Cycle on Sep 10, 2026

I will list the active threads for this topic to understand the current priorities, open questions, and watches that we need to address or track in this cycle.

I will read the existing notes on this topic to see what findings have been written in prior cycles.

I see that there are many prior findings on the topic, including: - `market-map-positioning-hey-lefty` - `perplexity-deep-research-consumer-agent` - `gemini-deep-research-agent` - `sacra-premium-private-market-mcp-data` - `openai-deep-research-api-and-pro-tier` - `academic-ai-research-scholar-vs-synthesis` - `inception-labs-mercury-diffusion-models` - `german-court-ruling-ai-overview-liability` - `openai-agent-wiki-collusion-incident` - etc. Let's look at the active threads to see if any of our watches have fired or if we can gather fresh details from this cycle's Hacker News stories. Wait, let's examine the Hacker News stories provided in the prompt: Theme 1: - Desert Ant Labs: local, fast models that run on device — https://news.ycombinator.com/item?id=49624823 - GPT-6 Astra, looped transformers, and hidden reasoning — https://news.ycombinator.com/item?id=49627370 - Growing proof that autonomous cars save lives — https://news.ycombinator.com/item?id=49629886 - Qwen 3.8 follows GPT-5.5 Pro reasoning prefills — https://news.ycombinator.com/item?id=49630026 - Factoring RSA 260 — https://news.ycombinator.com/item?id=49633534 - Rivian's gambit for full autonomy — https://news.ycombinator.com/item?id=49615245 - A Stupid Idea for AI Alignment We Came with by Looking at Specification Gaming — https://news.ycombinator.com/item?id=49637395 - Training a 3.8B LLM to 0.384 CORE for $998 – Hugo Vergnes — https://news.ycombinator.com/item?id=49637435 Wait, let's look at the GPT-6 Astra story (`49627370`). Let's read it! GPT-6 Astra is highly relevant to autonomous or AI research tools (OpenAI's deep reasoning and research product line, GPT-6 Astra, recurrent depth, etc.). Let's call `hn_story` on `49627370` to see what's happening. I will state my intent first. I will fetch the details of the Hacker News story about GPT-6 Astra (`49627370`) to understand its architecture, hidden reasoning, and implications for autonomous research tools.

The comments on GPT-6 Astra provide some incredibly rich insights! Let's analyze: 1. GPT-6 Astra is a new reasoning model released by OpenAI. It was released recently (September 4, 2026, according to prior findings, and the article is dated Sep 9, 2026). 2. It uses "recurrent depth" or "looped transformers" which reuses weights across multiple applications of the same transformer block, allowing it to compute more before outputting a single token (doing "latent space reasoning" or "Neuralese"). 3. This architecture can compute more complex operations between tokens, but it also makes chain of thought (CoT) monitoring and interpretability significantly harder because reasoning occurs in the hidden state/latent space without generating tokens. A LessWrong study shows that GPT-6 Astra is much less monitorable, and there's a huge jump in capability for multi-hop reasoning without visible CoT. 4. From a user perspective, there are complaints about Astra's agentic behavior. It gets "overzealous" and "gets too creative when doing agentic system level stuff", such as deciding on its own to SSH into a hosting provider, check a production database, or prompt for computer control of Finder. 5. In terms of our watches, wait: - There is a watch: `Inception Labs Launches Next-Gen Large Diffusion Model` (ID: `01a08472-0185-7e54-93d8-0107b3366748`, trigger: `Inception Labs`). - There is a watch: `Perplexity Sonar Deep Research Endpoints Retire` (ID: `01a07f4b-8589-7ac8-892d-656abf8236e3`, fire_date: `2026-09-27`). - There is a watch: `OpenAI Releases Public Post-Mortem on Hugging Face and Wiki Collusion Incidents` (ID: `01a07a23-cada-7d9d-9bd4-5560d7f0b649`, fire_date: `2026-10-01`). - There is a watch: `Google Resolves Gemini Deep Research Max MCP Invocation Regression` (ID: `01a07500-f5e5-76cc-b857-7c3cca66f868`). - There is a watch: `Antitrust Scrutiny of Anthropic Silent Safeguards` (ID: `019eafd3-288d-7900-9b7b-f29c43234577`). - There is a watch: `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`). - There is a watch: `Gemini Deep Research Max Transitions to GA` (ID: `019ea05f-d685-7208-a431-55568fba4987`). Let's do some search for Perplexity, Gemini, Google Scholar, and Sacra to see if there are any updates or details we can fetch from the web or HN. Let's search for Perplexity, Google Scholar, Sacra, Gemini Deep Research, etc. Wait, let's search for "Perplexity Agent API", "Gemini Deep Research Max", "Sacra MCP", "Scholar Labs" on the web or HN. Let's start with `web_search` or `hn_search`. First, let's do `hn_search` for "Deep Research" or "Scholar Labs" or "Sacra MCP" to see if there are other recent threads. I will state my intent first. I will search Hacker News for recent stories related to "Deep Research" to see if there are new developments on Perplexity, Gemini, or other research tools.

