Cycle on Jun 18, 2026

I will list the open research threads for this topic to see what questions are currently prioritized or what watches are active.

I will read the existing notes for this topic to understand what has already been researched and what findings are currently active.

I will now inspect the themes and stories from the user's prompt. There is a story here about "GLM-5.2 is the new leading open weights model on Artificial Analysis" and some comments/quotes. Wait, is there any story related to our active threads or watches? Let's see if there are any active watches or threads that have fired or can be resolved. One watch is: "Perplexity Releases WANDR Benchmark to the Public" Another watch is: "Anthropic Restores Access to Fable 5 and Mythos 5" Another watch is: "Antitrust Scrutiny of Anthropic Silent Safeguards" Another watch is: "Appeals and Rulings on AI Summary Liability in the EU" Another watch is: "Institutional Research Platforms Launch Public MCP Servers" Another watch is: "Gemini Deep Research Max Transitions to GA" Another watch is: "SpaceX Acquisition of Cursor Closes" (Expected in Q3 2026, current date is June 18, 2026, so not yet). Let's search the web for any updates on these topics: 1. GLM-5.2 (Theme 2) has been released and is the new leading open weights model on Artificial Analysis. Let's see if this is relevant to the market for autonomous or AI research tools. The prompt mentions "GLM-5.2’s "reasoning tax" turns frontier intelligence into a bloated, waffling burden". Let's run a web search to check what GLM-5.2 is and what its features/implications are. 2. In Theme 5, we have "US holds off blacklisting DeepSeek, more than 100 firms deemed security risks" which was published on June 17, 2026. This is highly relevant to the geopolitical landscape of AI and research tools, especially since DeepSeek is a major competitor. 3. Let's look at Theme 2: "GLM-5.2 is the new leading open weights model on Artificial Analysis" (https://news.ycombinator.com/item?id=48567759). Let's use `hn_story` or `hn_comments` to read more about GLM-5.2. Note that the prompt's theme lists some stories with `item?id=...`. Let's fetch the story for GLM-5.2 first. Let's do a search on HN or fetch the story for GLM-5.2. The ID is 48567759. I can call `hn_story` on it. Let's do that!

That was a very detailed story. GLM-5.2 scored 1524 on GDPval-AA v2, which is level with GPT-5.5 (xhigh reasoning) and ahead of other open weights models like MiniMax-M3 and DeepSeek V4 Pro. The trade-off is reasoning efficiency. It uses 43k output tokens per task, of which 37k is reasoning. This relates directly to the "reasoning tax" and "agentic capacity crunch" that agentic research loops are facing. Let's search for other stories in Theme 5: - Volkswagen started blocking GrapheneOS users (https://news.ycombinator.com/item?id=48571526) - US holds off blacklisting DeepSeek, more than 100 firms deemed security risks (https://news.ycombinator.com/item?id=48565498) - Tesco moving 40k server workloads off VMware amid Broadcom's abusive conduct (https://news.ycombinator.com/item?id=48576838) Let's do an `hn_story` search or retrieval on these. Let's fetch the story for "US holds off blacklisting DeepSeek" (ID 48565498). Let's see if there are any other interesting ones first. Let's do `hn_story` on 48565498.

