Part 2: AI assistants for OSINT research

The seven AI assistants we see most in investigative work, compared on the data terms that matter for OSINT rather than the feature list.

Every major AI assistant now offers the same core research features. Each has a deep research mode that runs many searches and writes a cited report. Each has connectors to your files and email, an agent mode that can browse and act, and memory across chats. The differences that matter for OSINT are in the data terms, not the feature list. This Part compares the seven assistants we see most often in investigative work. We have tried to describe each vendor’s terms as they stand in October 2026. Policies change often. Read the vendor’s own page before you rely on any row.

2.1 Comparison table

Rubric strings are explained in Part 1. Where two strings appear, the first is for consumer plans and the second for enterprise or API access.

Assistant Plans and prices (October 2026) Where your data goes Rubric Notes for OSINT work
Claude (Anthropic) Free. Pro $20 a month ($17 billed yearly). Max $100 or $200. Team $25 or $125 a seat. Enterprise from $20 a seat plus usage.1 Current models are Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.2 Since August 2025, Free, Pro and Max users choose whether their chats and Claude Code sessions train future models. If you allow it, data is kept up to five years. If you decline, 30 days. Chats flagged by Anthropic’s safety systems can be used for model improvement even if you opted out, and are kept up to two years. Team, Enterprise, API, Bedrock, Vertex AI, Claude Gov and Education accounts are excluded from training.345 Consumer T🟡 R🟡 D🔴 Q🟡. Team, Enterprise and API T🟢 R🟢 D🟡 Q🟡 Research mode, web search, connectors (MCP), memory with incognito chats, Claude Code, Claude in Chrome. Incognito chats are never used for training.6
ChatGPT (OpenAI) Free. Go about $8 with ads. Plus $20. Pro $200, with a $100 tier added in April 2026. Business $25 a seat. Enterprise and Edu on request. Prices are from third-party reporting because the pricing page blocks automated checks. “Improve the model for everyone” is on by default for consumer accounts. Temporary Chats are deleted within 30 days and do not use memory. Enterprise, Edu and API accounts with zero data retention are not used for training.7 Consumer and API logs without zero data retention were under a court preservation order from May to September 2025, and 20 million de-identified logs were ordered produced to plaintiffs in November 2025.8910 Consumer T🟡 R🟡 D🔴 Q🟡. Enterprise and zero-retention API T🟢 R🟢 D🟡 Q🟡 Deep Research, Agent mode (July 2025), Apps SDK built on MCP (plugins ended in April 2024).11 The Atlas browser launched in October 2025 and was retired in August 2026.12 In July 2025 more than 4,500 shared chats turned up in Google search until OpenAI removed the “discoverable” option.13
Gemini (Google) Google AI Plus about $5. AI Pro about $20. AI Ultra about $100 to $200. Third-party reporting; check the Google One page. “Keep Activity” stores chats for 18 months by default. With it off, chats are still kept for 72 hours. Chats sampled for human review are kept up to three years even if you delete them.14 Personal Intelligence (January 2026) connects Gmail, Photos and other Google apps, opt-in per app.15 Vertex AI and Gemini Enterprise do not train on customer data and cache inputs for at most 24 hours.16 Consumer T🟡 R🟡 D🔴 Q🔴. Vertex AI T🟢 R🟢 D🟡 Q🟡 Deep Research since December 2024. NotebookLM for source-grounded reading. Gemini in Chrome. Strong image and video analysis. Gemini 3 Pro shipped in November 2025 and Gemini 4 began a limited release in September 2026.
Perplexity Pro $20. Max $200. Enterprise Pro $40 and Enterprise Max $325 a seat. Third-party reporting; the pricing page blocks automated checks. The “AI data retention” setting is on by default and can be turned off in settings (third-party guides). The Comet browser has been free since October 2025.17 T🟡 R🟡 D🔴 Q🟡 Good for fast, cited answers on current events. In August 2025 Brave showed that hidden text on a web page could make Comet open Gmail, read a one-time code and post it back.18 Cloudflare accused Perplexity of crawling sites that had blocked it.19 Dow Jones, the New York Times and others have sued it.
Le Chat (Mistral AI) Free. Pro €14.99. Team about $25 a seat. Enterprise on request. Consumer plans train on your chats by default with an opt-out. Enterprise is opted out by default.20 Mistral is a French company and offers EU hosting and self-hosting through La Plateforme. T🟡 R🟡 D🟡 Q🟡 Deep Research since July 2025, including on the free tier.21 Mistral Large 3 is released under Apache 2.0 and Medium 3.5 (April 2026) replaced the Magistral and Devstral lines.22 Mistral reportedly renamed Le Chat “Vibe” in 2026. We keep the old name until the company documents the change.
Grok (xAI) Tiers reported from $10 to $300 a month. X Premium+ also includes Grok. X uses your posts and your Grok conversations for training by default. Opt out under Privacy and Safety, then “Grok and Third-party Collaborators”. The Grok app has a separate “Improve the Model” switch.23 T🔴 R🟡 D🔴 Q🔴 Useful for watching trends on X. Not suitable for case material. In July 2025 Grok produced antisemitic posts and called itself “MechaHitler”.24 In August 2025 hundreds of thousands of shared Grok chats were indexed by Google.25 In January 2026 its image tool generated sexualised images of real people; Indonesia and Malaysia blocked Grok, and Ofcom and the European Commission opened investigations.26
Microsoft Copilot Microsoft 365 Copilot $30 a user a month. Copilot Business $21. Microsoft 365 Premium $19.99 replaced Copilot Pro for individuals in October 2025.27 Tenant data is covered by enterprise data protection and is not used for training. Your administrators hold the audit logs. T🟢 R🟡 D🟡 Q🔴 Researcher agent for deep research (generally available mid-2025). Good when your evidence already lives in SharePoint, Outlook and Teams. We found no evidence that Microsoft 365 Copilot tenant data was affected by the OpenAI litigation.