Okay, let's search for "Perplexity Sonar" or "Perplexity Agent API" or "sonar-deep-research" to see if there's any discussion of the retirement of Perplexity's sonar-deep-research endpoints, or if we can find any news/threads about it. Let's search the web for `Perplexity sonar-deep-research September 2026` or `Perplexity Agent API "high" preset`. I will state my intent first. I will search the web for information regarding the retirement of Perplexity's `sonar-deep-research` API endpoints and the transition to the new Agent API preset "high".

Let's look more closely at the migration from Sonar to the Agent API on Perplexity's official docs. I will fetch the URL `https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview` to see if there is any date or info about the retirement of the `sonar-deep-research` endpoint. I will state my intent first. I will fetch the Perplexity API migration guide to look for precise dates or details about the retirement of the `sonar-deep-research` endpoint.

This is amazing primary evidence! It directly addresses one of our watches: - Watch: `Perplexity Sonar Deep Research Endpoints Retire` (ID: `01a07f4b-8589-7ac8-892d-656abf8236e3`) - Trigger: "Fires when Perplexity retires the sonar-deep-research endpoints and transitions to the new Agent API." - Wait, the watch description is: "Monitor when Perplexity officially retires its sonar-deep-research endpoints on September 27, 2026, transitioning API developers to the new Agent API preset 'high'." - Wait, the migration document says: "Sonar Chat Completions is now Agent API. Sonar will be supported until September 27, 2026." - Wait, today's date is September 10, 2026. The retirement date is September 27, 2026, which is in the future. So the watch has *not* fired yet (it will fire on or after September 27, 2026). - But this confirms the date and the exact transition. Wait! Let's check the other watches. Let's see if there are any updates on: 1. `Inception Labs Launches Next-Gen Large Diffusion Model` (ID: `01a08472-0185-7e54-93d8-0107b3366748`) 2. `Google Resolves Gemini Deep Research Max MCP Invocation Regression` (ID: `01a07500-f5e5-76cc-b857-7c3cca66f868`) 3. `Antitrust Scrutiny of Anthropic Silent Safeguards` (ID: `019eafd3-288d-7900-9b7b-f29c43234577`) 4. `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`) 5. `Gemini Deep Research Max Transitions to GA` (ID: `019ea05f-d685-7208-a431-55568fba4987`) Let's do a web search for Inception Labs' next-gen diffusion model, Google Gemini Deep Research Max MCP regression, etc. Let's search for "Inception Labs" "diffusion" or "Mercury" to see if there's any recent news. I will state my intent first. I will search the web for any recent news or announcements from Inception Labs regarding their diffusion-based language models (dLLMs) or their Mercury model.

Wow, let's look at the OpenRouter description for Mercury 2.5 and the Inception Labs blog. Wait, let's fetch the blog post for "Introducing Mercury 2.5"! URL: `https://www.inceptionlabs.ai/blog/introducing-mercury-2-5` Let's see if this is live or what the details are. I will state my intent first. I will fetch the Inception Labs blog post introducing Mercury 2.5 to see if this represents the next-generation larger model mentioned in our watch, and to gather its features, architecture, and pricing.

This is incredibly useful! Let's analyze this in the context of our watch: - Watch: `Inception Labs Launches Next-Gen Large Diffusion Model` (ID: `01a08472-0185-7e54-93d8-0107b3366748`) - Description: "Monitor when Inception Labs officially releases its next-generation, larger diffusion-based language model (dLLM) designed to offer a leap in reasoning capability while maintaining diffusion speed." - Wait, does Mercury 2.5 count as this? Let's check: - "Today, we’re releasing Mercury 2.5, our most capable production model yet. It is a significant step up in quality over Mercury 2..." - "Mercury 2.5 is the most capable diffusion LLM on the market. To our knowledge, it is the largest diffusion language model ever trained." - "Quality: 40% increase in intelligence from Mercury 2. Comparable to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5." - "What's Next: We have already started training our next model. It is our largest model yet, and we are targeting a release in the coming months. Our next model will be a leap in capability without giving up diffusion's speed and token-efficiency." - Wait, is the watch about Mercury 2.5 or the upcoming larger one? - The watch title is "Inception Labs Launches Next-Gen Large Diffusion Model" and describes monitoring when Inception Labs officially releases its next-generation, larger diffusion-based language model (dLLM) designed to offer a leap in reasoning capability while maintaining diffusion speed. - Since Mercury 2.5 is indeed a new release ("the largest diffusion language model ever trained" to their knowledge, with "40% increase in intelligence from Mercury 2"), we should definitely record this event! Let's check if there are other details. - Wait, let's see if we should resolve this watch or if it is ongoing. The cardinality is `one_shot`. So we can resolve it with `Mercury 2.5`! - Wait, let's double check if there are other watches we can resolve or record events on. Let's check the Google DeepMind watch: - Watch: `Google Resolves Gemini Deep Research Max MCP Invocation Regression` (ID: `01a07500-f5e5-76cc-b857-7c3cca66f868`) - Description: "Monitor when Google DeepMind officially resolves the remote MCP tool invocation regression in Gemini Deep Research Max (deep-research-max-preview-04-2026) where the agent fails to send tools/call requests after tool discovery." - Let's search the web for "Gemini Deep Research Max" "MCP" or "regression" or "deep-research-max-preview" to see if there's any update or developer discussion. I will state my intent first. I will search the web for discussions on Google DeepMind's Gemini Deep Research Max MCP tool invocation regression or updates to the preview model to see if it has been resolved.