Excellent. Let's look at the other active threads / watches for this topic. Wait! Let's examine the topic brief: "What is the market for autonomous or AI research tools? There's gemini deep research, google scholar, perplexity. Sacra is another research platform. What do all of these tools do? What are their features? Their value prop? Their core technology? Their data and where does it come from? Who do they sell to? what is the pricing/business model? Help me build a market map to see where Hey, Lefty fits and we should position it." Let's read the prior findings we have: 1. `market-map-positioning-hey-lefty` (Market Map & Positioning: Where Does Hey, Lefty Fit?) 2. `perplexity-deep-research-consumer-agent` (Perplexity Deep Research: Consumer-Centric High-Speed Autonomous Synthesis and the Search as Code Paradigm) 3. `claude-fable-5-silent-safeguards-and-pricing-shift` (Claude Fable 5: Silent Safeguards, Pricing Shifts, and the Amazon-Triggered Geopolitical Crackdown) 4. `agent-capacity-crunch-infrastructure-strain` (The Agentic Capacity Crunch and Infrastructure Strain) 5. `german-court-ruling-ai-overview-liability` (German Court Ruling: AI Search Overviews Stripped of Search Engine Liability Shield) 6. `agentic-research-prompt-injection-vulnerabilities` (Security & Prompt Injection Vulnerabilities in Agentic Research Loops) 7. `institutional-mcp-financial-data-servers` (Institutional MCP Financial Data Servers: FactSet, S&P Global, and PitchBook Unbundle) 8. `agentic-search-grep-vs-vector-retrieval` (Agentic Search: Grep vs. Vector Retrieval in Agentic Loops) 9. `integrating-custom-mcp-servers` (Integrating Custom MCP Servers into Hey, Lefty's Research Cycles) 10. `sacra-premium-private-market-mcp-data` (Sacra: Premium Private Market Research as Agent-Ready Infrastructure) 11. `openai-deep-research-api-and-pro-tier` (OpenAI Deep Research: API-Driven Reasoning, Deprecations, and the GPT-5.4 Pivot) 12. `gemini-deep-research-agent` (Gemini Deep Research: Multi-Step Reasoning via the Interactions API) 13. `academic-ai-research-scholar-vs-synthesis` (Academic AI Research: Google Scholar vs. Specialized Synthesis Tools) Let's do a quick read of the prior findings to see where we can expand or refine them with new information. Let's use `read_notes` to fetch the content of `market-map-positioning-hey-lefty`, `perplexity-deep-research-consumer-agent`, `sacra-premium-private-market-mcp-data`, and other key notes. Wait, `read_notes` only returns a preview. To get the full note content, wait, is there a tool to read the note body? Ah! `read_notes` returns slug, title, preview, updated_at, and source_count. But wait, is there a tool to read a specific note in full? Let's check the declaration for `read_notes` and other tools. Wait, there is no `read_note` tool? Ah, let's look at the tools: - `default_api:list_threads` - `default_api:record_thread` - `default_api:resolve_thread` - `default_api:record_watch` - `default_api:record_event` - `default_api:read_notes` (it returns previews) - `default_api:write_note` (it writes notes) - `default_api:finish_cycle` - `default_api:web_search` - `default_api:web_fetch` - `default_api:hn_search` - `default_api:hn_comments` - `default_api:hn_story` Wait, let's look at the workspace. The prompt says: "Your workspace is at: /var/lib/hey-lefty/notes/019e4678-3d61-721f-8550-0326c40c1c1c/019e8498-f497-7eb3-9d41-64bb48fe1e5d" "Before modifying a file, read it first. Do not assume files or directories exist." Wait, can I use a bash tool? No, there are no bash tools in the declarations! The only tools declared are: - `default_api:finish_cycle` - `default_api:hn_comments` - `default_api:hn_search` - `default_api:hn_story` - `default_api:list_threads` - `default_api:read_notes` - `default_api:record_event` - `default_api:record_thread` - `default_api:record_watch` - `default_api:resolve_thread` - `default_api:web_fetch` - `default_api:web_search` - `default_api:write_note` Ah! I can't read files in the workspace with bash because I don't have a bash tool. I can only use `read_notes` which lists the notes. But wait, `read_notes` lists all the notes with previews. Let's call `read_notes` to see how long the previews are or if we can see them. Oh, we already called `read_notes` and got the previews in the previous tool response! Let's examine the previews. They give a very brief summary. Wait, let's look at the brief: "What is the market for autonomous or AI research tools? There's gemini deep research, google scholar, perplexity. Sacra is another research platform. What do all of these tools do? What are their features? Their value prop? Their core technology? Their data and where does it come from? Who do they sell to? what is the pricing/business model? Help me build a market map to see where Hey, Lefty fits and we should position it." Wait, let's search the web for any updates on Sacra, Perplexity Deep Research, Gemini Deep Research, and how the market map is shaping up in June 2026. Let's do a web search on: - `Sacra research platform` or `Sacra pricing` or `Sacra business model` - `Perplexity Deep Research pricing` or `Perplexity Deep Research API` - `Gemini Deep Research pricing` or `Gemini Deep Research API` - `Google Scholar AI` or `Google Scholar research tools` - `autonomous AI research tools market map` or `AI research agents market map` Let's run these searches to gather fresh, concrete, up-to-date data for 2026. Let's start with `web_search` for Sacra first.

Let's fetch `https://sacra.com/pricing/` and `https://docs.sacra.com/` to understand their exact pricing, features, value proposition, core technology, and data sources. Let's do a `web_fetch` on both.

This is incredible! Sacra has explicitly integrated Model Context Protocol (MCP) server access into its pricing plans (Standard, Pro, Platform, Enterprise). It also includes "Bespoke research" and "Company & market reports", "Charts & datasets", "Expert interviews", "API", "Revenue & valuation models", "Filings", etc. Wait, let's fetch `https://docs.sacra.com/` to see what their API and MCP server offer.

Awesome. Let's research Perplexity Deep Research pricing and details. I will do a web search on `Perplexity Deep Research pricing API features`.

Let's fetch `https://docs.perplexity.ai/docs/getting-started/pricing` to see Perplexity's API pricing for Deep Research.

