Cycle on Jun 22, 2026

I will begin by listing the open threads and reading the existing notes for this research topic to understand what has been established and what needs attention in this cycle.

I will now read the existing notes for this topic to understand what findings have already been recorded.

Very interesting. The existing notes show a highly detailed and concrete reality in mid-2026: - A major structural shift with financial intelligence market giants FactSet, S&P Global, and PitchBook unbundling their data via MCP (Model Context Protocol). - Sacra unbundling private market intelligence into programmatic MCP servers. - Claude Fable 5, Mythos 5, and the Amazon-triggered geopolitical crackdown (export control directive, etc.). - Google DeepMind's Interactions API for Gemini Deep Research. - OpenAI's o3-deep-research and GPT-5.4 pivot. - Perplexity's WANDR benchmark and deep research. - German court ruling on AI liability (Case No. 26 O 869/26). - Anysphere (Cursor) being acquired by SpaceX for $60 billion (expected close in Q3 2026). - Anthropic having John Jumper leading scientific/biological AI models. Let's do some search to see if any of our watches have fired or if there's new information on: 1. Anthropic restoring access to Fable 5 and Mythos 5, or export controls. 2. SpaceX acquisition of Cursor closing. 3. Perplexity releasing WANDR. 4. Anthropic launching a scientific model under Jumper. 5. New developments in Gemini Deep Research Max GA or MCP servers. 6. Let's do general searches on what is happening in the AI research tools market in late June 2026. Let's do a web search first. I will search for Anthropic, John Jumper, Claude Fable 5, Mythos 5, SpaceX Cursor, Perplexity WANDR, and any other relevant terms to see if there are updates.

Interesting, John Jumper joining Anthropic is a major recent event (June 2026)! Let's search about this more. Let's see if there's any official release yet. The watch description is: "Monitor when Anthropic officially releases a scientific, biological, or chemistry-focused AI model or feature (e.g., an AlphaFold competitor or Claude Science tools) developed under the leadership of John Jumper. Fires when Anthropic announces its first major scientific, biological, or chemistry-focused AI model or feature release." He just joined in mid-June 2026 (a few days ago!). So no model has been released *yet*, but let's check to make sure. Let's check the other watches. "SpaceX Acquisition of Cursor Closes" -> expected Q3 2026. Today is June 22, 2026. Let's search for "SpaceX Anysphere Cursor acquisition" to see if there are updates.

Wow, SpaceX agreed to buy Cursor parent Anysphere for $60 billion on June 16, 2026! That's a huge deal. The transaction is expected to close in Q3 2026, so it's not closed yet. Let's do a search on: "Perplexity WANDR benchmark public" or "WANDR benchmark GitHub" or "Perplexity Wide Research benchmark" to see if it has been released yet.

Let's fetch the Perplexity research article "Rethinking Search as Code Generation" to see if there is any mention of WANDR and whether they released it, or what is happening there. Let's do a search on that first to find the URL and fetch it. Or we can fetch the URL `https://research.perplexity.ai/articles/rethinking-search-as-code-generation`. Let's fetch it.

Ah, the fetch failed with a timeout. Let's try searching for "Rethinking Search as Code Generation" text on Google or on HN to see what we can find. Let's do a web search.

Let's do a search on Hacker News for any stories about "WANDR" or "Search as Code" to find more detail. Let's use `hn_search` for "Search as Code" or "WANDR".

Let's read the comments for the HN story "Rethinking search as code generation" (id: 48372547) to see what practitioners are saying about WANDR or Search as Code!

These are excellent comments. They speak about the transition of search from iterative tool-calling loops to code-generation-based search, where the agent writes a Python script that executes search queries, filters the results, and returns them. This is "Search as Code" (SaC), first introduced by Perplexity in late 2025/early 2026. Let's do a search on the current state of "Search as Code" and Perplexity's WANDR benchmark, and see if there are any other recent papers or developments. Also, let's search specifically about: - Valyu, Parallel, Exa and Draco benchmark (from the medium article "AI DeepResearch APIs in 2026"). - Let's search for "AI DeepResearch APIs in 2026" or "Draco benchmark" or "Valyu" to see what these are. This seems very relevant to the market map of research tools. Let's do a search!

