Skip to content
DigIntoAI/Model intelligence
Model intelligence feed

Model intelligence.

Follow the changes. Understand what comes next.

Editor’s pickGoogle
Google/Gemini 4 Neon

Google Launches Gemini 4 Neon in Antigravity With 1M Context, Four Thinking Levels, and Benchmarks vs. Claude Opus 5.5 and GPT-6.2 Sol

Announced on November 19, 2026 at 05:51 Beijing time, Google's Gemini 4 Neon features a 1M-token context window, four thinking levels, and planner-executor pairing with Gemini 4 Hydrogen in Google Antigravity, outperforming Claude Opus 5.5 and GPT-6.2 Sol across most of 18 published benchmarks.

Latest stories349

Newest published
205

AWS Opens Infinity, a Research Runtime for Extreme Agent Concurrency

AWS released Infinity under Apache 2.0 as part of Hydro with UC Berkeley. The research runtime suspends an agent during a tool call, persists the conversation, and resumes when the result arrives—so an idle agent mostly costs its history, and a serverless function can end while it waits. AWS’s own blog benchmarks cite about 100 KB per idle agent and about 75,000 concurrent agents in 8 GB. Early research release seeking feedback, not production GA. Separate from OpenAI Agents API scheduled automations.

InfinityRead story
206

Replit Monitoring Agent Can Investigate, Fix and Auto-Publish Production Issues

Replit introduced Monitoring Agent—an autonomous on-call for published apps. It can read production logs and metrics, investigate an alert, make and validate a fix, and email a summary. With auto-publish on, it can deploy a validated fix to production. Earlier App Monitoring covered health checks and downtime alerts; this release adds the fix-and-ship step. Enable monitoring alerts and auto-fix in Publishing settings; auto-publish is optional. Separate from Replit Agent’s default model choice.

Monitoring AgentRead story
207

Thariq Leaves Anthropic After a Year on Claude Code

Thariq Shihipar (@trq212) posted that they have left Anthropic after more than a year on Claude Code. It’s a personal farewell—not an Anthropic company announcement. The post thanks teammates and the X community, says a short break comes first, then something new. For Claude Code updates, it points to @addyosmani, @lydiahallie, @_catwu, @ClaudeDevs, and @bcherny.

Claude Code / Thariq ShihiparRead story
208

Block Releases Dessa, a Full-Duplex Speech Model Trained Only on Speech

Block’s small AI research lab shipped Dessa: speech in, speech out, full-duplex. It says the model was trained from scratch on human conversation audio only—no text, no pretrained base. Research preview for now: a small group of AI safety researchers get early access; everyone else joins a waitlist on the official page. Block says inference beats real time and the model runs locally on a MacBook Pro M4 Max or better. Training, per the page, used one NVIDIA GB200 NVL72.

DessaRead story
209

Why Agents API Scheduled Automations Show Completed Before Work Finishes

OpenAI’s Agents API Scheduled Automations can start or continue agent sessions on a cron schedule with a time zone, or once at a Unix timestamp. An invocation’s status moving to completed means the task was delivered to the target session—not that the agent finished the work. Periodic runs add a small randomized delay; a busy session queues the task without interrupting an active turn. Pause or delete does not stop work already accepted. The interface is still beta for developers with Agents API access, not a ChatGPT subscription benefit. It is not the same product as ChatGPT Team Tasks.

Agents API Scheduled AutomationsRead story
210

Simular Agent S4 Evolves Interfaces Without Retraining the Model

On October 21, 2026, Simular introduced Agent S4, which evolves the interface repository computer-use agents work through while the task agent’s model and harness stay fixed. With Qwen3.8-27B and an evolved repository (no scratch VM), it reaches 21.3% binary success on the full 108-task OSWorld V2 set—about 3.8× the same model with fixed interfaces (5.56%), and above GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%) at a small fraction of their reported task cost. Evolving an interface repository is not the same as Cursor’s self-hosted desktop computer use.

Agent S4Read story
211

UK Opens World's First AI Biosecurity Lab

On October 24, 2026, the UK Cabinet Office announced the AI Biosecurity Lab (AIBL)—the world’s first dedicated AI biosecurity organisation, per the government. It has £100 million in government funding and up to £200 million in commitments from OpenAI and Anthropic (combined package up to about £300 million). Compute and model credits go to the lab, not to personal API or coding quotas. Partners also include Google DeepMind and Basecamp Research. AIBL will be an independent non-profit focused on earlier detection of biological threats and faster medical countermeasures.

AI Biosecurity Lab (AIBL)Read story
212

OpenAI Tightens Isolation for Frontier Model Training

On October 23, 2026, OpenAI published how it secures frontier-model training and evaluation. The company says frontier training now runs with no live internet access, using an “offline world” of indexed and cached web content, plus defense-in-depth sandboxes and perimeters. After a September 20 incident and a pause, it resumed frontier RL training and internal inference on October 1. Red-teaming found open-source flaws disclosed under Patch the Planet, including Brotli CVE-2026-111497 covered in a separate note.

Frontier training environmentRead story
213

Brotli Unauthenticated RCE Affects Node, Bun and Envoy

CVE-2026-111497 is an out-of-bounds write in the Google Brotli decoder. Crafted compressed input can crash a process, leak memory, or enable remote code execution when an app decompresses attacker-controlled Brotli data. Fixed lines include Node.js 22.23.4 / 24.22.1 / 26.12.1, Bun 1.4.4, Envoy 1.40.0 and backports, Istio patched lines, and upstream Brotli 1.0.10 / 1.1.1 / 1.2.1. OpenAI Codex Security researchers reported the issue; Patch the Planet tracks mitigations. Until you can upgrade, disable Brotli decompression of untrusted data.

Brotli / CVE-2026-111497Read story
214

Claude Code Adds MCP Tasks and Mid-Job User Prompts

At 02:24 Beijing time on October 24, 2026, ClaudeDevs said Claude is adding MCP Tasks and Elicitations. Tasks let Claude hand plugins long-running jobs such as deep research and get results when they finish. The posts say both are available today in the Claude Code CLI, with more Claude products to follow. Official docs call MCP tasks a public beta that needs Claude Code v2.1.304 or later; elicitation is a mid-job dialog for structured input (form or URL).

Claude Code MCP Tasks / ElicitationsRead story
215

Codex for Open Source Offers Maintainers Six Months of ChatGPT Pro

OpenAI’s Codex for Open Source program supports open-source maintainers with six months of ChatGPT Pro that includes Codex, conditional access to Codex Security, and API credits through the Codex Open Source Fund. Applications are reviewed on a rolling basis; selected applicants are notified by email. The application asks for a ChatGPT-account email and a public GitHub username and repository URL. This program is separate from our notes on the Codex Ambassador Program.

Codex for Open SourceRead story
216

Cursor Self-Hosted Workers Can Now Drive Your Desktop

Cursor’s self-hosted workers can now use Computer Use on macOS and Linux: the agent can click, type, take screenshots, and drive desktop apps (Cursor’s post cites the iOS Simulator; docs also cover a browser when Chrome or Chromium is installed). Enable with agent worker --computer-use start. This is not the same as our notes on iPhone Remote Control for local agents, or cloud-agent Canvas and Environments.

Cursor Self-Hosted / Computer UseRead story