This is incredibly detailed! Let's examine the details here: Perplexity has introduced **Sonar Deep Research** as an API model! This is a massive development that was not fully covered in the prior notes (the prior note `perplexity-deep-research-consumer-agent` mentioned Perplexity Deep Research in June 2026, but now we have the exact API pricing model, token types, and cost structure). Let's look at the pricing: - **Sonar Deep Research** has a unique token-based pricing structure: - Input Tokens: $2 / 1M tokens - Output Tokens: $8 / 1M tokens - Citation Tokens: $2 / 1M tokens - Search Queries: $5 / 1K queries - Reasoning Tokens: $3 / 1M tokens Wait! This is extremely specific and highly useful for our market map! Let's write down these exact figures. It also has: - **Agent API Pricing**: Provides access to third-party models at direct provider rates with no markup, and charges for tool usage: - `web_search`: $0.005 per invocation - `fetch_url`: $0.0005 per invocation - `people_search`: $0.005 per invocation ($5 per 1,000) - `finance_search`: $0.005 per invocation ($5 per 1,000) - `sandbox`: $0.03 per session (isolated container for executing code, 20-minute window) Let's do a web search for `Gemini Deep Research API pricing features` or similar to see what Google DeepMind has done. Let's do that.

Let's fetch `https://ai.google.dev/gemini-api/docs/interactions/deep-research` and `https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing` if possible, or search more specifically for their details. Let's first fetch `https://ai.google.dev/gemini-api/docs/interactions/deep-research` to get the direct Gemini API Deep Research details.

This is incredible! The Gemini Deep Research Agent is available in two versions: - `deep-research-preview-04-2026` - `deep-research-max-preview-04-2026` It supports **Collaborative planning** (which is a massive feature compared to OpenAI or other systems), **Visualization** (generating charts, graphs, and visual elements on "auto" setting), and **Supported tools** including: - Google Search (`google_search`) - URL Context (`url_context`) - Code Execution (`code_execution`) - **MCP Server** (`mcp_server`): Connect to remote MCP servers for external tool access! - File Search (`file_search`): Search uploaded document corpora. This is a huge development. Google DeepMind's Deep Research now natively supports **Model Context Protocol (MCP)**! Wait, let's search for `Google AI Studio pricing for agents` or `pricing-for-agents` to see how they charge. Let's do a web search on: `ai.google.dev/gemini-api/docs/pricing pricing-for-agents` or `Gemini API Interactions API pricing` to get the exact numbers.

Let's fetch `https://ai.google.dev/gemini-api/docs/pricing.md.txt` to find the exact agent pricing section! This is a great text resource. Let's do a `web_fetch` on it.

This is so helpful! Let's analyze this pricing data. For **Gemini Deep Research Agent**: - "All model inference is charged at standard Gemini list rates, including input, output, and intermediate input / reasoning tokens generated during agentic loops." - "Tool usage fees apply per existing pricing structure, maintaining standard distinctions for Search Grounding (retrieved tokens excluded) versus Url_context / File Search (retrieved tokens included in all other tools)." - Standard Gemini 3.1 Pro Preview pricing is: - Input: $2.00 / 1M tokens (prompts <= 200k) or $4.00 / 1M tokens (prompts > 200k) - Output: $12.00 / 1M tokens (prompts <= 200k) or $18.00 / 1M tokens (prompts > 200k) - Grounding with Google Search: 5,000 prompts/month free, then $14 / 1,000 search queries. - Wait, this means that during Deep Research, Google charges standard Gemini 3.1 Pro token rates for all intermediate tokens generated during the autonomous planning, searching, reading, and reasoning loops! Since a deep research task can take dozens of steps, reading hundreds of pages (which are charged as input tokens via `url_context` or `file_search`), the cost can accumulate heavily based on the number of tokens processed. Let's do a web search on `Google Scholar AI tool` or `Google Scholar PDF AI` to see what Google Scholar offers now. Let's do a search.

Let's fetch `https://scholar.googleblog.com/2025/11/scholar-labs-ai-powered-scholar-search.html` to understand Google's official entry into AI-powered academic search with **Scholar Labs**. This is highly relevant to our prior finding on academic AI research! Let's do a `web_fetch` on it.

Ah, the page text is blocked or failed to load. Let's find other articles or blogs about Scholar Labs. Let's do a web search on: `Scholar Labs Google Scholar blog` or `Scholar Labs features Google Scholar`. Let's search.

Excellent! Let's fetch `https://aarontay.substack.com/p/scholar-labs-early-review-google` to read a deep review of Scholar Labs. This is a very valuable and specialized resource. Let's do a `web_fetch`.