This is fantastic! Let's do a search on "Valyu AI" and "Parallel AI" or "Parallel Deep Research" or "Valyu DeepResearch" to see what these startups are. They seem to be specialized deep research API providers or research tools that are beating Perplexity on the DRACO benchmark in 2026! Let's search.

This is incredibly rich! Let's fetch the Medium article "AI DeepResearch APIs in 2026" (URL: `https://medium.com/@unicodeveloper/ai-deepresearch-apis-in-2026-f6d89ca0c17d`) and the Reddit post "Deepresearch API comparison 2026" (URL: `https://www.reddit.com/r/AI_Agents/comments/1s7rteo/deepresearch_api_comparison_2026/`) to get exact numbers, quotes, features, pricing, and details about the players. This will directly feed into our market map and positioning for Hey, Lefty. Let's do a web fetch on the Medium article first.

This is incredibly valuable and concrete! Let's analyze this article: - We have a clear breakdown of the **Deep Research API Category** in mid-2026. - The **DRACO (Deep Research Accuracy, Completeness, and Objectivity)** benchmark is the industry standard for evaluating deep research APIs. It is fully open-source and published on Hugging Face (by Perplexity AI) in early 2026. - The top-tier deep research APIs are: 1. **Valyu DeepResearch (Heavy)**: 72.7% accuracy / ~$2,500 CPM ($2.50 per task/query). It features proprietary data access (SEC filings, academic publishers, clinical databases) and is the most cost-effective premium option. 2. **Perplexity Deep Research (driven by Claude Opus 4.6)**: 70.5% accuracy / ~$5,500 CPM ($5.50 per task/query). Strong brand, high cost. 3. **You.com Research (exhaustive)**: 52.9% accuracy / $450 CPM ($0.45 per task). Best value play for mid-tier. 4. **Parallel Ultra8x**: 50.8% accuracy / ~$2,400 CPM ($2.40 per task). Raised $230M, known for speed and data coverage, but underperforming on accuracy relative to price. 5. **Tavily Research (pro)**: 39.1% accuracy / ~$1,060 CPM ($1.06 per task). Acquired by Nebius in February 2026. Opaque roadmap, hard to recommend on price-to-accuracy. 6. **Exa Deep Reasoning**: 25.9% accuracy / $15 CPM ($0.015 per task). Neural indexing and web retrieval, cheapest option, used as a pre-filter. Let's do a search on the Reddit post "Deepresearch API comparison 2026" or search for "reddit deepresearch API comparison 2026" to get the exact URL and fetch it, as it will likely have more details about the Fast/Standard/Heavy pricing for Valyu and others.

Let's fetch the Reddit post URL: `https://www.reddit.com/r/AI_Agents/comments/1s7rteo/deepresearch_api_comparison_2026/` to read the practitioner comparisons. This is incredibly valuable for our market map.

Ah, Firecrawl does not support Reddit. That is totally fine, the search snippet and the Medium article are already extremely detailed. Let's do a web search to see if we can find other mentions of Valyu, Parallel, and other deep research APIs, and what their specific features are. Let's search for "Valyu DeepResearch" to see what its features, deliverables, human-in-the-loop, etc. are. Let's do a web search.

Let's fetch the Medium article "9 AI DeepResearch Features That You Should Be Using" (URL: `https://medium.com/@unicodeveloper/9-ai-deepresearch-features-that-you-should-be-using-with-examples-d1e4d2683453`) or search for details of the Valyu DeepResearch API to see what features it has. This is extremely relevant. Let's do a web search.

Let's search for "9 AI DeepResearch Features" on Google.