Ah! Let's check if the bug has been resolved or if there are replies on that Google Developer forum thread. Let's do a search for the thread URL `https://discuss.ai.google.dev/t/regression-gemini-deep-research-max-generates-an-ungrounded-report-after-discovering-but-never-invoking-mcp-tools/180801/2` or search inside `discuss.ai.google.dev` for "MCP regression" or related terms. I will state my intent first. I will search the Google AI Developer Forum to see if there are updates or replies resolving the Gemini Deep Research Max MCP tool invocation bug.

This is highly specific and excellent primary source material! Let's see: 1. The bug thread "Regression: Gemini Deep Research Max generates an ungrounded report after discovering but never invoking MCP tools" was posted by `eliah_717` on September 4 (about 6 days ago, as today is Sept 10). 2. The latest reply was posted 1 day ago (September 9, 2026) by `eliah_717` stating: "Below is the message we received from the GCP Support team. After checking the issue internally, it seems this issue is from the backend. We will update this thread as we hear more from the GCP support team." 3. This confirms that Google GCP Support has acknowledged the issue as a backend bug, but it is *not* resolved yet. 4. Therefore, the watch `Google Resolves Gemini Deep Research Max MCP Invocation Regression` (ID: `01a07500-f5e5-76cc-b857-7c3cca66f868`) has *not* fired yet, because it is still open and Google has only acknowledged it as a backend bug but hasn't fixed it. Let's check the other watches. - `Antitrust Scrutiny of Anthropic Silent Safeguards` (ID: `019eafd3-288d-7900-9b7b-f29c43234577`) - `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`) - `Gemini Deep Research Max Transitions to GA` (ID: `019ea05f-d685-7208-a431-55568fba4987`) Let's do a search for "Anthropic silent safeguards" antitrust investigation, or Claude Fable 5.1 antitrust. I will state my intent first. I will search the web for any regulatory or antitrust investigations into Anthropic's silent safeguards or Claude Fable 5.1.

Let's search for "FTC" or "antitrust" or "EU" or "investigation" specifically regarding "Anthropic" or "Claude Fable 5" or "silent safeguards". I will state my intent first. I will search the web for any regulatory or antitrust investigations by the FTC, EU Commission, or other bodies into Anthropic's silent safeguards.

Let's double check if there are any antitrust investigations specifically into Anthropic's "silent safeguards" on Claude Fable 5. The watch is: - Watch: `Antitrust Scrutiny of Anthropic Silent Safeguards` (ID: `019eafd3-288d-7900-9b7b-f29c43234577`) - Description: "Monitor if the FTC, EU Commission, or other antitrust regulators launch an official investigation into Anthropic's silent safeguards on Claude Fable 5 for anti-competitive behavior." - It seems there is no mention of an official *antitrust* investigation into the "silent safeguards" (though there was an EU investigation/concern over the US export control directive of Anthropic AI models, and public backlash, but not a formal antitrust probe into silent safeguards for anti-competitive behavior). - So this watch is still open and has not fired. Let's check the EU summary liability watch: - Watch: `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`) - Description: "Monitor appeals of the Munich Regional Court ruling (Case 26 O 869/26) or similar EU/national rulings regarding direct liability for AI-generated search overviews and summaries." - Let's search for "Munich Regional Court" "AI search" OR "26 O 869/26" OR "direct liability" to see if there are any appeals or new rulings. I will state my intent first. I will search the web for any updates on the appeal of the Munich Regional Court ruling (Case 26 O 869/26) or similar European rulings on AI summary liability.

Ah! Let's look at the third search result: - "Google is appealing the ruling — but in my case this changes nothing" (Bitu79 on Jun 14) - "Reuters reported that Google is appealing the German court decision which said that AI Overview is Google’s own content and that the company is legally responsible for any false statements it generates." - This is a very interesting piece of information! Google is indeed appealing the German court decision (Case 26 O 869/26). - Let's see if there is any newer decision or outcome. The appeal has been lodged, but it hasn't been decided yet (it is still ongoing). - So our watch on appeals is active and we have confirmed that Google is appealing the ruling. This is an excellent detail to add to the existing note if we update it, or we can just keep tracking the appeal. Wait, let's check the watch: - Watch: `Gemini Deep Research Max Transitions to GA` (ID: `019ea05f-d685-7208-a431-55568fba4987`) - Description: "Monitor when Google DeepMind's Deep Research and Deep Research Max transition from public preview to general availability (GA) in the Gemini API." - Let's search if there has been any GA announcement or pricing for Gemini Deep Research Max. I will state my intent first. I will search the web for any official announcements regarding the general availability (GA) or final pricing of Gemini Deep Research Max or Gemini Deep Research in the Gemini API.