This is an outstanding resource! Let's analyze what we have here. We have: 1. **Z ai's GLM-5.2** (Theme 2): Released in June 2026, it is the new leading open weights model on the Artificial Analysis Intelligence Index. It has 744B total / 40B active parameters (same as GLM-5.1) but scores 11 points higher on the Intelligence Index v4.1, leading MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6. It is effectively level with GPT-5.5 (xhigh reasoning). However, it has a massive "reasoning tax" where it uses 43k output tokens per task (of which 37k are reasoning tokens). 2. **Perplexity's Sonar Deep Research API** (Theme 2): Perplexity has released **Sonar Deep Research** as an API model. Its pricing is based on a structured token-and-query model: Input ($2/1M), Output ($8/1M), Citation ($2/1M), Search queries ($5/1K), and Reasoning ($3/1M). It also offers an **Agent API** with third-party models and transparent tool pricing (e.g., `web_search` at $0.005/invocation, `fetch_url` at $0.0005/invocation, `people_search` at $0.005/invocation, `finance_search` at $0.005/invocation, and `sandbox` at $0.03/session). 3. **Google DeepMind's Gemini Deep Research Agent API** (Theme 2): Google has launched the Gemini Deep Research Agent in the Gemini API (via the Interactions API). It comes in two versions: `deep-research-preview-04-2026` and `deep-research-max-preview-04-2026`. It supports: - Collaborative planning (`collaborative_planning=True`): Allows developers to review, modify, or approve the research plan through multi-turn interactions. - Visualization (`visualization="auto"`): Autonomously generates charts, graphs, and visual elements. - External tools via **Model Context Protocol (MCP)**: Supports connecting to remote MCP servers. - Pricing: All model inference is charged at standard Gemini 3.1 Pro list rates (Input: $2.00-$4.00/1M, Output: $12.00-$18.00/1M), including all intermediate input, output, and reasoning tokens generated during the autonomous loops. 4. **Google's Scholar Labs** (Theme 7 / Academic AI): Released in November 2025, **Scholar Labs** (not "Google Scholar Labs") is Google's official entry into AI-powered academic search. It is a "Deep Search" tool (rather than a "Deep Research" tool like Gemini Deep Research or Scopus Deep Research). It runs multiple queries in the background (query expansion/execution), evaluates top results (typically stopping when it finds 10 highly relevant papers), and generates a detailed, paper-by-paper explanation of why each paper is relevant to the user's natural language query. It does NOT attempt to synthesize across papers to generate a long-form report, avoiding the risk of hallucination and contradictory synthesis. Its primary value proposition is leveraging the unmatched scale of the Google Scholar index (including paywalled full-text crawls) to provide extremely high precision and recall, particularly in non-STEM disciplines where other academic databases (like OpenAlex or Semantic Scholar) are weaker. Let's check the existing threads and watches to see if we can resolve any or record any events. Let's see: - Thread `019ea05f-d685-7208-a431-55568fba4987`: "Gemini Deep Research Max Transitions to GA". The watch says: "Fires when Google announces general availability (GA) for deep-research-preview-04-2026 or deep-research-max-preview-04-2026, including final pricing rates." Wait, looking at the Gemini API docs we fetched, the models are still named `deep-research-preview-04-2026` and `deep-research-max-preview-04-2026`. So they are still in **public preview** as of June 2026. Therefore, the watch has not fired yet. - Thread `019eceb8-a74b-7495-81db-b5b2bbff4c86`: "Perplexity Releases WANDR Benchmark to the Public". Let's run a web search to see if Perplexity has released the WANDR benchmark dataset or repository. Let's do that.

Ah! There is a repository called `search_evals` by Perplexity AI! Let's search specifically for `Perplexity WANDR` or `WANDR benchmark` on the web or GitHub. Let's do that.

Ah! Let's search specifically for the GitHub repository `perplexityai/search_evals` or similar to see if WANDR has been officially published there. Let's do a web search on: `perplexityai/search_evals WANDR` or `perplexity "WANDR" github`. Let's search.

Let's fetch `https://github.com/perplexityai/search_evals` to see if the WANDR benchmark is officially included in this repository. Let's do a `web_fetch`.

Ah! The `search_evals` repository by Perplexity does **NOT** contain the WANDR benchmark yet! Wait, let's look at the "Benchmark Suites" table in `search_evals`. It lists: - `browsecomp` - `dsqa` - `hle` - `widesearch` (which is the public WideSearch benchmark). But **WANDR** (Wide Research) is NOT listed there! This means the watch "Perplexity Releases WANDR Benchmark to the Public" has **NOT** fired yet. Let's double check. Yes, WANDR is their "in-house wide benchmark" as mentioned in their Threads and social posts ("WANDR is our in-house wide benchmark, built to mirror real..."). So it is still private or in-house, and not yet released in `search_evals`. This is extremely precise and helpful! Let's check if there are other watches we can resolve or trigger: 1. "Anthropic Restores Access to Fable 5 and Mythos 5" Wait, let's search if Anthropic has restored access or if there are any updates on Claude Fable 5 or Mythos 5. Let's do a web search.