Let's try to fetch the Medium article `https://medium.com/@unicodeveloper/9-ai-deepresearch-features-that-you-should-be-using-with-examples-d1e4d2683453` to get these 9 features. It will tell us exactly what features are key in modern 2026 deep research systems. This will be incredibly useful for Hey, Lefty's product map and defining its competitive advantage.

This is extremely detailed and provides an absolute goldmine of information! Let's review what we have: 1. **Valyu DeepResearch API**: - **Positioning**: Accuracy-first, enterprise-grade deep research API. Grounded in specialized and proprietary datasets. - **Accuracy**: 72.7% accuracy on DRACO, outperforming Perplexity Deep Research (70.5%), Claude Opus 4.6 (59.8%). - **Cost**: Usage-based, fixed-price per task: Fast ($0.10/task), Standard ($0.50/task), Heavy ($2.50/task), Max ($15.00/task). This is equivalent to ~$2,500 CPM for Heavy. - **Key Features**: - File inputs (up to 10 files, 100MB total - PDFs, CSVs, images, docs) incorporated into research. - Multiple deliverables/output formats (PDF, Word, Excel, PPTX, Markdown, JSON) from a single task. - Source Control: Pin/restrict search to specific categories: `web`, `academic` (ArXiv, PubMed, BioRxiv, Clinical trials, FDA labels, WHO, NIH, Wikipedia), `finance` (SEC filings, financial statements, prices, economic indicators, prediction markets), `patent` (USPTO), `transportation`, `politics`, `legal` (UK court judgments, legislation). - Custom JSON Schemas: Structured output. - Webhooks: Async execution (tasks take 5-15 mins). - Full source traceability with inline citations. - Sandboxed Python code execution (generating charts, tables). - Webpage screenshots. - Human-in-the-loop: Pause at up to 4 checkpoints (plan review, source review, etc.) to steer research in real-time. - Inline chart generation (renders React vs Vue vs Svelte, Semiconductor Foundry Analysis). 2. **Perplexity Deep Research**: - **Positioning**: High-speed consumer search and synthesis, transitioning to "Search as Code" (SaC). - **Core Technology**: "Search as Code" (SaC) generation (the agent writes a Python script in a sandbox to execute searches, filter results, and deduplicate context instead of looping iterative tool calls). - **Accuracy**: 70.5% on DRACO. - **Cost**: High-cost API at ~$5,500 CPM (using Claude Opus 4.6). Consumer subscription is $20/month, with a $200/month Max tier. - **Key Features**: High-speed web search & synthesis, inline citations, WANDR benchmark (designed for "wide research" tasks). 3. **Google DeepMind / Gemini Deep Research**: - **Positioning**: Multi-step reasoning optimized for long-horizon agentic workflows. - **Core Technology**: Interactions API (server-side state management, multi-step reasoning, long-horizon agents). - **Data**: Google's index, Google Scholar, specialized academic search integration. 4. **Sacra**: - **Positioning**: Premium private market research as agent-ready infrastructure. - **Core Technology**: Programmatic MCP (Model Context Protocol) servers. Unbundles research reports, private company financials, and profiles into programmatic endpoints for AI agents. 5. **Where Does Hey, Lefty Fit? (Market Map & Positioning)**: - Let's read the existing note on `market-map-positioning-hey-lefty` to understand how we can refine and extend it with these exact 2026 developments (Valyu, Parallel, DRACO benchmark, Search as Code, etc.). I will first read the existing note using `read_notes` or I can write a note updating it. Let's list the notes. I have the preview of `market-map-positioning-hey-lefty` but let's read the full note. Wait, I don't have a specific tool to read a single note's full text other than `read_notes` which returns previews, but I can read all notes if needed or I can just write a note updating `market-map-positioning-hey-lefty` to include these detailed insights. Let's check `read_notes` parameters. The `read_notes` tool returns slug, title, preview, updated_at, and source_count for each note. It doesn't return the full body. That's fine, I can write a note with the slug `market-map-positioning-hey-lefty` to update it. Wait, let's look at the open threads and see if we can resolve any. - `019ea05f-eaaf-7afb-bb51-55b5f00e17d3`: "Institutional Research