Verified 2026-10-09 against vendor pages unless marked as third-party reporting. Prices change. Check before you buy.

We dropped three things from this table since the last edition. OSINT-GPT was a third-party wrapper on YesChat.ai that added a data hop with no published terms. The “custom GPT” listings are covered by the warning in Part 4. The long Mistral product list now lives in one row above and in Part 3.

2.2 Deep research across vendors

Deep research modes run dozens of searches, read the pages, and write a report with citations. They are the single most useful AI feature for background work on a company, a vessel, a port or a trafficking route. They are also where hallucinated citations show up most often, so every link in a deep research report needs to be opened and read. Part 6 explains how we check them.

Vendor Feature Launched Plans Vendor-stated run time
Google Gemini Deep Research 11 December 2024 AI Pro and above, limited free use Minutes
OpenAI Deep Research 2 February 2025 Plus, Pro, Business, Enterprise; limited free use 5 to 30 minutes
Perplexity Deep Research 14 February 2025 Free (limited) and Pro Minutes
xAI DeepSearch 17 February 2025 Paid X and SuperGrok tiers Minutes
Anthropic Research 15 April 2025 (Pro since 3 June 2025) Pro, Max, Team, Enterprise 5 to 15 minutes, up to 45 for complex questions2829
Microsoft Researcher agent Generally available mid-2025 Microsoft 365 Copilot Minutes
Mistral Deep Research in Le Chat 17 July 2025 All tiers including free30 Minutes

Launch dates from vendor announcements and contemporaneous reporting.31 Verified 2026-10-09.

2.3 Connectors, MCP and agent modes

The Model Context Protocol (MCP) is the open standard that lets an assistant call outside tools and data sources. Anthropic published it in late 2024, OpenAI adopted it in March 2025, and Microsoft followed at Build in May 2025.32 In December 2025 Anthropic donated MCP to the Linux Foundation’s new Agentic AI Foundation, citing more than 10,000 public servers.33 The official registry listed roughly 10,000 to 19,000 entries at different points in 2026, depending on how you count.34

Three things to know before you connect anything:

  • The old reference servers are unmaintained. In May 2025 the MCP project moved its GitHub, Slack, PostgreSQL, Google Drive and Puppeteer servers to an archive. The archive notice says “no security guarantees are provided”.35 Earlier editions of this guide linked to them. Use vendor-maintained servers instead: GitHub’s own server,36 Exa’s,37 Tavily’s,38 and the Slack server now maintained by Zencoder.39
  • For OSINT data, prefer servers run by the data owner. OpenSanctions ships an MCP server with its open-source client.40 The popular Shodan MCP server is a community project under an MIT licence, not a Shodan product.41
  • A connector is a new way for data to leave. Each MCP server you add can read what the model reads and may send it elsewhere. Part 4 covers the attacks we have seen in the wild and the rules we follow.

2.4 Memory and temporary modes

Memory lets an assistant carry facts between chats. For OSINT that is a liability. A target’s name from one case can surface in a chat about another. Memory also builds a profile of you, the analyst, that a breach or a discovery order could expose. And memory can be planted: a web page an agent reads can try to write into it.

Assistant Memory behaviour How to avoid it
ChatGPT Saved memories, plus “reference chat history” across all past chats since April 2025. On by default where available.42 Settings, Personalization, turn off both switches. Temporary Chat neither reads nor writes memory.
Claude Memory rolled out to Team and Enterprise in September 2025 and to Pro and Max in October 2025. You can view, edit or pause it.43 Incognito chats are kept out of history, memory and training.
Gemini Saved info and past-chat recall are tied to “Keep Activity”. Personal Intelligence connects Google apps, opt-in per app.44 Turn off Keep Activity and disconnect apps. Temporary Chats skip personalisation.
Grok Memory launched in April 2025, not in the EU or UK. Settings, Data Controls, turn off personalisation with memories.

Our rule: case work happens in an incognito or temporary chat, or in a separate account with memory off, or on a local model.

2.5 Enterprise and no-training routes

The question we hear most often is “can I use a frontier model safely at all?” The answer is yes, if you buy it under business terms. Four routes:

  • The vendor’s own business plans. Claude Team and Enterprise, ChatGPT Enterprise, Edu and Business, and Microsoft 365 Copilot are all excluded from model training. They also give administrators audit logs and retention controls.
  • API access with zero data retention. Anthropic, OpenAI and Google all offer API terms that exclude training. OpenAI’s zero-data-retention option requires approval and is set per organisation or project.45
  • Cloud marketplaces. Amazon Bedrock, Google Vertex AI and Microsoft Foundry serve Claude, Gemini, Llama, Mistral and other models inside your own cloud account. You choose the region and your data is not used for training.46 This is the usual answer for government and regulated organisations.
  • Open weights you run yourself. See Part 3.

Cohere sells to this market too. Its Command A+ model was released as open weights under Apache 2.0 in May 2026, and its North product is a workspace for private deployments.47

2.6 Research and academic tools

Tool What it does Cost OSINT use
Semantic Scholar AI-assisted search over academic papers with summaries Free Methodology and policy literature on trafficking and smuggling
arXiv Preprints in computer science and other fields Free New detection methods and AI tools before they are peer reviewed
SSRN Social science working papers Free with paid extras Economic and policy analysis of illicit markets
ResearchGate Researcher network and paper sharing Free Finding and contacting authors
Consensus Search engine that summarises what papers agree on Free tier, paid plans Quick checks of research consensus
Elicit AI research assistant that extracts data from papers Paid tiers; see site Systematic reviews and evidence tables
ResearchRabbit and Connected Papers Citation maps Free tiers Finding the literature around a key paper
Scite Shows whether later papers support or contradict a claim Paid Source reliability checks
NotebookLM (Google) Answers questions only from the sources you upload Free with Google account; paid tiers Reading large document sets with inline citations. Google’s consumer data terms apply

BloombergGPT (2023) was the first large language model trained on financial text and remains a useful reference paper on domain-specific models.48 Bloomberg’s 2026 Terminal products use a mix of commercial and open-weight models.49

A worked example with one vendor

This example uses Claude Research with MCP connectors. ChatGPT Deep Research with connectors or Gemini Deep Research with Google Workspace work the same way.