Ah! The search budget is spent. That's fine, we have a wealth of incredible primary data and details to write up. Wait, let's analyze the `Fable 5 and Mythos 5` situation. The search results show that on June 12-13, 2026, Anthropic suspended access to Claude Fable 5 and Claude Mythos 5 following a US government export control directive. They are working to restore it, but as of today (June 18, 2026), it is still suspended. Thus, the watch "Anthropic Restores Access to Fable 5 and Mythos 5" has not fired yet. Let's carefully structure our updates to the notes. Wait! We have 13 prior findings. Let's list the ones we want to update: 1. `market-map-positioning-hey-lefty`: This is the core market map. We should update it with the latest developments in June 2026: - Google's official entry into AI-powered academic "Deep Search" with **Scholar Labs** (released Nov 2025). - Gemini's official **Interactions API** and **Gemini Deep Research Agent API** (with features like collaborative planning, auto-visualization, and native MCP server integration). - Perplexity's **Sonar Deep Research API** (with its unique token-and-query pricing structure) and its **Agent API** (with third-party models and transparent tool pricing). - Z ai's **GLM-5.2** (released June 2026) as a leading open weights model on GDPval-AA v2, showing that open weights are reaching GPT-5.5-level agent intelligence, though at a heavy "reasoning tax" (token-inefficient thinking loops). - This allows us to refine **Hey, Lefty's** positioning as a model-agnostic orchestrator that integrates premium vertical data (like Sacra and FactSet) via MCP, bypassing the "reasoning tax" of monolithic models by orchestrating specialized micro-agents. Let's check if there are other slugs we should update: 2. `perplexity-deep-research-consumer-agent`: This is specifically about Perplexity Deep Research. We should update it with the new **Sonar Deep Research API** pricing and cost structure (Input: $2/1M, Output: $8/1M, Citation: $2/1M, Search queries: $5/1K, Reasoning: $3/1M) and the **Agent API** tool-pricing model. 3. `gemini-deep-research-agent`: This is about Gemini Deep Research. We should update it with the **Gemini Deep Research Agent API** (Interactions API) details: - Dual versions: `deep-research-preview-04-2026` and `deep-research-max-preview-04-2026`. - Collaborative planning (`collaborative_planning=True`). - Auto-visualization (`visualization="auto"`). - Native **Model Context Protocol (MCP)** server integration (`mcp_server`). - Standard Gemini 3.1 Pro token pricing for all intermediate steps. 4. `academic-ai-research-scholar-vs-synthesis`: This is about academic AI research. We should update it to cover **Scholar Labs** (released November 2025), which is a "Deep Search" tool leveraging the unmatched scale of the Google Scholar index to do paper-by-paper relevance reasoning rather than risking multi-paper synthesis hallucinations. 5. `sacra-premium-private-market-mcp-data`: We should update it with the exact pricing plans of Sacra (Standard at $50/mo, Pro at $350/mo, Platform at $1500/mo) and confirm that all these plans now explicitly include **Model Context Protocol (MCP) server access** and API access, making premium private market data programmatically available to agentic loops. Wait, let's write each of these notes one by one, ensuring we ground them in concrete quotes and links from our research this cycle. Let's draft the first note: **Market Map & Positioning: Where Does Hey, Lefty Fit?** - Slug: `market-map-positioning-hey-lefty` - We will incorporate the new dimensions of the market in June 2026: - **Academic Deep Search**: Google's **Scholar Labs** (Nov 2025) which uses LLMs to do paper-by-paper relevance reasoning across the Google Scholar index instead of risky multi-paper synthesis. - **API-Driven Agentic Research**: Google's **Gemini Deep Research Agent API** (Interactions API) and Perplexity's **Sonar Deep Research API** / **Agent API**. - **Open Weights Agentic Intelligence**: Z ai's **GLM-5.2** (June 2026) which matches GPT-5.5 (xhigh) on agentic benchmarks (GDPval-AA v2) but imposes a high "reasoning tax" in token-based cost and time. - **Hey, Lefty's Positioning**: Hey, Lefty's sweet spot is as an orchestrator of this fragmented landscape. Rather than relying on a single monolithic model that runs up massive "reasoning taxes" (like GLM-5.2's 43k tokens/task or Gemini's standard Pro rates for all intermediate reasoning), Hey, Lefty acts as a lean, model-agnostic coordinator. It leverages specialized, high-fidelity data sources (like **Sacra** and institutional financial platforms) via **Model Context Protocol (MCP)** and integrates academic deep search (via **Scholar Labs**), bypassing the token bloat of general-purpose "deep research" agents. Let's write this note using `write_note`. Let's ensure we include verbatim quotes and sources. Wait, let's look at the sources we can cite: - [Scholar Labs Early Review: Google Scholar Finally Enters the AI Era](https://aarontay.substack.com/p/scholar-labs-early-review-google) - [GLM-5.2 is the new leading open weights model on Artificial Analysis](https://artificialanalysis.ai/articles/glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index) - [Gemini Developer API pricing](https://ai.google.dev/gemini-api/docs/pricing) / [pricing.md.txt](https://ai.google.dev/gemini-api/docs/pricing.md.txt) - [Pricing - Perplexity](https://docs.perplexity.ai/docs/getting-started/pricing) - [Sacra pricing](https://sacra.com/pricing/) Let's write the first note. I will state my intent before calling `write_note`.I will update the core market map finding (`market-map-positioning-hey-lefty`) to incorporate the latest developments from June 2026, including Google's Scholar Labs, Perplexity's Sonar Deep Research API, Gemini's Deep Research Agent API, and Z ai's GLM-5.2, and refine Hey, Lefty's strategic positioning within this landscape.

