Recent AI Model/{roduct Releases

Scope note: Here are several releases from Apr. 23–26, plus a few major “edge-of-window” items from Apr. 21–22 that are still highly relevant for this week’s genealogy work.

A. Named releases & features

• OpenAI: GPT-5.5 in ChatGPT: OpenAI’s Apr. 23 release note describes GPT-5.5 as its “smartest frontier model” for professional work, stronger at messy multi-step tasks, research, data analysis, document creation, spreadsheets, coding, retrieval, and PDF/report-grounded Q&A; GPT-5.5 Thinking is available to eligible paid plans, and GPT-5.5 Pro is available to Pro, Business, Enterprise, and Edu plans subject to workspace settings.

• DeepSeek: DeepSeek-V4-Pro and DeepSeek-V4-Flash Preview: DeepSeek’s V4 preview is live, open-sourced, and available through chat and API, with V4-Pro at 1.6T total / 49B active parameters, V4-Flash at 284B total / 13B active parameters, 1M-token context, Thinking and Non-Thinking modes, and OpenAI ChatCompletions plus Anthropic API compatibility.

• DeepSeek: deepseek-chat and deepseek-reasoner retirement notice: DeepSeek says deepseek-chat and deepseek-reasoner will be fully retired after July 24, 2026, 15:59 UTC, and are currently routing to DeepSeek-V4-Flash non-thinking/thinking modes.

• OpenAI: Fast answers in ChatGPT: OpenAI’s Apr. 22 release adds faster responses for common information-seeking questions across web, iOS, and Android for logged-in and logged-out users on all plans, but these answers do not use past chats or memory and can be disabled in Personalization settings.

• Google: gemini-embedding-2 GA: Google’s Gemini API changelog says gemini-embedding-2 became generally available on Apr. 22, making it the newest production embedding model for retrieval, clustering, similarity search, and document-indexing style workflows.

• Google: Deep Research agent previews: Google released deep-research-preview-04-2026 and deep-research-max-preview-04-2026 on Apr. 21, adding collaborative planning, visualization support, MCP server integration, and File Search, with one version optimized for speed and the other for maximum comprehensiveness.

• OpenAI: ChatGPT Images 2.0 and images with thinking: OpenAI’s Apr. 21 release introduced a new ChatGPT image-generation model on all ChatGPT plans, plus “images with thinking” for paid plans using select Thinking and Pro models so image outputs can be planned and refined before generation.

• Qwen: Qwen3.6-27B open-weight model: Qwen’s Hugging Face model card describes Qwen3.6-27B as the first open-weight Qwen3.6 variant, a 27B-parameter model with 262,144-token native context extendable to 1,010,000 tokens, text/image/video input, thinking mode by default, tool-calling capabilities, and stronger agentic coding and repository-level reasoning.

• Anthropic / Claude: no new Apr. 23–26 release note found: Anthropic’s official Claude release notes show no Apr. 23–26 entries; the latest April entries remain Claude Design on Apr. 17 and Claude Opus 4.7 on Apr. 16, just outside this briefing window.

• xAI / Grok: no new Apr. 23–26 release note found: xAI’s official API release notes show no Apr. 23–26 entries; the latest April entry visible is Apr. 15, when Speech to Text became generally available for 25 languages with batch and streaming modes.

• Meta / Llama and Mistral: no qualifying Apr. 23–26 official release found: Meta’s AI blog page did not show a qualifying recent Llama announcement for Apr. 23–26, and the Mistral changelog check did not surface a qualifying Apr. 23–26 model/product launch (AI at Meta Blog, Mistral Docs).

B. Implications for genealogists this week

The biggest practical shift is long-context research work becoming more ordinary. GPT-5.5 is aimed at multi-step professional research and document-grounded work, while DeepSeek-V4 and Qwen3.6-27B both emphasize very large context windows; for genealogists, that means you can increasingly place a whole research packet into one working session: deed abstracts, census extracts, county histories, tax lists, prior notes, and a draft proof argument, then ask for conflict analysis, missing-record suggestions, and a revised research plan tied to the evidence (OpenAI Help Center, DeepSeek API Docs, Hugging Face).