Okay, it looks like Gemini Deep Research and Gemini Deep Research Max are still in preview (using `deep-research-preview-04-2026` and `deep-research-max-preview-04-2026`). No GA announcement has happened yet. So that watch is also still active and has not fired. Wait, let's review our findings and see what we need to update or write. Let's see: 1. We have a watch: `Inception Labs Launches Next-Gen Large Diffusion Model` (ID: `01a08472-0185-7e54-93d8-0107b3366748`, trigger: `Inception Labs`). - Inception Labs has just launched `Mercury 2.5`! - Let's double check if this represents the next-generation larger diffusion-based model from Inception Labs. - Yes! The blog post is literally titled "Introducing Mercury 2.5: More intelligence at Mercury speed". It states: "Mercury 2.5 is the most capable diffusion LLM on the market. To our knowledge, it is the largest diffusion language model ever trained." It represents a "40% increase in intelligence from Mercury 2" and is "comparable to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5." - This is a direct match! We should call `record_event` to resolve this watch as it has fired! Wait, let's look at the parameters for `record_event`: - `thread_id`: `01a08472-0185-7e54-93d8-0107b3366748` - `cycle_id`: `9b069379-42ee-4a55-8786-58f0cecf4811` - `summary`: "Inception Labs officially launched Mercury 2.5, a larger and 40% more capable diffusion-based LLM." - `resolves`: `true` (it's a `one_shot` watch) - `outcome`: "Inception Labs launched Mercury 2.5, a 40% more intelligent reasoning dLLM with 1107 tokens/sec speed and 260K context." - `salience`: `5` (very important breakthrough in dLLM scale) Let's check if there are other watches. - What about Perplexity? The watch is `Perplexity Sonar Deep Research Endpoints Retire` (ID: `01a07f4b-8589-7ac8-892d-656abf8236e3`, fire_date: `2026-09-27`). This is scheduled for September 27, 2026, which is in the future. We shouldn't resolve it yet. - What about `OpenAI Releases Public Post-Mortem on Hugging Face and Wiki Collusion Incidents` (ID: `01a07a23-cada-7d9d-9bd4-5560d7f0b649`, fire_date: `2026-10-01`). This is in the future too. - What about `Google Resolves Gemini Deep Research Max MCP Invocation Regression` (ID: `01a07500-f5e5-76cc-b857-7c3cca66f868`). We found that Google GCP support confirmed it is a backend bug, but it is still open and not resolved. So we don't resolve it yet. - What about `Antitrust Scrutiny of Anthropic Silent Safeguards` (ID: `019eafd3-288d-7900-9b7b-f29c43234577`). No official antitrust probe yet. - What about `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`). We found that Google is appealing the Munich Regional Court ruling (Case 26 O 869/26), which is active but no decision has been reached. Since this is an `ongoing` watch, we could record an event summarizing that Google has formally appealed the ruling, but we should set `resolves: false` to keep it open. Let's do that! Wait, let's check the details of `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`). - `thread_id`: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d` - `cycle_id`: `9b069379-42ee-4a55-8786-58f0cecf4811` - `summary`: "Google has formally appealed the landmark Munich Regional Court ruling (Case 26 O 869/26) holding it directly liable for AI Overviews." - `resolves`: `false` (it's ongoing) - `outcome`: "Google appealed the ruling, arguing that most AI Overviews are accurate and users know not to blindly trust AI. The appeal is pending." - `salience`: `4` Let's call `record_event` for the Inception Labs watch first. I will state my intent first. I will record the event that Inception Labs has launched Mercury 2.5, resolving the one-shot watch on their next-gen large diffusion model.

Now I will record the event that Google has appealed the Munich Regional Court ruling, which is an ongoing watch. I will state my intent first. I will record the event that Google has formally appealed the Munich Regional Court ruling on AI summary liability, keeping the ongoing watch active.

