ChatGPT, Claude and Gemini for Indian trading tasks: a documentation comparison

Start with the evidence your task needs

A trader reading a company report needs a traceable summary. Someone auditing a journal needs the entries and their own rules. Someone checking a charge calculation needs the dated fee schedule and a reproducible formula. These are different jobs; a fluent answer does not show that an assistant completed any of them correctly.

The comparison below describes the providers' published tools and limits. It makes no measured claim about accuracy, speed, Indian-market knowledge or relative reasoning quality. No account was used to benchmark any model. Model names in a subscription menu can change, and a consumer chat app is a different product from its developer API.

For a useful comparison, give each assistant the same material, ask the same narrow question and inspect the answer against an external reference. Prefer tasks where you can identify an error without trusting another chatbot to judge it. A source-grounded document summary is easier to check than an unsupported prediction about tomorrow's market.

What the official documentation establishes

Documentation checked 14 October 2026. Features are not measured task results.

Assistant	Files and visuals	Sources and access
ChatGPT	Free and paid document, spreadsheet and image uploads, subject to limits. Enterprise PDF visual retrieval; other plans extract digital text rather than embedded images.	Search on Free and listed paid plans; may search automatically. Results and citations may be incomplete, outdated or wrong.
Claude	Documents and images; XLSX requires code execution and file creation. PDFs of up to 100 pages include visuals; 101–1,000 pages are text only.	Web search and direct URL fetch with citations. Workspace controls and usage limits apply; the new experience has no search toggle.
Gemini	Documents, spreadsheets and other supported inputs; spreadsheet chart creation documented. Large inputs can lose connections or details.	Source and related-content links may appear. Some answers have no links, and a related source need not support every sentence.

Sources: [1], [2], [4], [5], [6], [7]

Download this table (CSV)

Upload size does not measure reading accuracy

ChatGPT's guide lists a 512 MB document limit, a two-million-token text limit, approximately 50 MB for spreadsheets and 20 MB for an image. Free users have three uploads per day, and limits can be reduced at peak times. Claude lists 500 MB per chat file and up to twenty files per chat, with a separate 30 MB project-file limit. Gemini allows up to ten files in a prompt, subject to availability, and 100 MB per non-video file. These are distinct limits on different products; they cannot be turned into a quality ranking. OpenAI uploads, Claude uploads, Gemini uploads.

Prepare an extract that includes the relevant header, unit, table and notes. If a report's chart matters, ensure the visual is actually an input the selected plan can read. A successful upload notification does not prove the model used the chart. Ask it to identify the source page and any unreadable element before interpreting it.

Ten tasks you can verify without believing a model

This is a task specification written for this guide, not a report of completed runs. The checks are proposed methods. Each row tells you what to supply and what would count as evidence of a satisfactory answer; it gives no score to any provider.

Keep the input fixed when comparing assistants. Record the visible model label, subscription tier, date, tools enabled and full conversation, including corrections. An answer repaired after you pointed out a mistake should not be presented as a correct first response. If one assistant cannot access a source, record the access limitation before comparing its reasoning.

The tasks are educational and retrospective. They do not ask for market direction, a recommended contract or a trading signal. For more ready-to-adapt wording, use the trading prompt collection. The distinction between a published capability and a demonstrated result matters throughout the sheet.

Educational task sheet and proposed verification

Written 14 October 2026. Not run, not scored; no raw chatbot outputs exist for these tasks. Source IDs support tool availability and the education boundary.