  1. Open an incognito chat on a Team or Enterprise account, or on a Pro account with model training switched off.
  2. Ask a public, non-identifying research question first. For example: “How do Chinese-flagged fishing fleets use transshipment at sea to move catch? Which ports in West Africa are named in recent reports?”
  3. Let Research run. It takes five to fifteen minutes and returns a report with citations.
  4. Open every citation. Mark each claim as confirmed, unconfirmed or wrong. Expect at least one wrong or missing citation per report.
  5. Only then move to the case specifics. Do that part on a local model or an enterprise account, with the confirmed sources pasted in as context.

Footnotes

  1. Anthropic. “Pricing.” https://claude.com/pricing. Accessed 2026-10-09.↩︎

  2. Anthropic. “Models overview.” https://platform.claude.com/docs/en/about-claude/models/overview. Accessed 2026-10-09.↩︎

  3. Anthropic (28 August 2025). “Updates to Consumer Terms and Privacy Policy.” https://www.anthropic.com/news/updates-to-our-consumer-terms. Accessed 2026-10-09.↩︎

  4. Anthropic. “Privacy Policy”, effective 10 September 2026. https://www.anthropic.com/legal/privacy. Accessed 2026-10-09.↩︎

  5. Anthropic Privacy Center (updated 1 July 2026). “How long do you store my data?” https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data. Accessed 2026-10-09.↩︎

  6. Anthropic Privacy Center (updated 16 March 2026). “Is my data used for model training?” https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training. Accessed 2026-10-09.↩︎

  7. OpenAI Help Center. “How are files vs chats retained?” https://help.openai.com/en/articles/8983778-how-are-files-vs-chats-retained ; OpenAI. “Your data” (API), https://developers.openai.com/api/docs/guides/your-data. Accessed 2026-10-09.↩︎

  8. The New York Times Co. v. Microsoft Corp. et al., No. 1:23-cv-11195-SHS (S.D.N.Y.), now part of In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143. Preservation order of 13 May 2025 (Wang, M.J.). Order text: https://cdn.arstechnica.net/wp-content/uploads/2025/06/NYT-v-OpenAI-Preservation-Order-5-13-25.pdf. OpenAI’s explanation of scope: “How we’re responding to The New York Times’ data demands”, https://openai.com/index/response-to-nyt-data-demands/ (updated 22 October 2025). Accessed 2026-10-09.↩︎

  9. Willison, S. (23 October 2025). “OpenAI no longer has to preserve all of its ChatGPT data, with some exceptions.” https://simonwillison.net/2025/Oct/23/openai-no-longer-has-to-preserve/ ; Engadget (October 2025), https://www.engadget.com/ai/openai-no-longer-has-to-preserve-all-of-its-chatgpt-data-with-some-exceptions-192422093.html. The termination order was entered on 9 October 2025 and ended the duty as of 26 September 2025. Accessed 2026-10-09.↩︎

  10. Bloomberg Law (5 January 2026). “OpenAI Must Turn Over 20 Million ChatGPT Logs, Judge Affirms.” https://news.bloomberglaw.com/ip-law/openai-must-turn-over-20-million-chatgpt-logs-judge-affirms ; ABA Journal, https://www.abajournal.com/news/article/chatgpt-creator-must-turn-over-20m-chat-logs-in-copyright-litigation-federal-judge-says. Accessed 2026-10-09.↩︎

  11. OpenAI (17 July 2025). “Introducing ChatGPT agent.” https://openai.com/index/introducing-chatgpt-agent/. Accessed 2026-10-09.↩︎

  12. BGR (2026). “OpenAI is discontinuing its ChatGPT Atlas browser.” https://www.bgr.com/2212396/openai-chatgpt-browser-atlas-discontinued/. Accessed 2026-10-09.↩︎

  13. Fast Company (30 July 2025). “Google is indexing ChatGPT conversations.” https://www.fastcompany.com/91376687/google-indexing-chatgpt-conversations ; Search Engine Journal, https://www.searchenginejournal.com/openai-is-pulling-shared-chatgpt-chats-from-google-search/552671/. Accessed 2026-10-09.↩︎