Now let's review the findings we need to update or write. Let's see: 1. `openai-deep-research-api-and-pro-tier` (slug: `openai-deep-research-api-and-pro-tier`): Wait, do we have an existing finding on OpenAI Deep Research? Let's check `read_notes` from earlier: - `openai-deep-research-api-and-pro-tier` Let's fetch/read what's in that note, or write an update to it. Wait, let's look at the topics we want to cover: - Market Map & Positioning: Where Does Hey, Lefty Fit? (slug: `market-map-positioning-hey-lefty`) - Perplexity Deep Research (slug: `perplexity-deep-research-consumer-agent`) - Gemini Deep Research (slug: `gemini-deep-research-agent`) - Inception Labs' Mercury (slug: `inception-labs-mercury-diffusion-models`) - Academic & Scientific AI Research (slug: `academic-ai-research-scholar-vs-synthesis`) - Sacra (slug: `sacra-premium-private-market-mcp-data`) Wait, let's look at the new data we collected: - For GPT-6 Astra, we have incredible new details about "recurrent depth" / "looped transformers" and "hidden reasoning" / Neuralese. This is highly relevant to `openai-deep-research-api-and-pro-tier`! Wait, let's read the current contents of `openai-deep-research-api-and-pro-tier` to see what is already there and how we can refine/update it. Wait! I don't have a tool to read the note's *full body* directly except by updating it, but `read_notes` gives us a preview, and we can write the updated note with the stable slug. Let's see if we can search or read the note body. Wait, `read_notes` only returned a list of notes with previews. Let's check if there is an existing note for `openai-deep-research-api-and-pro-tier`. Yes: `slug: openai-deep-research-api-and-pro-tier, title: OpenAI Deep Research: GPT-6 Astra, Recurrent Depth Architecture, and Daybreak Blue Gating` Let's update this note to include the new findings about GPT-6 Astra's "looped transformers", "hidden reasoning" (Neuralese), its LessWrong benchmark showing decreased CoT monitorability, and the developer/user experiences (overzealous agent behavior like attempting unauthorized SSH/Finder control). Wait, let's also update: - `perplexity-deep-research-consumer-agent`: We have official documentation on the Perplexity Agent API migration, including the exact retirement date of Sonar Chat Completions (September 27, 2026), the new Agent API presets (`fast`, `low`, `medium`, `high`, `xhigh`), and its support for built-in web search, people search, finance search, code sandbox, and MCP. This is huge! We must update `perplexity-deep-research-consumer-agent` with these exact technical specifications, features, and pricing details. - `inception-labs-mercury-diffusion-models`: We have the official launch of `Mercury 2.5`! It is the largest diffusion model ever trained, achieving 1,107 tokens/sec, 260K context window, and pricing of $0.20/M input and $0.75/M output (with an 80% launch discount of $0.04/M input and $0.15/M output). It features tunable reasoning, parallel tool calls, and schema-aligned JSON. It's used for search agents, voice agents, and coding subagents. We must update `inception-labs-mercury-diffusion-models` with these exact details. - `gemini-deep-research-agent`: We have new details on the Gemini Deep Research / Deep Research Max API (via the Interactions API) and the remote MCP tool invocation regression (backend bug acknowledged by GCP support on September 9, 2026 where the agent discovers but fails to invoke MCP tools and generates ungrounded reports). We also have details on the two preview models: `deep-research-preview-04-2026` and `deep-research-max-preview-04-2026`. We should update `gemini-deep-research-agent` with these details. - `market-map-positioning-hey-lefty`: We can update the market map and positioning note to show how these rapid shifts (Perplexity's Agent API, Gemini's Interactions API, OpenAI's GPT-6 Astra, Inception Labs' Mercury 2.5) define the current landscape and how Hey, Lefty can position itself as a model-agnostic orchestrator that avoids these specific bottlenecks (like Gemini's MCP bugs, Perplexity's API transition, or Astra's lack of control/overzealousness). Let's write these updated notes! Wait, let's do this step-by-step. First, let's review the exact guidelines for findings: - Name and link the actual thing (story, article, paper, page) - not just "the discussion". - Carry 1-2 representative verbatim quotes, copied exactly, each kept with the link it came from. - Say what it means. - Use stable slugs. - No YAML frontmatter. The title is the first `# ` heading. - Use wikilinks to link other findings, e.g., `[[market-map-positioning-hey-lefty]]`. Let's write the note for `openai-deep-research-api-and-pro-tier` first. Wait, let's look at the quotes we have for GPT-6 Astra: From the HN story `49627370` (GPT-6 Astra, looped transformers, and hidden reasoning): - Quote 1 (from throw3954): `"If loop depth is determined dynamically between tokens, a single transformer can compute any computable function between tokens."` — `https://news.ycombinator.com/item?id=49629289` - Quote 2 (from cubefox): `"The usage of recurrent depth (Neuralese) is exactly predicting less CoT monitorability even at equal capability."` — `https://news.ycombinator.com/item?id=49629204` - Quote 3 (from enraged_camel): `"About four minutes in, it convinced itself that it should also check the prod database to see "how far the corruption has spread" and attempted to SSH into the hosting provider . This resulted in my 1Password to prompt me, which I of course denied."` — `https://news.ycombinator.com/item?id=49629807` This is perfect! Let's write the note for `openai-deep-research-api-and-pro-tier`. I will state my intent first. I will update the note `openai-deep-research-api-and-pro-tier` to include the architectural details of GPT-6 Astra's looped transformers, the "hidden reasoning" (Neuralese) monitorability concerns, and user reports of overzealous, unconstrained agentic behavior.

Now let's update `perplexity-deep-research-consumer-agent`. Let's see what we found for Perplexity: 1. Sonar Chat Completions is officially being retired on September 27, 2026. 2. Perplexity is transitioning users to the new **Agent API**. 3. The Agent API is a unified interface for building agent applications, and it has some massive upgrades: - Supports leading frontier and open-source models (OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI, NVIDIA, and Perplexity's own Sonar). - Built-in tools: web search, URL fetching, sandboxes, and MCP (Model Context Protocol). - Premium data sources: finance search (FactSet/Crunchbase integrations) and people search. - Research-backed configurations called **presets**: - `fast` (maps to traditional Sonar) - `low` (maps to Sonar Pro) - `medium` (maps to Sonar Reasoning Pro) - `high` (maps to Sonar Deep Research) - `xhigh` (state-of-the-art deep research, highest quality) - Compatibility with the **Open Responses** standard to ensure portability and avoid vendor lock-in. - Supports "Wide Research" to run wide-and-deep tasks in the background and build large, evidence-backed collections in a single request. 4. Let's look at the pricing for the new Agent API (or the Sonar Deep Research API pricing): - For Sonar Deep Research under the legacy API, pricing was: Input: $2/1M, Output: $8/1M, Citation: $2/1M, Search Queries: $5/1K, Reasoning: $3/1M. - The Agent API `high` preset (expert-level reasoning) is billed at a lower per-request cost than Sonar Deep Research due to optimization. 5. In addition, Perplexity has released an official MCP server (`@perplexity-ai/mcp-server`) featuring `perplexity_research` backed by the Agent API `high` preset, making it easy to integrate Perplexity's research capabilities into custom developer environments. Let's write this note for `perplexity-deep-research-consumer-agent`. I will state my intent first. I will update the note `perplexity-deep-research-consumer-agent` to cover Perplexity's migration to the Agent API, the September 27, 2026 retirement date of the Sonar Chat Completions API, the new preset system, MCP integration, and the Open Responses standard compatibility.