Task	Material and request	Your independent check
1. Read an old chart	Attach a chart dated at least thirty days before your session. Ask for a transcription of visible axes, labels and the plotted period, followed by a list of unreadable items. Do not ask for the next candle, a target or an entry.	Compare each transcribed number with the image. A description of a trend is separate from a forecast. Reject precise prices copied from an indistinct axis, and keep the original screenshot beside the answer. A missing scale should produce uncertainty rather than a guessed value.
2. Explain an option payoff	Provide an entirely hypothetical set of strikes, premiums and quantities. Request the formula, units, maximum loss under the stated assumptions and any missing contract information. Specify that the numbers are invented and ask for no trade suggestion.	Recompute the algebra separately. Check whether quantities mean units or lots and whether the answer silently inserts a current lot size. A valid explanation should identify the missing multiplier and separate a payoff before costs from a broker's final bill.
3. Read an exchange notice	Upload the actual notice and ask for its issuer, reference number, publication date, effective date and contracts affected. Ask the assistant to separate the rule from its own explanation and mark information that is not in the notice.	Open the original and any later amendment yourself. A correct old notice can still be superseded. Keep the source chronology with the answer so a model cannot turn a transition rule into a permanent statement about every expiry.
4. Explain charge arithmetic	Supply your own dated fee inputs, order count and clearly defined turnover. Ask for a line-by-line calculation with a formula beside each amount. Keep statutory rates in the linked charges guide rather than asking the chatbot to recall them.	Use a calculator or spreadsheet independently. Separate executed orders from attempted orders, brokerage from data fees and the charged buy or sell side. A neat total with an undocumented tax base should remain an unresolved calculation, even if the arithmetic adds up.
5. Summarise a concall	Attach a public earnings-call transcript and ask for management's statements on demand, margins and debt, with speaker and page references. Ask for uncertainty and guidance to remain in the original conditional language. Do not request a stock-price reaction.	Find the cited passages in the supplied file. Identify whether a number is historical performance, a question from an analyst or management guidance. Keep 'we expect' separate from a completed result. Do not let a summary turn a tentative statement into a promise.
6. Compare report sections	Give the same relevant annual-report extracts to each assistant. Ask it to compare the reporting periods, units, segment definitions and any restatements before discussing changes. Request explicit 'not found' entries when a figure is missing.	Check the unit header and reporting basis before subtracting values. Standalone and consolidated accounts need to remain distinct. Confirm the comparison uses matching periods and includes footnotes, rather than treating a visually large number as a comparable figure.
7. Inspect a historical CSV	Upload an old candle file with a description of its columns, timezone and instrument identity. Ask for duplicate timestamps, missing fields and suspicious gaps. Specify whether you want a descriptive audit, without predictions or performance optimisation.	Reproduce the reported row counts and examples in the original file. A trading holiday and a missing data interval are different explanations; the assistant should not decide between them without an appropriate calendar and your sampling rules.
8. Review code without execution	Provide a small, isolated example that calculates a journal metric. Ask for an explanation of inputs, edge cases, rounding and assumptions, followed by a proposed correction if necessary. Remove tokens, account details and any order-placement connection.	Read the revised code and verify it on invented cases before using it. A model explaining a function has not run it. Ask the response to identify which findings come from inspection and which, if any, come from an actual execution record.
9. Check an option-chain extract	Use an archived extract old enough for the education rule, with the expiry, timestamp and field definitions. Ask for a transcription and a list of unavailable fields. Do not request a current chain, an invented Greek or a suggested strike.	Compare contracts, call and put columns and the timestamp with the original. Open interest is not volume, and a missing implied-volatility field should stay missing. Demand the data source for each field instead of accepting a plausible fabricated chain.
10. Review a sample journal	Provide a redacted sample of your own older notes with the rules you intended to follow. Ask the assistant to identify recorded rule breaks, cite entry identifiers and separate observed patterns from unanswered questions. Ask for no new trade ideas.	Read every flagged entry. An unrecorded stop, mood or reason is unknown, rather than evidence that you ignored it. Check whether a repeated pattern uses all relevant entries or selected anecdotes. The assistant can organise the record; you decide whether its description matches it.

Sources: [1], [4], [6], [12]

Download this table (CSV)

If you later turn this into a real comparison

1. 1

   Freeze the reference answer first

   Choose the documents and expected calculations before opening the chat apps. Define which statements require exact matches and which allow a faithful paraphrase. Retain the dated originals so later source edits do not change the grading basis.

2. 2

   Keep first answers and repairs separately

   Save the full output with its input, model label, tools and session date. Record failed uploads, refusals and inaccessible websites as events. Do not replace an inconvenient first answer with a polished retry while calling the result a first-pass test.

3. 3

   Use independent readers

   Ask two reviewers to check the same outputs against the fixed references and explain disagreements. They should distinguish factual correctness, completeness and clarity. A combined score can conceal a serious factual error, so publish the actual findings beside any future summary.

Seven Indian-market facts to check before trusting an answer

The following are possible error categories, not errors we observed in ChatGPT, Claude or Gemini. We have no dated screenshots or raw outputs establishing provider mistakes. They are included because an Indian trading task needs an appropriate authority for each field.

Check the lot size against the contract's applicable exchange notice, rather than a model's remembered multiplier. Check the expiry date against the relevant contract calendar and subsequent changes. Check a statutory charge against the dated rate and taxable base in the charges changelog, rather than a rounded brokerage estimate.

For a ban-period status, use the exchange's dated report and the exact security. For a price band or circuit, identify the market segment and instrument rather than applying a generic percentage to every derivative. For an option-chain value, require an actual source extract and timestamp. For a broker capability, consult the broker's own current documentation and distinguish a read-only connection from an order tool.