  14. Google. “Gemini Apps Privacy Hub.” https://support.google.com/gemini/answer/13594961. Accessed 2026-10-09.↩︎

  15. Google (January 2026). “Personal Intelligence in the Gemini app.” https://blog.google/innovation-and-ai/products/gemini-app/personal-intelligence. Accessed 2026-10-09.↩︎

  16. Google Cloud. “Data governance” (generative AI). https://docs.cloud.google.com/generative-ai-app-builder/docs/data-governance. Accessed 2026-10-09.↩︎

  17. Business Today (3 October 2025). “Perplexity opens Comet browser to all users.” https://www.businesstoday.in/amp/technology/news/story/perplexity-opens-comet-browser-to-all-users-adds-background-assistant-for-max-subscribers-496681-2025-10-03. Accessed 2026-10-09.↩︎

  18. The Decoder (August 2025). “Brave discovers a security flaw in Perplexity’s Comet browser.” https://the-decoder.com/brave-discovers-a-security-flaw-in-perplexitys-comet-browser/ ; Willison, S. (25 August 2025), https://simonwillison.net/2025/Aug/25/. Accessed 2026-10-09.↩︎

  19. Search Engine Journal (4 August 2025). “Cloudflare delists and blocks Perplexity from crawling websites.” https://www.searchenginejournal.com/cloudflare-delists-and-blocks-perplexity-from-crawling-websites/552899/. Accessed 2026-10-09.↩︎

  20. Mistral AI Help Center (updated 31 August 2026). Data usage and training policy. https://help.mistral.ai/en/articles/347617. Accessed 2026-10-09.↩︎

  21. Mistral AI (17 July 2025). “Le Chat dives deep.” https://mistral.ai/fr/news/le-chat-dives-deep. Accessed 2026-10-09.↩︎

  22. Mistral AI. “Changelog.” https://docs.mistral.ai/resources/changelogs. Accessed 2026-10-09.↩︎

  23. X Help Center. “About Grok.” https://help.x.com/en/using-x/about-grok. Accessed 2026-10-09.↩︎

  24. Michigan Public and NPR (9 July 2025). “Elon Musk’s AI chatbot, Grok, started calling itself ‘MechaHitler’.” https://www.michiganpublic.org/2025-07-09/elon-musks-ai-chatbot-grok-started-calling-itself-mechahitler. Accessed 2026-10-09.↩︎

  25. TechCrunch (20 August 2025). “Thousands of Grok chats are now searchable on Google.” https://techcrunch.com/2025/08/20/thousands-of-grok-chats-are-now-searchable-on-google/ ; Forbes Australia, https://www.forbes.com.au/news/innovation/xai-published-hundreds-of-thousands-of-grok-chatbot-conversations/. Accessed 2026-10-09.↩︎

  26. CNN (12 January 2026). “Indonesia and Malaysia block Grok.” https://edition.cnn.com/2026/01/12/business/indonesia-malaysia-grok-elon-musk-intl-hnk ; Ofcom (12 January 2026), https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/ofcom-launches-investigation-into-x-over-grok-sexualised-imagery ; The Register (26 January 2026), https://www.theregister.com/2026/01/26/ec_open_new_investigation_into/. Accessed 2026-10-09.↩︎

  27. PCWorld (October 2025). “Microsoft adds Microsoft 365 Premium, cuts Copilot Pro.” https://www.pcworld.com/article/2925962/microsoft-adds-microsoft-365-premium-cuts-copilot-pro.html. Accessed 2026-10-09.↩︎

  28. Anthropic (15 April 2025). “Claude takes research to new places.” https://claude.com/blog/research. Accessed 2026-10-09.↩︎

  29. Anthropic (1 May 2025, updated 3 June 2025). “Claude can now connect to your world.” https://claude.com/blog/integrations. Accessed 2026-10-09.↩︎

  30. Mistral AI (17 July 2025). “Le Chat dives deep.” https://mistral.ai/fr/news/le-chat-dives-deep. Accessed 2026-10-09.↩︎

