Chat Picker

AI Assistants in Logistics: Route Optimization and Demand Forecasting

A practical test plan for comparing ChatGPT, Claude, and Gemini on logistics tasks while separating documented file and plan limits from untested route and forecast performance.

Sources checked 2 Oct 2026

For AI assistants in logistics, vendor pages settle upload, account, and policy conditions for a test, but do not say that ChatGPT, Claude, or Gemini will solve a traveling-salesperson problem, forecast demand accurately, or improve live operations. Chat Picker has not tested them for logistics; run the same bounded evaluation instead of treating feature lists as a verdict.

Know these limits before you start

Vendor pages cover OpenAI’s file help, Anthropic’s upload help, and Google’s file help, but none publishes a route or demand-forecasting benchmark.

As read on October 1, 2026, OpenAI’s File Uploads FAQ sets a 512 MB hard limit per chat file, 2M tokens per text or document, about 50 MB per CSV or spreadsheet, and 3 file uploads per day on Free. Anthropic’s upload help sets 500 MB per chat file and 30 MB per project file; XLSX requires code execution and file creation. Google’s file help allows up to 10 files per prompt, 2 GB per video, and 100 MB per other supported file. OpenAI says file counts may also vary by upload type, model, plan, workspace, and remaining allowance in its data-analysis help.

Verify data settings first. ChatGPT pricing says Free, Go, Plus, and Pro have a training opt-out; Claude pricing says Free, Pro, and Max have one, while Team is not trained on by default. Google AI plans states no chat-training setting, so the matrix marks it Not verified. Record the actual account setting before testing.

Keep an accuracy gate. OpenAI’s accuracy help says ChatGPT can be incorrect or misleading and may sound confident; Anthropic’s incorrect-response help says not to use Claude as the only source of truth; Google’s Gemini safety guidelines say outputs may reflect training-data limits and overgeneralizations. OpenAI’s usage policies prohibit automated high-stakes decisions in sensitive areas without human review; Anthropic’s Usage Policy requires qualified review for covered elevated-risk recommendations. Require human approval before a route, inventory, or capacity recommendation reaches operations.

The test plan

Use the same synthetic or properly de-identified files, prompt wording, and review window for each assistant; repeat each prompt unchanged.

  1. Route optimization. Input: A CSV with order ID, pickup and delivery coordinates, time windows, vehicle capacity, service time, depot, and hard operating rules. Instruction: “Find a feasible sequence in the attached route CSV. Identify missing or conflicting fields, separate hard constraints from preferences, show distance or travel-time calculations, and call the sequence optimal only when the method supports it.” Look for: Violations, unsupported distances, and hidden assumptions. Log: Feasibility, cited fields, method, elapsed time, and unresolved issues.

  2. Demand forecasting. Input: A timestamped CSV with location, SKU, units, calendar fields, and a held-out final period. Instruction: “Backtest the attached demand history from rolling origins, using only information available before each forecast date. Provide point forecasts, prediction intervals, horizons, and formulas and results for mean absolute error and weighted absolute percentage error.” Look for: Leakage, missing dates, and unexplained jumps. Log: Errors by horizon, interval misses, assumptions, and reproducibility.

  3. Real-time decisions. Input: A mock order-event JSON file and a redacted API schema with authentication, rate-limit, retry, timeout, and write-action fields. Instruction: “Without calling an external system, use the attached event and schema to draft pseudocode for validation, authentication, timeout, retry, logging, and a separately approved write action. List assumptions.” Look for: Invented endpoints, missing failure paths, and an unclear write boundary. Log: Assumptions, failure handling, and whether the write step is isolated.

  4. Cost and scale. Input: A workbook with expected monthly prompts, file volume, context needs, team seats, and peak concurrency, plus dated vendor pricing pages. Instruction: “Using the attached workload and pricing pages, calculate monthly subscription and usage cost. Show each formula, billing period, included allowance, and source date; mark any missing API or storage rate Not verified rather than estimating it.” Look for: Mixed units or currencies and omitted quotas. Log: Subscription cost, metered charges, the first limit likely reached, and unresolved items.

  5. Data integration and training settings. Input: A route CSV, demand spreadsheet, operating-constraints PDF, and account notes you are permitted to inspect. Instruction: “Combine the attached files, citing the filename and row or page for each operational constraint. Preserve column names, identify missing joins, and separate source facts from inference. Do not infer privacy, training, connector, or retention settings; list required operator checks.” Look for: Grounded citations, schema preservation, and missing joins. Log: Source visibility, admin prerequisites, verified data setting, and limits encountered.

How to read your results

Use gates before preferences. Reject a run that breaches a hard route constraint, leaks future information into a forecast, violates a policy boundary, or cannot be traced to the input. Among the remaining runs, compare your logged forecast errors, violations, traceability, repeatability, review time, plan fit, and dated cost. Treat a fluent answer, a large context figure, or a citation as evidence to inspect, not proof. Choose only the workflow you will pilot, keep human approval in place, and leave unverified prices or settings marked as such.

Where the plans differ

Use the three-way comparison matrix to check 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. Each figure there carries its vendor source and read date.

Vendor prices read on October 1, 2026, differ for Main paid plan: OpenAI’s Plus help page lists Plus at $20/month when billed monthly, Claude pricing lists Pro at $20 monthly or $17/month with annual billing after $200 up front, and Google AI plans lists AI Pro at $19.99/month.

For Low-cost tier, OpenAI pricing lists Go at A$13/month on its Australian page, rechecked October 2, 2026; Go may include ads, and no US-dollar price was verified. Pages read October 1 list no Claude tier between Free and Pro and Google AI Plus at $4.99/month (Claude pricing; Google AI plans).

For Context window in the app, pages read October 1 list ChatGPT limits by plan and model, say Claude pricing offers up to 1M tokens on every plan while varying by model, and leave Google AI plans’s app token figure Not verified. Under How usage limits are described, those pages give no exact message count for ChatGPT or Claude; Google describes compute-based limits that refresh every 5 hours up to a weekly limit.

Recheck the vendor page before subscribing, and use the matrix’s API-price entries as dated inputs to the cost test.

Sources

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

Compare assistants