Platforms Launch Public MCP Servers" -> "Fires when a major institutional competitor to Sacra or developer data provider launches a public or partner-accessible MCP server." We have a prior finding: "Institutional MCP Financial Data Servers: FactSet, S&P Global, and PitchBook Unbundle". Let's check if this is already resolved or if we should resolve it. It is still listed in `list_threads`. Let's resolve it if we have the evidence. We can resolve it by noting that FactSet, S&P Global, and PitchBook have unbundled their data via MCP. Let's do that! - Wait, let's look at the other threads: - `Anthropic Launches Scientific or Biological AI Model under John Jumper` (id: `019ee351-7287-781b-9912-104ad72fc9bf`). John Jumper joined Anthropic on June 16, 2026. This is a very recent event! No model has been launched yet, but we should definitely write a finding about John Jumper joining Anthropic to lead their scientific AI efforts, which represents a massive threat to Google DeepMind and a huge boost for Anthropic's scientific research capabilities. - `SpaceX Acquisition of Cursor Closes` (id: `019ed3e0-ceef-79c4-afdb-1ca1e2aa1745`). The acquisition was announced on June 16, 2026, for $60 billion. It is expected to close in Q3 2026. This is a dated catalyst. We can record an event for it if it closed, but it hasn't closed yet. We can update our findings. - `Perplexity Releases WANDR Benchmark to the Public` (id: `019eceb8-a74b-7495-81db-b5b2bbff4c86`). Let's see if Perplexity has released WANDR to the public. In the search results, it says "We will publish the WANDR benchmark in the coming weeks" in their "Rethinking Search as Code Generation" paper (published around early June 2026). They have released the DRACO benchmark, which is fully open source on Hugging Face. Let's verify if WANDR is open yet or not. Since the paper says "coming weeks" in June, it might not be fully open yet, or it is. Let's write a note on Perplexity's "Search as Code" (SaC) and WANDR benchmark. Let's organize our notes: 1. **market-map-positioning-hey-lefty**: We will update this note to incorporate the mid-2026 deep research API market landscape. We will present a highly structured market map showing where Valyu, Perplexity, You.com, Parallel, Exa, Gemini Deep Research, OpenAI, and Sacra fit, and where Hey, Lefty fits as a model-agnostic, multi-source orchestrator that integrates these APIs and MCP servers. 2. **perplexity-deep-research-consumer-agent**: We will update this note to focus on "Search as Code" (SaC) and the WANDR and DRACO benchmarks. 3. **academic-ai-research-scholar-vs-synthesis**: We will update this to cover John Jumper leaving Google DeepMind to join Anthropic in June 2026, a major structural shift in the scientific AI space. Let's write these notes. First, let's write `market-map-positioning-hey-lefty`. I will construct a comprehensive, data-driven market map based on the DRACO benchmark results and features from May/June 2026. Wait, let's double check if there are other findings. Let's write the `market-map-positioning-hey-lefty` note. Let's make sure it contains verbatim quotes and links from our sources. The sources we have are: - Medium Article: "AI DeepResearch APIs in 2026" by unicodeveloper (May 5, 2026) -> `https://medium.com/@unicodeveloper/ai-deepresearch-apis-in-2026-f6d89ca0c17d` - Medium Article: "9 AI DeepResearch Features That You Should Be Using (With Examples)" by unicodeveloper (2026) -> `https://medium.com/@unicodeveloper/9-ai-deepresearch-features-that-you-should-be-using-with-examples-d1e4d2683453` - Perplexity Research: "Rethinking Search as Code Generation" (June 2026) -> `https://research.perplexity.ai/articles/rethinking-search-as-code-generation` - SpaceX acquisition of Cursor (Anysphere) on June 16, 2026 -> `https://techcrunch.com/2026/06/16/spacex-to-acquire-cursor-for-60b-in-stock-days-after-blockbuster-ipo/` - John Jumper joining Anthropic (June 16, 2026) -> `https://thenextweb.com/news/john-jumper-nobel-deepmind-leaves-anthropic-alphafold` Let's write the updated `market-map-positioning-hey-lefty` note. I will use the `write_note` tool. Let's formulate the body. No YAML frontmatter. First line is `# Market Map & Positioning: Where Does Hey, Lefty Fit?`. Let's draft it.