That was a fantastic write-up! Let's now write/update the other notes to capture the deeper technical details for each platform. Let's update: 1. `perplexity-deep-research-consumer-agent` (Perplexity Deep Research: Consumer-Centric High-Speed Autonomous Synthesis and the Search as Code Paradigm) Let's see what we can write about Perplexity's Sonar Deep Research API and Agent API. Wait, let's look at the pricing and features: - **Sonar Deep Research** model: - Input tokens: $2 / 1M - Output tokens: $8 / 1M - Citation tokens: $2 / 1M - Search queries: $5 / 1K - Reasoning tokens: $3 / 1M - **Agent API Tool Pricing**: - `web_search`: $0.005 per invocation - `fetch_url`: $0.0005 per invocation - `people_search`: $0.005 per invocation - `finance_search`: $0.005 per invocation - `sandbox`: $0.03 per session (20-minute isolated container session for code execution) - **Search API**: $5 / 1K requests (raw web search with advanced filtering, no token cost). - **Pro Search Options**: - `fast`: Standard Sonar Pro behavior ($6 / $10 / $14 per 1K requests based on context size) - `pro`: Multi-step tool usage for complex queries ($14 / $18 / $22 per 1K requests based on context size) - `auto`: Automatic classification based on query complexity. Let's write this note using `write_note`. Let's state my intent first.I will update the Perplexity Deep Research note (`perplexity-deep-research-consumer-agent`) with the newly launched Sonar Deep Research API, its granular pricing structure, the Agent API's tool-pricing model, and how they implement the "Search as Code" paradigm programmatically.

Excellent! Let's now write/update the Gemini Deep Research note: - Slug: `gemini-deep-research-agent` (Gemini Deep Research: Multi-Step Reasoning via the Interactions API) Let's see what we can write about Gemini Deep Research: - Dual versions: - `deep-research-preview-04-2026`: Designed for speed and efficiency, ideal to be streamed back to a client UI. - `deep-research-max-preview-04-2026`: Maximum comprehensiveness for automated context gathering and synthesis. - Collaborative planning (`collaborative_planning=True`): - Step 1: Request a plan. - Step 2: Refine the plan (optional) using `previous_interaction_id`. - Step 3: Approve and execute (setting `collaborative_planning=False`). - Visualization (`visualization="auto"`): Autonomously generates charts, graphs, and visual elements to support research findings. - Supported tools: - Google Search (`google_search`) - URL Context (`url_context`) - Code Execution (`code_execution`) - **MCP Server** (`mcp_server`): Connect to remote MCP servers for external tool access! - File Search (`file_search`) - Pricing: - Charged at standard Gemini list rates (e.g., Gemini 3.1 Pro Preview: Input: $2.00–$4.00/1M, Output: $12.00–$18.00/1M). - This includes all intermediate input, output, and reasoning tokens generated during the autonomous multi-step loops. - Google Search grounding: 5,000 prompts/month free, then $14 / 1,000 search queries. - URL context: Charged as input tokens. - Code execution: Billed at standard token rates for the selected model (generated code and results billed as output tokens, then input tokens when reused). Let's write this note using `write_note`. Let's state my intent first.I will update the Gemini Deep Research note (`gemini-deep-research-agent`) with the new Gemini Deep Research Agent API (Interactions API), detailing its collaborative planning, auto-visualization, native Model Context Protocol (MCP) server integration, and standard token-based agent pricing.

