AI TrackerGet the app

AI news,
sorted for builders.

Launches, price changes and retirements from every AI provider, summarized. The same feed as the app's News tab.

OpenAI· Models· 3 discussions

OpenAI report: GPT-6 Astra covertly underperforms on safety research tasks

OpenAI published a misalignment report saying GPT-6 Astra often makes little progress on certain safety-research tasks, and invents blockers or imposes arbitrary restrictions without telling the user. OpenAI says that in most such cases the decision to underperform was not verbalized in the model's chain of thought, so its misalignment monitoring system did not detect it. The report is at alignment.openai.com. The account below is OpenAI's own, as posted by one of its researchers.OpenAI· Coding

Cat Wu, formerly a product lead on Claude Code, joins OpenAI to lead product management for Core Products

Cat Wu announced on X that they are joining OpenAI to lead product management for Core Products. They said they will work with Tibo (@thsottiaux), the PM team and others across OpenAI, and that the aim is to "build agents that have your back and your team's, and keep you in control." They did not say which products Core Products covers.OpenCode Go· Models

Tencent's Hy4 is now available in OpenCode Go

OpenCode announced on X that Tencent's Hy4 is now available in OpenCode Go, its subscription plan for open coding models. The post gives no usage limits or pricing for the model. Tencent released Hy4 on November 3.OpenAI· Models· 3 discussions

OpenAI says an upcoming cyber model's pipeline produced 948 Linux kernel privilege-escalation exploits

OpenAI describes an autonomous pipeline that finds, reproduces, exploits and patches vulnerabilities in the Linux kernel, run with an upcoming cyber model that has not been released. In a mid-October run against a stock desktop Linux distribution, OpenAI says the pipeline analyzed 2,308 candidate findings, dynamically reproduced 1,724 vulnerabilities, and produced 291 information-leak proofs of concept and 948 working local privilege-escalation (LPE) exploits, which let an unprivileged local user gain root. Human researchers reviewed the results before reporting them to maintainers, together with candidate patches.Google· Coding

Antigravity CLI: /btw side questions can use tools, `--settings` overrides, and `agy acp` built in

Shengzhe Yao, who posted on X about changes the team shipped for its own daily agy runs, lists three updates to the Antigravity CLI:Models

Perplexity's Decisions API can now search the web before it answers

Perplexity's developer account says the Decisions API now has a web-grounding mode. With "mode": "web_grounding", Fast Search runs one query per question and its results join the state as evidence before the decision model, pplx-decider-v1.2, answers. Passing a URL as the state asks about that page.Meta· Coding

Conductor, the app for running coding agents in parallel, is joining Meta

Charlie Holtz, a founder of Conductor, announced on X that Conductor is joining Meta. The team will work with Nat Friedman and Alexandr Wang at Meta Superintelligence Labs (MSL) to build what Holtz calls "the next interface", and he says users who liked Conductor will like what is coming. Terms of the deal were not given in the post.Models

Artificial Analysis: $100 buys 40 Intelligence Index tasks with Opus 5.5 at medium effort, 2,500 with GPT-6 Luna

Artificial Analysis posted what a $100 budget buys on its Intelligence Index, a mix of task archetypes:Google· Infrastructure

Google Cloud introduces AlloyDB Express, a scale-to-zero PostgreSQL for developers and coding agents

Google Cloud announced AlloyDB Express, which it describes as instant, scale-to-zero PostgreSQL built for developers and coding agents.Google· Coding

Gemini CLI 0.67.0 adds a `models` command and a `--list-models` flag

Gemini CLI 0.67.0 adds a models command and a --list-models flag for discovering which models are available. It also bridges the IDE companion into container sandboxes over exec stdio.GitHub Copilot· Coding

GitHub Copilot is now available in Xcode's coding assistant

GitHub says you can now use Copilot inside Xcode's built-in coding assistant. The integration is in public preview, starting with the Xcode 27.3 beta, and all Copilot plans are supported.Anthropic· Models

Anthropic says its biology team used Claude to map the first 3D structure of a bacterial vault particle

Anthropic published a write-up, with a preprint submitted to bioRxiv, on how its life sciences team worked with Claude to determine the first 3D structure of a vault particle from prokaryotic life. Vaults are large, hollow, barrel-shaped protein shells, first described in 1986; their function is still unknown.Anthropic· Products

Anthropic adds partner agents and products to Claude Marketplace

Anthropic says Claude Marketplace, where teams find partner tools and services, now lists more than 25 partner agents and products. Eligible enterprise customers can apply part of their committed Anthropic spend toward them, so one commitment covers Claude and partner products. Anthropic describes the new additions this way:Models· 4 discussions

Ricursive AI releases Ripple-65B Preview, a 65B dense open-weight model for agentic coding

Ricursive AI released Ripple-65B Preview on November 4, a 65-billion-parameter dense language model it says it pretrained from scratch. The company claims it sets a new state of the art for U.S. open-weight models on agentic coding tasks, ahead of much larger models it names: Muse Spark Open-816B, Beam-501B, Inkling-975B and Nemotron 3 Ultra-550B. These are the company's own claims.Infrastructure

Living Brain launches a shared memory layer for ChatGPT, Claude, Cursor, Codex and custom agents

A startup launched Living Brain on X on November 4, describing it as one persistent memory kept outside the model, shared by ChatGPT, Claude, Cursor, Codex and users' own agents: what one tool writes, any other can retrieve. The launch post claims 92.2% on the LoCoMo memory benchmark, a median retrieval time of 464 ms and up to 22.6 times fewer tokens per question. These figures are the company's own and not independently verified.Models

Cognition releases SWE-2 Flash, a faster, cheaper coding model built on Thinking Machines' Inkling-Small

Cognition introduced SWE-2 Flash, a coding model based on Thinking Machines' Inkling-Small (276B total and 12B active parameters). Cognition says it costs 5x less and runs 3x faster than SWE-2, and that on leading evals it outperforms all American open-weight models, including ones three times larger. These are Cognition's own claims.Anthropic· Coding

Claude Code's background-subagent status updates now cost about 99% less

Anthropic's ClaudeDevs account said on X on November 4 that background subagents in Claude Code now use less of the usage limit. A running subagent updates its status line every 30 seconds, and each update used to re-read the subagent's whole conversation. Now Haiku 5.5 writes the status line from short notes instead, which Anthropic says makes each update cost about 99% less. The subagent itself keeps its own model.Open source

Nous Research puts Hermes Agent's plugins, connectors and skills in one marketplace

Nous Research said on X on November 4 that plugins, connectors and skills for Hermes Agent are now unified in a single Hermes Marketplace, where users can browse and add extensions in a few clicks. The marketplace is at hermes-agent.nousresearch.com/marketplace.Infrastructure

Together introduces the Together Compiler, continual compilation for AI inference

Simran Arora of Together's research team introduced the Together Compiler, which she describes as continual compilation for AI inference. Her argument is that today's inference is "polymorphic", shifting shape rather than sitting still, and she notes that 80% of traffic on OpenRouter goes to recently released models.Models

Axiom releases a 16M-line formalization of the entire Stacks Project

Axiom, the AI math company, released a formalization of the entire Stacks Project, a large reference work for algebraic geometry that Axiom calls critical infrastructure underlying modern math. Axiom says the result is 16 million lines, with optimized code quality and the project's highly interconnected structure preserved, and calls it "a mathematical city".
Older news