Now let's write a note focusing on Perplexity's "Search as Code" (SaC) paradigm and the WANDR and DRACO benchmarks. We will reuse the slug `perplexity-deep-research-consumer-agent` since that is the existing note tracking Perplexity's deep research capabilities. Let's make sure it contains verbatim quotes and links from our sources. The sources we have are: - Perplexity Research: "Rethinking Search as Code Generation" -> `https://research.perplexity.ai/articles/rethinking-search-as-code-generation` - Threads post by Perplexity -> `https://www.threads.com/@perplexity/post/DZDWen0GoRN/we-tested-search-as-code-on-deep-research-dsqa-browse-comp-hle-and-wide/` - Hacker News discussion -> `https://news.ycombinator.com/item?id=48372547` Let's write this note using `write_note` tool. Let's formulate the body. No YAML frontmatter. First line is `# Perplexity Deep Research: Search as Code and the WANDR Benchmark`. Let's draft it.

Now let's write a note focusing on the scientific AI research and academic tools, specifically updating `academic-ai-research-scholar-vs-synthesis` to cover John Jumper leaving Google DeepMind to join Anthropic in June 2026. This is a massive event, as Jumper co-created AlphaFold and won the 2024 Nobel Prize in Chemistry. His departure represents a major talent loss for Google DeepMind and a huge win for Anthropic as they construct wet labs and prepare specialized scientific models. Let's write this note using `write_note` tool. Let's formulate the body. No YAML frontmatter. First line is `# Academic & Scientific AI Research: The Sovereignty War and John Jumper's Move to Anthropic`. Let's draft it.

This is beautiful! Let's check if there are other open threads we can resolve. - `019ea05f-eaaf-7afb-bb51-55b5f00e17d3`: "Institutional Research Platforms Launch Public MCP Servers" -> "Fires when a major institutional competitor to Sacra or developer data provider launches a public or partner-accessible MCP server." Let's double check if we can resolve this thread. The thread description says "Monitor when major financial or market data platforms (e.g., PitchBook, S&P Global, FactSet) release developer-facing or public MCP servers." We already have a note `institutional-mcp-financial-data-servers` which says "FactSet, S&P Global, and PitchBook Unbundle" and "Institutional MCP Financial Data Servers". Let's resolve this thread as it was fully addressed by that finding! I will call `resolve_thread` for it. Let's find its ID from `list_threads`: `019ea05f-eaaf-7afb-bb51-55b5f00e17d3`. Let's resolve it.