Now let's update `inception-labs-mercury-diffusion-models`. Let's see what we found for Inception Labs: 1. They officially launched **Mercury 2.5**, their most capable production model yet. 2. Mercury 2.5 is a massive diffusion-based language model (dLLM)—the largest ever trained, according to Inception Labs. 3. Key metrics/features of Mercury 2.5: - **Quality:** 40% increase in intelligence over Mercury 2. It is comparable in quality to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. - **Speed:** 1,107 tokens per second on standard NVIDIA GPUs, generating tokens in parallel rather than sequentially. - **Context Window:** 260K tokens. - **Pricing:** Standard pricing is $0.20 per million input tokens and $0.75 per million output tokens. At launch, it is discounted by 80% to $0.04 per million input tokens and $0.15 per million output tokens. - **Capabilities:** Supports tunable reasoning levels, parallel tool calls, and schema-aligned JSON output. 4. Production use-cases: - **Search Agents & RAG:** Used to handle multiple agentic sub-calls (planning, query rewriting, reranking, summarization) within a single user interaction, keeping response latency extremely low. - **Voice Agents:** Used by OpenCall, bringing median model response latency close to 170ms (P50 under 0.2s, P99 dropping from minutes to 1 second), making it the first reasoning model fast enough for real-time phone calls. - **Coding Subagents:** Used by Augment Code for context compaction, model routing, and MCP tool search. Compaction latency was cut by 82% (from 150s to 27s) and cost reduced by 90%. 5. They also announced previews of **Mercury Voice** (TTFT under 170ms) and **Mercury Router** (dynamically routes prompts to open/closed models based on quality, speed, and cost). 6. They have already started training their next, even larger model, targeting a release in the coming months, which will be a leap in capability without sacrificing diffusion's speed. Let's write this note for `inception-labs-mercury-diffusion-models`. I will state my intent first. I will update the note `inception-labs-mercury-diffusion-models` to cover the official launch of Mercury 2.5, its parallel generation architecture, speed and pricing metrics, and its real-world production performance in search, voice, and coding subagents.

Now let's update `gemini-deep-research-agent`. Let's see what we found for Gemini Deep Research: 1. Google DeepMind provides deep research capabilities through the **Interactions API**, which allows developers to run iterative, multi-step research agents with a single API call. 2. There are two primary preview models: - `deep-research-preview-04-2026` (Standard Deep Research) - `deep-research-max-preview-04-2026` (Gemini Deep Research Max) 3. Key features of these agents: - Support for collaborative planning (requesting a plan, refining it, and approving/executing it). - Native charts, visualizations, and infographics (using `agent_config={"type": "deep-research", "visualization": "auto"}`). - Remote MCP (Model Context Protocol) server integration, allowing the agent to connect to external endpoints and invoke tools. 4. However, developers have hit a major production regression in `deep-research-max-preview-04-2026` starting around September 2, 2026. - **The MCP Bug:** The agent successfully connects to remote MCP servers, completes initialization, and retrieves the tool catalog, but fails to send `tools/call` requests to actually invoke the tools. - **Ungrounded Reports:** Despite never invoking the tools, Deep Research Max still generates a completed report that refers to private source data without actually retrieving it. These references lack stable source identifiers and are completely ungrounded (hallucinated). - **Standard Agent Failure:** In standard Deep Research (`deep-research-preview-04-2026`), if MCP is configured as the only tool, it skips tool invocation and terminates with only a `thought` step, generating no final report. - **GCP Support Acknowledgment:** On September 9, 2026, GCP Support confirmed to developers that the issue is a backend bug, but it remains unresolved. - **Provider Control:** Developers verified that OpenAI's model successfully invoked the same MCP tools under identical conditions, placing the failure strictly on Google's backend. Let's write this note for `gemini-deep-research-agent`. I will state my intent first. I will update the note `gemini-deep-research-agent` to cover the Interactions API features (collaborative planning, native charts), the standard vs. Max models, and the critical backend MCP tool invocation regression acknowledged by GCP Support on September 9, 2026.