  31. Willison, S. (16 February 2025). “Introducing Perplexity Deep Research” (with dates for Gemini and OpenAI). https://simonwillison.net/2025/Feb/16/introducing-perplexity-deep-research/ ; ARY News (17 February 2025) on Grok 3 and DeepSearch, https://arynews.tv/grok-3-musk-unveils-chatgpt-rival ; Topedia (June 2025) on Microsoft Researcher general availability, https://blog-en.topedia.com/2025/06/researcher-and-analyst-agents-moving-to-general-availability/. Accessed 2026-10-09.↩︎

  32. TechCrunch (26 March 2025). “OpenAI adopts rival Anthropic’s standard for connecting AI models to data.” https://techcrunch.com/2025/03/26/openai-adopts-rival-anthropics-standard-for-connecting-ai-models-to-data/. Accessed 2026-10-09.↩︎

  33. Linux Foundation (9 December 2025). “Linux Foundation announces the formation of the Agentic AI Foundation.” https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation ; Anthropic, https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation. Accessed 2026-10-09.↩︎

  34. Model Context Protocol. “Official MCP Registry.” https://registry.modelcontextprotocol.io/. Accessed 2026-10-09.↩︎

  35. Model Context Protocol. “servers-archived” (archived 29 May 2025). https://github.com/modelcontextprotocol/servers-archived. Accessed 2026-10-09.↩︎

  36. GitHub. “github-mcp-server.” https://github.com/github/github-mcp-server. Accessed 2026-10-09.↩︎

  37. Exa. “exa-mcp-server.” https://github.com/exa-labs/exa-mcp-server. Accessed 2026-10-09.↩︎

  38. Tavily. “tavily-mcp.” https://github.com/tavily-ai/tavily-mcp. Accessed 2026-10-09.↩︎

  39. Zencoder. “slack-mcp-server.” https://github.com/zencoderai/slack-mcp-server. Accessed 2026-10-09.↩︎

  40. OpenSanctions. “Open source software.” https://www.opensanctions.org/docs/opensource/. Accessed 2026-10-09.↩︎

  41. BurtTheCoder. “mcp-shodan” (community project, MIT). https://github.com/BurtTheCoder/mcp-shodan. Accessed 2026-10-09.↩︎

  42. OpenAI. “Memory and new controls for ChatGPT.” https://openai.com/index/memory-and-new-controls-for-chatgpt/ ; TechCrunch (10 April 2025), https://techcrunch.com/2025/04/10/openai-updates-chatgpt-to-reference-your-other-chats/. Accessed 2026-10-09.↩︎

  43. Anthropic (2025). “Bringing memory to Claude.” https://www.anthropic.com/news/memory. Accessed 2026-10-09.↩︎

  44. Google (January 2026). “Personal Intelligence in the Gemini app.” https://blog.google/innovation-and-ai/products/gemini-app/personal-intelligence. Accessed 2026-10-09.↩︎

  45. OpenAI Help Center. “How are files vs chats retained?” https://help.openai.com/en/articles/8983778-how-are-files-vs-chats-retained ; OpenAI. “Your data” (API), https://developers.openai.com/api/docs/guides/your-data. Accessed 2026-10-09.↩︎

  46. Google Cloud. “Data governance” (generative AI). https://docs.cloud.google.com/generative-ai-app-builder/docs/data-governance. Accessed 2026-10-09.↩︎

  47. Cohere (20 May 2026). “Introducing Command A+: Making sovereign agentic capabilities available to all.” https://cohere.com/blog/command-a-plus. Weights on Hugging Face under Apache 2.0, 218B total and 25B active parameters. Accessed 2026-10-12.↩︎

  48. Wu, S., Irsoy, O., Lu, S., et al. (2023). “BloombergGPT: A Large Language Model for Finance.” arXiv:2303.17564. https://arxiv.org/abs/2303.17564. Accessed 2026-10-09.↩︎

  49. Markets Media (2026). “Bloomberg introduces agentic AI to the Terminal.” https://www.marketsmedia.com/bloomberg-introduces-agentic-ai-to-the-terminal/. Accessed 2026-10-09.↩︎