This checklist deliberately avoids printing unsourced current values. The answer you need is tied to your record's date and contract. A chatbot that says the input is insufficient can be easier to audit than one that supplies a precise but undocumented number. When an official source conflicts with another official page, retain both and resolve applicability before treating either as settled.

Before sharing a journal or statement

• Remove identity and secrets

  Redact account numbers, PAN, addresses, tokens, passwords and authentication codes. Use anonymous trade identifiers and only the fields needed for the task.

• Read the setting and the exception

  Training opt-outs, temporary chats, feedback and deletion are different controls. Check the policy for your consumer or business product rather than assuming a subscription tier guarantees a particular retention rule.

• Keep your original record

  Save the unedited source privately and mark any transformations. A summary should not become your only record of a trade or replace the broker's statement.

Consumer privacy controls have different exceptions

Official policies checked 14 October 2026; not a security ranking.

Provider	Published control	Relevant limitation
OpenAI	Model improvement off: new conversations excluded from training. Temporary chats are not used for improvement while temporary.	Feedback can allow the associated conversation to be used for training. Temporary chats may remain up to 30 days for safety; saving one makes it a regular chat.
Anthropic	Consumer training includes consent, safety review or explicit opt-in. Incognito chats are not used to improve Claude.	Feedback can retain the entire related conversation up to five years and may be used for training. Raw connector data is excluded, but material copied into chat can be included.
Google	Keep Activity on includes training and human review. Off or temporary chats have a training exclusion with feedback conditions.	Off/temporary chats retained 72 hours. Protective processing including human review continues; reviewed data may remain up to three years. This is not a zero-retention promise.

Sources: [9], [10], [11]

Paying for access does not prove trading skill

The official plan pages explain subscriptions, but this research did not establish comparable India consumer checkout totals across all three services. The fetched ChatGPT and India-localised Google pages did not expose a rupee amount. Claude's page separates consumer, workplace and developer products. We have not converted dollar prices, applied assumed taxes or used an enterprise seat price as a consumer quote. Check your own currency, billing period, tax and renewal conditions before paying. ChatGPT plans, Claude plans, Google India plans.

A higher allowance can let you upload more material; it does not establish that a model reads Indian exchange notices correctly. Start by checking whether the feature you need is available in your account, then assess outputs on material you understand. This guide makes no assertion that a paid tier reduces trading losses or that a free tier is sufficient for your workload.

Order access requires separate broker tools, permissions and authentication. A chat subscription should never be treated as evidence that the assistant can place an Indian exchange order. The official broker MCP comparison documents that separate connection layer; permissions need checking even when the chat app recognises a broker's name.

“Confidence isn’t reliability”

Judge the journal review by the entries it can cite

For retrospective review, context comes from your record: intended setup, rule, entry, exit and reason for changing the plan. Ask for references to individual entries and let the assistant leave gaps unresolved. Separate 'the note says I moved the stop' from 'the model inferred I was anxious'. The second statement needs evidence you may never have recorded.

OneTradeJournal uses manual trade entry, with no import, broker sync or order connection. Its paid AI can be wrong and only works with the information supplied. It does not send signals or stop trades. A trading journal with AI can help organise a review, but every conclusion still needs checking against the entries.

Keep the model's suggested descriptions alongside your corrections. A useful review tells you where the record supports a pattern and where more notes are needed. That is an auditable outcome even when you never choose a single winning chatbot.

Sources

Every figure on this page comes from one of these. Each was checked on the date shown.

1. [1]OpenAI: uploading files and audio · checked 14 Oct 2026
2. [2]OpenAI: searching the web · checked 14 Oct 2026
3. [3]OpenAI: accuracy and hallucinations · checked 14 Oct 2026
4. [4]Anthropic: upload files to Claude · checked 14 Oct 2026
5. [5]Anthropic: enable and use web search · checked 14 Oct 2026
6. [6]Google: upload and analyse files · checked 14 Oct 2026
7. [7]Google: sources and related content · checked 14 Oct 2026
8. [8]Google: responsible use of responses · checked 14 Oct 2026
9. [9]OpenAI: data controls · checked 14 Oct 2026
10. [10]Anthropic: consumer training and feedback · checked 14 Oct 2026
11. [11]Google: Gemini Apps Privacy Hub · checked 14 Oct 2026
12. [12]SEBI PR 48/2026, 17 August 2026 · checked 14 Oct 2026
13. [13]ChatGPT plans and pricing · checked 14 Oct 2026
14. [14]Claude plans and pricing · checked 14 Oct 2026
15. [15]Google AI plans, India page · checked 14 Oct 2026