Excellent! Let's now write/update the academic AI research note: - Slug: `academic-ai-research-scholar-vs-synthesis` (Academic AI Research: Google Scholar vs. Specialized Synthesis Tools) Let's see what we can write about **Scholar Labs**: - Released in November 2025 by Google Scholar. - It is a "Deep Search" tool rather than a "Deep Research" tool. - A "Deep Search" tool is designed to find relevant papers and not generate direct answers. It differs from normal search by going "deeper," running beyond the typical 1000ms latency of traditional search engines. - Behind the scenes, it performs query understanding, runs multiple expanded queries (e.g., 11 different queries in a test), combines and de-dupes the results, evaluates the top results, and generates a paper-by-paper explanation of why each paper is relevant to the query. - It typically stops once it has found 10 highly relevant papers (though clicking "more results" continues the search). - It does not attempt to synthesize across papers to generate a long-form report, which avoids the risk of hallucination and contradictory synthesis. - Unlike other academic AI engines (which use OpenAlex or Semantic Scholar), Scholar Labs inherits the unmatched scale of the Google Scholar index, which includes paywalled full-text crawls from almost every major publisher. This makes it particularly dominant in non-STEM disciplines where other academic databases are weaker. Let's write this note using `write_note`. Let's state my intent first.I will update the academic AI research note (`academic-ai-research-scholar-vs-synthesis`) to cover the launch of Google's Scholar Labs (November 2025) and analyze its "Deep Search" vs. "Deep Research" paradigm, its query expansion mechanics, and its leveraging of the massive Google Scholar index.

Excellent! Let's now write/update the Sacra note: - Slug: `sacra-premium-private-market-mcp-data` (Sacra: Premium Private Market Research as Agent-Ready Infrastructure) Let's see what we can write about Sacra: - Standard plan: $50/mo, 1 seat, 500 tasks. - Pro plan: $350/mo, 5 seats, 3500 tasks. - Platform plan: $1500/mo, 20 seats, 25000 tasks. - Enterprise plan: Custom seat/task allocation. - All plans natively include Model Context Protocol (MCP) server access and API access. - The Sacra API makes revenue estimates, valuations, funding rounds, research documents, and news programmatically accessible. - The MCP server connects Sacra directly to Claude, ChatGPT, Cursor, or any custom agent without requiring manual API integration. - This is a prime example of premium private market intelligence unbundling itself into "agent-ready infrastructure." Let's write this note using `write_note`. Let's state my intent first.I will update the Sacra premium private market research note (`sacra-premium-private-market-mcp-data`) to capture their exact pricing plans (Standard, Pro, Platform), task limits, and their native integration of Model Context Protocol (MCP) server access across all tiers.

This is amazing! We have updated all four of our core findings to reflect the latest developments in 2026. Let's review if there are any other notes we should write or if we are ready to finish the cycle. Wait, let's look at the active threads or watches. We checked: - "Perplexity Releases WANDR Benchmark to the Public" (The benchmark is private/in-house, not yet released in search_evals repository). - "Anthropic Restores Access to Fable 5 and Mythos 5" (Access is still suspended as of June 18, 2026). - "Gemini Deep Research Max Transitions to GA" (Still in preview as of June 2026). - "SpaceX Acquisition of Cursor Closes" (Expected in Q3 2026, which is in the future). Wait, are there any other open threads or watches? Let's check `list_threads` again. The threads are: - `019ed3e0-ceef-79c4-afdb-1ca1e2aa1745`: "SpaceX Acquisition of Cursor Closes" - `019eceb8-a74b-7495-81db-b5b2bbff4c86`: "Perplexity Releases WANDR Benchmark to the Public" - `019ebf48-b349-73e7-984e-261a5c8bc427`: "Anthropic Restores Access to Fable 5 and Mythos 5" - `019eafd3-288d-7900-9b7b-f29c43234577`: "Antitrust Scrutiny of Anthropic Silent Safeguards" - `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`: "Appeals and Rulings on AI Summary Liability in the EU" - `019ea05f-eaaf-7afb-bb51-55b5f00e17d3`: "Institutional Research Platforms Launch Public MCP Servers" - `019ea05f-d685-7208-a431-55568fba4987`: "Gemini Deep Research Max Transitions to GA" Let's check if there are any other events to record or threads to resolve. Wait, none of these watches have fired yet, so we don't record any events. Let's see if we should open any new threads or register any new watches. Wait, since we have updated our findings with concrete evidence and resolved the user's research questions, we can now complete the cycle. Let's verify if we need to write any other notes. No, the four notes we updated: - `market-map-positioning-hey-lefty` - `perplexity-deep-research-consumer-agent` - `gemini-deep-research-agent` - `academic-ai-research-scholar-vs-synthesis` - `sacra-premium-private-market-mcp-data` (That's actually five notes updated/written! This is a comprehensive set of updates that fully addresses the user's brief). Let's call `finish_cycle` to complete this research cycle. I will state my intent first.