The second shift is better retrieval and personal research-library organization. Google’s gemini-embedding-2 GA matters less as a chat feature and more as infrastructure: it can power smarter search across Zotero notes, transcriptions, locality guides, surname files, and research logs, especially where keyword search misses variant spellings or conceptually similar records.

The third shift is visual and teaching output improving quickly. ChatGPT Images 2.0 with “images with thinking” is especially useful for genealogy educators and bloggers because it can plan before generating, which is exactly what you want for record-type diagrams, land-record teaching visuals, migration-path illustrations, and step-by-step classroom aids rather than generic decorative images.

C. Plug-and-play micro-workflows

1. GPT-5.5 evidence-to-plan workflow
   Paste a research-log excerpt, a census timeline, and 3–5 conflicting clues. Ask: “Create a prioritized research plan for proving or disproving this relationship. Separate direct evidence, indirect evidence, negative evidence, and next-record targets.” This uses GPT-5.5’s stated strength in messy multi-step professional research and document-grounded work.

2. DeepSeek-V4 long-county-packet review
   Put a county history chapter, deed abstracts, tax-list notes, and probate snippets into one long-context session. Ask DeepSeek-V4-Pro: “Identify every person, place, land description, date, officeholder, and possible kinship clue; then produce a FAN-club table and unresolved-question list.” DeepSeek says V4 supports a 1M context window and Thinking / Non-Thinking modes.

3. DeepSeek-V4 retirement-proof workflow audit
   If you use any saved scripts, API workflows, or third-party tools that call deepseek-chat or deepseek-reasoner, check them now and change the model name to deepseek-v4-pro or deepseek-v4-flash. DeepSeek states the older routes retire after July 24, 2026, and the new V4 API is available today.

4. Gemini embedding research-library index
   Use gemini-embedding-2 to index Zotero Better Notes exports, locality guides, surname studies, and old research logs, then search by meaning rather than exact wording. Try queries like “records that might explain disappearance after 1890” or “evidence of land ownership before statehood.” Google lists gemini-embedding-2 as GA as of Apr. 22.

5. Google Deep Research locality packet
   Use the new Gemini Deep Research preview for a focused locality question: “Build a research packet for Richland County, Ohio land and probate records, 1815–1850, with repositories, record gaps, maps, and a suggested search order.” Google says the Apr. 21 Deep Research previews add collaborative planning, visualization support, MCP server integration, and File Search.

6. ChatGPT Images 2.0 teaching diagram
   Ask: “Create a clean classroom diagram showing how a metes-and-bounds deed becomes an abstract, then a plat, then a neighborhood/FAN-club clue map. Use readable labels and no fictional names.” ChatGPT Images 2.0 is available on all ChatGPT plans, and images with thinking is available on paid plans with select Thinking and Pro models.

7. Qwen3.6-27B local/open-weight transcription assistant
   For a privacy-conscious workflow, run Qwen3.6-27B where available through open-weight tooling and feed it page images or text extracts for first-pass transcription cleanup, table reconstruction, or record-type classification. Qwen’s model card says it is the first open-weight Qwen3.6 variant, supports text/image/video input, and has long-context capabilities.

8. Fast answers for quick blog fact checks
   Use ChatGPT Fast answers only for low-stakes factual lookups like “What county was formed from what parent county?” or “What does et al. mean in a deed index?”, then verify against your locality guide before publishing. OpenAI says Fast answers are designed for common information-seeking questions and do not use memory or past chats.

9. GPT-5.5 proof-argument polishing pass
   Draft your conclusion first, then ask GPT-5.5: “Improve this genealogical proof argument without adding unsupported claims. Flag every sentence that needs a citation, every inference that needs explanation, and every possible alternative interpretation.” OpenAI positions GPT-5.5 for professional knowledge work, document creation, research, checking outputs, and polished artifacts.

10. DeepSeek-V4 vs. GPT-5.5 cross-check
    Give the same evidence summary to GPT-5.5 and DeepSeek-V4-Pro, then ask each to list “the three strongest arguments against my conclusion.” Compare disagreements before updating your research log. GPT-5.5 is framed for high-accuracy professional work, while DeepSeek-V4-Pro is positioned as a strong open-source reasoning and long-context model (OpenAI Help Center, DeepSeek API Docs).