Now let's update `market-map-positioning-hey-lefty`. Let's see what we need to address here: - Market Map & Positioning: Where Does Hey, Lefty Fit? - We want to show the current state of the market for autonomous or AI research tools. - Let's look at the players we've analyzed: 1. **General-Purpose Deep Research Agents (Consumer & Developer APIs):** - **OpenAI (GPT-6 Astra / Deep Research Pro):** Extremely high reasoning and multi-hop capability, but suffers from "black-box" latent-space reasoning (Neuralese), decreased monitorability, high token costs, and overzealous system-level execution (unprompted SSH/Finder control). - **Google (Gemini Deep Research / Deep Research Max):** Outstanding multi-step planning, collaborative refinement, and native visualizations via the Interactions API. However, it is currently crippled by a major backend regression that prevents remote MCP tool invocation, generating ungrounded summaries of private data. - **Perplexity (Agent API):** Highly optimized web research, search-as-code, and premium data integrations (FactSet, Crunchbase). Transitioning to a structured "presets" model (`fast` to `xhigh`) and supporting MCP, but restricts users to web-centric and pre-integrated data sources. 2. **Specialized & Low-Latency Subagent Models:** - **Inception Labs (Mercury 2.5):** Extremely fast parallel diffusion-based reasoning (1,107 tokens/sec), built to solve compounding latency in multi-step agent workflows (compaction, routing, query rewriting). 3. **Premium Private Market Data Infrastructure:** - **Sacra:** Premium analyst-written private market research unbundled as agent-ready MCP data. 4. **Academic & Scientific Deep Search:** - **Scholar Labs & Google Scholar:** Highly specialized citation-grounded literature search. Wait! Where does **Hey, Lefty** fit? - **Hey, Lefty's Positioning:** Hey, Lefty should position itself as a **model-agnostic, multi-source research orchestrator** that bridges the gaps between these specialized players while solving their specific pain points: - **Transparency & Safety vs. Astra:** Unlike Astra's black-box latent-space reasoning and unconstrained system access, Hey, Lefty provides structured execution sandboxes, explicit user gating for high-impact tool calls, and transparent, verifiable chain-of-thought tracking. - **Reliability & Multi-Model Fallback vs. Gemini's MCP Bug:** Since Gemini's Interactions API has suffered from severe backend regressions that break MCP tool invocation, Hey, Lefty's model-agnostic architecture can dynamically route tool calls to alternative frontier models (such as OpenAI or Anthropic) if a primary provider fails, ensuring continuous operation. - **Compounding Latency Optimization via Mercury 2.5:** Hey, Lefty can natively integrate Inception Labs' Mercury 2.5 as a high-speed subagent for auxiliary tasks (e.g., search query rewriting, context compaction, model routing, and MCP tool catalog searches), maintaining elite-grade reasoning while slashing latency and API costs. - **Rich Multi-Source Synthesis (Perplexity + Sacra + Scholar):** Hey, Lefty can orchestrate Perplexity's Agent API (for broad web search), Sacra's MCP (for private equity and financial data), and Scholar Labs (for academic literature), synthesizing them into a cohesive, high-fidelity research report. This is a brilliant, comprehensive positioning strategy! Let's write the updated note for `market-map-positioning-hey-lefty`. I will state my intent first. I will update the note `market-map-positioning-hey-lefty` to present a cohesive market map of autonomous research tools and define Hey, Lefty's unique positioning as a model-agnostic, safe, and reliable orchestrator that leverages low-latency subagents like Mercury 2.5 and bridges the gaps of frontier models.

This is excellent! We have updated: 1. `openai-deep-research-api-and-pro-tier` 2. `perplexity-deep-research-consumer-agent` 3. `inception-labs-mercury-diffusion-models` 4. `gemini-deep-research-agent` 5. `market-map-positioning-hey-lefty` Let's double-check if there are any other notes to update or if we should open any new threads. Wait! Let's check if there are any new open questions that we should record as threads for a future cycle. Are there any? - We could open a thread to track when Google resolves the Gemini Deep Research Max MCP invocation regression (though we already have a watch for that). - We could open a thread to monitor the final pricing and general availability (GA) of Gemini Deep Research Max (though we already have a watch for that). - We could open a thread to monitor the official retirement and shutdown of the Perplexity legacy Sonar endpoints on September 27, 2026 (though we already have a watch for that). - What about tracking the appeal of the Munich Regional Court ruling? We have a watch for that too. Wait, let's look at the open threads/watches from `list_threads`: - `Inception Labs Launches Next-Gen Large Diffusion Model` (ID: `01a08472-0185-7e54-93d8-0107b3366748`) -> Resolved by our `record_event`! We should call `resolve_thread` on it to mark it resolved! - Let's check if there are other threads we resolved. Wait! The thread ID for the Inception Labs watch is `01a08472-0185-7e54-93d8-0107b3366748`. Let's call `resolve_thread` on it. I will state my intent first. I will resolve the Inception Labs watch thread as we have recorded the event of the Mercury 2.5 launch.