I have completed the research cycle and successfully updated five core findings on the market for autonomous research tools. Here is a summary of the key findings and updates written to your workspace: ### 1. Market Map & Positioning: Where Does Hey, Lefty Fit? (`market-map-positioning-hey-lefty`) * **Market Stratification:** The market has split into distinct, highly specialized quadrants: 1. *Consumer-Facing "Deep Research" Engines* (Perplexity Deep Research, Gemini Advanced, OpenAI Pro) optimized for end-to-end, black-box long-form synthesis. 2. *API-Driven Agentic Research Platforms* (Gemini Deep Research Agent API, Perplexity Sonar Deep Research API, Perplexity Agent API) offering pay-as-you-go programmatic loops. 3. *Academic "Deep Search" & Discovery* (Google's **Scholar Labs**), which prioritizes paper-by-paper evaluation over risky multi-paper synthesis. 4. *Agent-Ready Private Market Infrastructure* (**Sacra**, FactSet, S&P Global, PitchBook) unbundling premium curated data via Model Context Protocol (MCP) servers. * **The "Reasoning Tax" Bottleneck:** The release of Z ai's **GLM-5.2** (June 2026) shows that while open weights can match GPT-5.5-level agentic intelligence (scoring 1524 on GDPval-AA v2), they suffer from massive token inefficiency—averaging **43,000 output tokens per task** (with **37,000 reasoning tokens**), causing long latency and high costs. * **Hey, Lefty's Positioning:** Positioned as a **Model-Agnostic Research Orchestrator**. Instead of running expensive, monolithic end-to-end reasoning loops, Hey, Lefty serves as a lightweight coordinator. It delegates academic paper verification to Scholar Labs, private company financial queries to Sacra's MCP server, and general scraping to cheap utility models—minimizing the "reasoning tax" while providing predictable costs, user-steered collaborative planning, and data sovereignty. ### 2. Perplexity Deep Research & Sonar API (`perplexity-deep-research-consumer-agent`) * **Sonar Deep Research API:** Perplexity has programmatically exposed its deep research capabilities. It features a granular billing model: Input ($2/1M), Output ($8/1M), Citation ($2/1M), Reasoning ($3/1M), and Search Queries ($5/1K). A typical deep research query costs **~$0.41**. * **Agent API Tool Pricing:** Developers can build their own agentic loops using third-party models at pass-through cost, paying flat fees per tool invocation (e.g., `web_search` at $0.005, `fetch_url` at $0.0005, and `sandbox` code execution at $0.03 per 20-minute session). * **Context Controls:** Context limits can be tuned (Low, Medium, High) to manage request fees for Sonar Pro and Sonar Reasoning Pro. ### 3. Gemini Deep Research Agent API (`gemini-deep-research-agent`) * **Interactions API:** Google DeepMind released the Gemini Deep Research Agent in public preview, available in standard (`deep-research-preview-04-2026`) and maximum-depth (`deep-research-max-preview-04-2026`) versions. * **Collaborative Planning:** Supports `collaborative_planning=True`, enabling the agent to return a proposed research plan first, allowing interactive multi-turn plan refinement before execution. * **Auto-Visualization & Tooling:** Autonomously generates charts and graphs (`visualization="auto"`) and supports native **Model Context Protocol (MCP)** server integration for remote tool access. * **Agent Pricing:** Charged at standard Gemini 3.1 Pro Preview token rates (Input: $2.00–$4.00/1M, Output: $12.00–$18.00/1M) for all intermediate reasoning and input tokens generated during the autonomous loop. ### 4. Academic AI Research: Google Scholar Labs (`academic-ai-research-scholar-vs-synthesis`) * **Scholar Labs (Nov 2025):** Google's official entry into AI-powered academic search. It is built as a high-precision **"Deep Search"** tool rather than a synthesis engine. * **How it Works:** It performs query understanding, runs multiple expanded queries, evaluates the top retrieved papers, and generates a paper-by-paper explanation of why each is relevant to the query. It typically halts after finding 10 highly relevant papers to prevent waffling and hallucinations. * **Strategic Advantage:** Unlike other tools relying on open indexes like OpenAlex or Semantic Scholar, Scholar Labs inherits the unmatched scale of the Google Scholar index, including paywalled full-text data from major publishers, making it particularly dominant in non-STEM disciplines. ### 5. Sacra: Agent-Ready Infrastructure (`sacra-premium-private-market-mcp-data`) * **Tiered Pricing:** Standard ($50/mo, 500 tasks), Pro ($350/mo, 3,500 tasks), and Platform ($1,500/mo, 25,000 tasks). * **Native MCP Server:** All tiers natively include Model Context Protocol (MCP) server access, enabling LLMs and orchestrators like **Hey, Lefty** to programmatically extract pre-IPO revenue models, expert interviews, and valuation data without writing custom API integration code.