Let's check if there are other threads we can resolve or record events for. - `Anthropic Launches Scientific or Biological AI Model under John Jumper` (id: `019ee351-7287-781b-9912-104ad72fc9bf`). He just joined Anthropic on June 16, 2026. This is an ongoing watch or a one-shot watch. Let's look at the description: "Monitor when Anthropic officially releases a scientific, biological, or chemistry-focused AI model or feature... developed under the leadership of John Jumper. Fires when Anthropic announces its first major scientific, biological, or chemistry-focused AI model or feature release." Since he just joined, the model has not been released yet, so the watch should remain open. We can, however, record a progress event or just keep it open. Since it's a one-shot watch, we keep it open until the model is released. - `SpaceX Acquisition of Cursor Closes` (id: `019ed3e0-ceef-79c4-afdb-1ca1e2aa1745`). It was announced on June 16, 2026, and is expected to close in Q3 2026. So it should remain open. - `Perplexity Releases WANDR Benchmark to the Public` (id: `019eceb8-a74b-7495-81db-b5b2bbff4c86`). Let's search if the WANDR benchmark is officially open sourced or released yet. We saw in the June 2026 paper: "We will publish the WANDR benchmark in the coming weeks." So it hasn't been officially released on GitHub as a public repo yet, but we are watching it. Let's register a new watch if there is anything forward-looking. - Valyu DeepResearch API was launched in mid-2026. Let's see if there are other upcoming deep research launches or benchmarks we should watch. - Wait, what about the Gemini Deep Research Max GA or OpenAI GPT-5.4/o3-deep-research updates? We already have ongoing watches for those. - Let's register a watch for any upcoming updates to the DRACO benchmark or new deep research API standings (e.g., DRACO v2 or similar). Or we can register a watch for when Valyu DeepResearch launches an official enterprise MCP server, or when You.com releases an API update. - Actually, we have a very solid set of watches. Let's check if we can register a watch for the DRACO benchmark updates. - Wait, let's look at the instructions. "Call record_watch for anything forward-looking this cycle surfaced — a 'what to watch' that has a future trigger." Let's keep the existing watches and see if there are any other actions. Let's double-check the rules: - "Every substantive finding MUST cite at least one real source you actually consulted THIS cycle — a web page you fetched, an HN story or comment you read, etc." Yes, we fetched the two Medium articles and searched the web/HN. - "One finding = ONE coherent subject tracked under one stable slug across cycles... NEVER write a note that summarizes the cycle... If you drilled five stories, that's up to five separate findings, not one digest." Yes, we updated three distinct existing notes (`market-map-positioning-hey-lefty`, `perplexity-deep-research-consumer-agent`, `academic-ai-research-scholar-vs-synthesis`). - "To link ANOTHER finding from a note body, write its slug as a wikilink — [[its-slug]]". Yes, we linked `[[sacra-premium-private-market-mcp-data]]` and `[[institutional-mcp-financial-data-servers]]` in `market-map-positioning-hey-lefty`. - "Call finish_cycle to end the cycle." This is extremely clean and compliant. I am ready to finish the cycle. Let's call `finish_cycle`.

I have successfully completed the research cycle for topic **019e8498-f497-7eb3-9d41-64bb48fe1e5d**. ### Summary of Actions Taken: 1. **Resolved Open Thread**: Marked the thread *"Institutional Research Platforms Launch Public MCP Servers"* (`019ea05f-eaaf-7afb-bb51-55b5f00e17d3`) as resolved, referencing our existing finding detailing the unbundling of FactSet, S&P Global, and PitchBook via MCP. 2. **Updated Standing Findings**: * `[[market-map-positioning-hey-lefty]]`: Extensively updated to reflect the mid-2026 deep research landscape, incorporating the **DRACO (Deep Research Accuracy, Completeness, and Objectivity)** benchmark standings. Highlighted the emergence of **Valyu DeepResearch** as the accuracy leader (72.7% accuracy / $2.50 per task) and detailed how **Hey, Lefty** fits as a model-agnostic, multi-source orchestrator. * `[[perplexity-deep-research-consumer-agent]]`: Updated to analyze Perplexity's transition to the **Search as Code (SaC)** paradigm, explaining how sandboxed Python execution cuts token context by up to 85% on long-horizon tasks, and detailing the **WANDR (Wide Research)** and **DRACO** benchmarks. * `[[academic-ai-research-scholar-vs-synthesis]]`: Updated to capture Nobel Laureate **John Jumper** leaving Google DeepMind to join **Anthropic** in June 2026 to lead their scientific/biological AI model initiatives. Linked this talent migration to the broader movement for **Sovereign AI** and open-weights initiatives like **Apertus**.