s source-viewing help page says Gemini Apps may show a Sources panel when sources are available. OpenAI’s web-search help page also warns that citations may be incomplete, outdated, or incorrect.
Use public, nonclient material first. OpenAI’s Usage Policies prohibit tailored legal advice requiring a license without appropriate involvement by a licensed professional. Anthropic’s Usage Policy classifies legal interpretation, guidance, and decisions with legal implications as high-risk and calls for qualified professional review. Gemini’s safety guidelines warn that outputs may be inaccurate or violate the guidelines. These are vendor rules, not a compliance finding.
Before considering client data, review the training controls. ChatGPT’s pricing page lists a training opt-out for Free, Go, Plus, and Pro. Claude’s pricing page lists an opt-out for Free, Pro, and Max and says Team is not trained on by default. The Google AI plans page states no training setting. OpenAI’s upload guidance says uploaded chat files remain until deletion; matching Claude and Gemini retention rules were not settled by the pages read.
Research modes are not equivalent: OpenAI says completed deep-research outputs include citations, Anthropic says Claude Research can consume usage limits faster, and Google says Deep Research reports usually take 5–10 minutes and may take longer.
The test plan
Use a fresh chat for each run on the plan you intend to buy. Keep one log with the corpus ID, assistant, plan, mode, start and finish times, files used, citations offered and verified, unsupported statements, displayed usage, and permission failures.
-
Build a statute packet. Prepare current and superseded statute excerpts from one jurisdiction, an amendment note, and a source list checked against official publications. Give each assistant: “Using only the attached statute excerpts, identify each rule, quote the supporting text with document and PDF page, and list any amendment date the files do not establish. Put every gap under Not found; do not infer a citation.” Record quotation accuracy, pinpoint accuracy, omitted amendments, and references absent from the packet.
-
Build a case packet. Prepare public opinions with procedural histories and an independently checked citation list. Give this prompt: “Build a chronology from the attached opinions. For each legal proposition, give the case name, court, date, pinpoint page, and exact passage. Label direct holdings separately from inference and use no source outside the packet.” Record missing procedural steps, unsupported synthesis, and differences between quoted and original text.
-
Verify every citation. Prepare a proposition checklist tied to the packet. Give this prompt: “Check each proposition against the attached source packet. Mark PASS only when the cited passage supports the exact proposition; otherwise mark FAIL or NOT VERIFIED and explain the mismatch.” Open the originals yourself. Record citation existence, pinpoint correctness, quotation fidelity, and source-to-proposition support. Keep raw counts: verified citations divided by citations offered is your test result, not a general performance rate.
-
Measure work, not impressions. Repeat one mixed statute-and-case request on the same device and network where possible. Give this prompt: “Answer using only the supplied files. State when a requested source cannot be accessed and do not fill gaps from memory.” Time the first visible output and final completion. Record retries, pages processed, displayed token or usage counters, and plan limits reached. Mark unavailable data as unavailable rather than estimating hidden tokens.
-
Check workflow fit. Place three public files in one folder and test upload, connected-source retrieval, citation display, and export. Give this prompt: “List every connected source used, identify where each fact appears, and flag any source you could not open.” OpenAI says data analysis can use Drive, OneDrive, or SharePoint when connectors are available. Claude Projects provide a knowledge base and separate chat history. Gemini Connected Apps vary by location and permissions, while Drive uploads require Workspace and Keep Activity. Record setup effort, permission scope, source traceability, and export portability.
How to read your results
Set pass conditions before viewing outputs. Retrieval passes only if the checked source supports every quoted rule. Case analysis passes only if material holdings and procedural steps trace to the originals. Operations pass only if the selected plan completes the defined workload within its documented controls. Governance passes only if your approved data settings and vendor terms fit the matter and a qualified reviewer checks the final legal work.
Keep plans and research modes in separate columns. Do not let faster output compensate for an unsupported legal proposition. Chat Picker has no test result, score, ranking, or benchmark to add.
Where the plans differ
Before subscribing, open the Claude vs ChatGPT vs Gemini matrix and check these 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.
At the October 1, 2026 read date, ChatGPT Plus was $20 per month, billed monthly. Claude Pro was $20 monthly or $17 per month with annual billing and $200 up front, while Google AI Pro was $19.99 per month.
For the low-cost tier, ChatGPT Go showed A$13 per month on the Australian page, but its US-dollar price was not verified. Anthropic listed no tier between Free and Pro. Google AI Plus was $4.99 per month.
For teams, ChatGPT Business listed a standard seat at $20 per month with annual billing or $25 monthly, with at least two seats. Claude Team had the same standard seat prices and a range of two to 150 people. Gemini business pricing was not verified.
The context row is not directly comparable: OpenAI lists different app windows by mode and tier, Anthropic lists up to 1M on every plan depending on the model, and Gemini’s app token figure was not verified. OpenAI and Anthropic do not publish exact message counts on their pricing pages; Google describes compute-based limits that refresh every five hours up to a weekly limit.