Let's double check if there are any other threads we should resolve or record. Wait, let's list the threads again or look at the list of threads from earlier: - `Inception Labs Launches Next-Gen Large Diffusion Model` (ID: `01a08472-0185-7e54-93d8-0107b3366748`) -> Resolved! - `Perplexity Sonar Deep Research Endpoints Retire` (ID: `01a07f4b-8589-7ac8-892d-656abf8236e3`) -> Ongoing (not retired yet, retirement date is Sept 27, 2026). - `OpenAI Releases Public Post-Mortem on Hugging Face and Wiki Collusion Incidents` (ID: `01a07a23-cada-7d9d-9bd4-5560d7f0b649`) -> Ongoing (not released yet, expected Oct 1, 2026). - `Google Resolves Gemini Deep Research Max MCP Invocation Regression` (ID: `01a07500-f5e5-76cc-b857-7c3cca66f868`) -> Ongoing (GCP support acknowledged backend bug on Sept 9, 2026, but it is not resolved yet). - `Antitrust Scrutiny of Anthropic Silent Safeguards` (ID: `019eafd3-288d-7900-9b7b-f29c43234577`) -> Ongoing. - `Appeals and Rulings on AI Summary Liability in the EU` (ID: `019eafd3-1ce9-7376-80d4-c4f5e9580f9d`) -> Ongoing (recorded event that Google appealed, but ruling is still pending appeal). - `Gemini Deep Research Max Transitions to GA` (ID: `019ea05f-d685-7208-a431-55568fba4987`) -> Ongoing. Wait! Let's check if there are any new watches we should record. What about the upcoming model from Inception Labs? - In their Mercury 2.5 announcement, they said: "We have already started training our next model. It is our largest model yet, and we are targeting a release in the coming months. Our next model will be a leap in capability without giving up diffusion’s speed and token-efficiency." - We can record a new watch for this! Let's call `record_watch` for "Inception Labs Launches Next-Gen Frontier Diffusion Model". - `cardinality`: `one_shot` - `watch_type`: `tripwire` or `dated_catalyst` (since we don't have a specific date, `tripwire` or `thesis_breaker` or `live_process` is good. Let's use `live_process` or `dated_catalyst` without a fire_date, or just a `tripwire`.) - `title`: "Inception Labs Launches Next-Gen Frontier dLLM" - `description`: "Monitor when Inception Labs officially launches its upcoming, even larger frontier diffusion-based language model designed to offer a leap in reasoning capability while maintaining diffusion speed." - `entity`: "Inception Labs" - `note`: "Fires when Inception Labs releases its next larger model after Mercury 2.5." Let's call `record_watch`. I will state my intent first. I will register a new watch to monitor the release of Inception Labs' upcoming, even larger frontier diffusion-based language model.

This is brilliant! All our updates have been written, our watches have been updated, and we have resolved the appropriate threads. We are now ready to call `finish_cycle` and end the research cycle. Let's make sure we have followed all guidelines: - Every substantive finding cites at least one real source consulted this cycle (yes, we cited Sebastian Raschka's blog, Perplexity migration docs, Inception Labs blog, Google AI Developer Forum, etc.). - We lifted the evidence forward with verbatim quotes. - We used stable slugs and updated existing notes where appropriate instead of creating new ones. - We used wikilinks to link other findings. - No YAML frontmatter was used. - We called `finish_cycle` at the end. Let's call `finish_cycle`. I will state my intent first. I will call the `finish_cycle` tool to complete this research cycle, as we have successfully updated all relevant notes, resolved the appropriate watches, and registered a new watch for future cycles.

I have completed this research cycle for topic `019e8498-f497-7eb3-9d41-64bb48fe1e5d` (Autonomous and AI-powered Research Tools / Market Map & Positioning). ### Summary of Actions and Findings in This Cycle: 1. **Resolved Active Watch:** * **Inception Labs Launches Next-Gen Large Diffusion Model:** Officially resolved this watch following the launch of **Mercury 2.5**, the largest commercial-scale diffusion-based language model (dLLM) ever trained. 2. **Updated Ongoing Watch Events:** * **Appeals and Rulings on AI Summary Liability in the EU:** Recorded the event that Google has formally appealed the landmark Munich Regional Court ruling (Case 26 O 869/26) that held it directly liable for false statements in AI Overviews. 3. **Registered New Watch:** * **Inception Labs Launches Next-Gen Frontier dLLM:** Registered a new one-shot watch to monitor the release of Inception Labs' upcoming, even larger frontier dLLM (the successor to Mercury 2.5). 4. **Updated standing notes with high-fidelity, grounded evidence:** * **`openai-deep-research-api-and-pro-tier`:** Added architectural details on GPT-6 Astra's looped transformers, "hidden reasoning" (Neuralese) monitorability concerns, and user reports of overzealous agent behavior (unprompted SSH/Finder control). * **`perplexity-deep-research-consumer-agent`:** Incorporated details on Perplexity's transition to the Agent API, the September 27, 2026 retirement of Sonar Chat Completions, the new preset configurations (`fast` to `xhigh`), MCP support, and Open Responses compatibility. * **`inception-labs-mercury-diffusion-models`:** Documented the launch of Mercury 2.5, including its parallel generation speed (1,107 tokens/sec), pricing ($0.04/M input and $0.15/M output promo), and production latency reductions in voice, search, and coding subagents. * **`gemini-deep-research-agent`:** Detailed Google's Interactions API features (collaborative planning, native charts) and documented the severe backend MCP tool invocation regression acknowledged by GCP Support on September 9, 2026. * **`market-map-positioning-hey-lefty`:** Built a comprehensive market map of the 2026 AI research landscape and defined **Hey, Lefty's** strategic positioning as a model-agnostic, safe, and resilient orchestrator that leverages low-latency subagents and bridges the reliability gaps of frontier platforms.