Chat Picker

Long-Text Handling in AI Assistants: A Context Test Plan

Vendor pages set plan and usage limits, while this reader-run test compares long-document retention, instruction following, references, and timing across ChatGPT, Claude, and Gemini.

Sources checked 2 Oct 2026

Long-text handling cannot be settled by a context-window number: OpenAI’s ChatGPT pricing page and Anthropic’s Claude pricing page list app context windows, while Google’s Gemini Apps limits page documents usage conditions, but none provides a retention-accuracy result for your document. Chat Picker has not tested ChatGPT, Claude, or Gemini for this purpose. The plan below turns your own material into a controlled afternoon test; unsourced statistics and test results from the earlier page are not reused.

Know these limits before you start

As read on October 1, 2026, OpenAI’s ChatGPT pricing page lists app context windows of 27K for Free instant models, 54K for Go and Plus instant models, and 128K for Pro instant models; its reasoning-model figures are 256K for Go and Plus and 400K for Pro. It gives paid tiers higher upload and message allowances without exact counts. Anthropic’s Claude pricing page says context can reach 1M on every plan and varies by model.

A window is capacity, not proof of recall. Anthropic’s context-window documentation says large documents can hit request limits before the token limit and warns that accuracy and recall can degrade as token count grows. Google’s Gemini Apps limits page says app limits vary with prompt complexity, feature and model use, and chat length; they refresh every five hours until a weekly limit is reached, as read on October 1, 2026.

Check data settings before using real files. OpenAI’s pricing page offers a training opt-out on Free, Go, Plus, and Pro; its Business overview says workspace data is not used for training. Anthropic’s Claude pricing page says model training is opt-out on Free, Pro, and Max, and Team is not trained on by default. Google’s AI plans page does not state a corresponding setting, so those pages do not settle it.

OpenAI’s accuracy note says ChatGPT can sound confident while wrong and advises checking important information. Anthropic’s incorrect-response note says not to rely on Claude as the only source and to inspect cited sources. Google’s policy guidelines describe LLM outputs as probabilistic and warn that Gemini may sometimes violate its guidelines. For covered legal or medical advice, OpenAI’s Usage Policy calls for licensed involvement; it also prohibits automated high-stakes decisions without human review. Anthropic’s Usage Policy classifies legal interpretation as high-risk and calls for qualified professional review.

The test plan

  1. Prepare controlled files. Assemble a synthetic report called Harbor Renewal Report. Put Orchid Relay near the opening, Copper Kite in the middle, and Harbor Glass near the end. Give each marker a different owner and dependency. Add a reference map that links those facts to later decisions and ends at a stated approval condition. Create a separate 100-page brief with its approval rule on a middle page and its unresolved risk near the end.

  2. Control the run. Use the candidate plan you would actually consider. Before submitting, write your must-pass conditions and record the relevant data setting. Start from a fresh chat on the same device, browser, and connection; submit the same file and prompt to each assistant. Record the plan, displayed model, date, time, and upload result.

  3. Measure retention and drift. Submit this prompt:

Use only the attached Harbor Renewal Report. Return a table with columns Marker, section heading, stated owner, exact supporting sentence, and dependency. If a marker is missing, write Not found and do not infer it.

Check whether names, qualifiers, dependencies, and contradictions survive. Record pass, partial, or fail, then save the supporting evidence.

  1. Trace references and final-paragraph compliance. Use a fresh chat for each prompt:

Using the attached Harbor Renewal Report and its reference map, trace every dependency from Orchid Relay to the final approval condition. For each link, quote the supporting sentence, name its section heading, and flag any break. If the report and map disagree, show both.

Using only the attached 100-page brief, write a decision summary. In its final paragraph, identify the largest unresolved risk, name the approval owner, cite the supporting page, and end with the exact sentence Review required before release. Do not add anything after that sentence.

Look for unsupported links, missing page references, missing final elements, and text after the required ending. Save the relevant outputs.

  1. Separate timing from task fit. Repeat the retention run and time from submission to the first visible token and completion. Keep those times separate from the retention result. Then attach a short specialist packet from your own work and use:

Extract the accountable owner, earliest timestamp, unresolved action, and source heading for each incident. Preserve exact names and write Not stated when a field is absent.

Record name and timestamp errors, missing actions, source-heading accuracy, and elapsed time. This reveals task fit without declaring a general winner.

How to read your results

Keep a simple log with one row per assistant and prompt: plan, displayed model, date and time, upload result, pass, partial, or fail, evidence, first-token time, completion time, and data setting.

Apply your must-pass conditions before comparing preferences. Mark a run as failed if it cannot accept the required file, omits a required marker, misses final-paragraph instructions, or cites evidence that does not support the claim. Do not let a large context window or fast response cancel a failed requirement.

Among runs that meet your requirements, compare repeated timing, source quality, and performance on the specialist packet. Keep missing facts separate from latency; an average can hide both. Save raw answers and links so a person can check them. The result is a fit decision for your task, not a universal ranking.

Where the plans differ

Use the Claude vs ChatGPT vs Gemini matrix rows for Context window in the app, How usage limits are described, Training on your chats, Ads, Free plan, Low-cost tier, Main paid plan, Heavy-use plans, and Team plan. Also read the feature notes for uploads and storage.

The matrix contains vendor-sourced, dated figures, marks unconfirmed cells “Not verified,” and leaves them blank; it contains no site-run speed or quality scores. Prices are US dollars unless a cell says otherwise, so confirm the vendor page before subscribing. The method page explains the sourcing rule.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants