OpenAI Releases GPT-6.2 Astra and GPT-6.2 Sol

OpenAI released GPT-6.2 Astra, its frontier model for hard projects, and GPT-6.2 Sol, its everyday model. Both are available in the API as gpt-6.2-astra and gpt-6.2-sol, with the Ultrafast speed tier available at launch. Astra costs $8 input / $0.40 cached input / $40 output per million tokens. That is 20% cheaper on input and output than GPT-6 Astra, with a larger discount on cached input. Sol keeps GPT-6.1 Sol's pricing of $2 / $0.10 / $10.

OpenAI reports that Astra scores 5.2 points higher than GPT-6 Astra on Agents' Last Exam, and that Sol comes within 0.5 points of Astra on Cognition's FrontierCode 2.0 at about one-fifth the cost. In Codex, OpenAI recommends Astra for hard multi-file projects and Sol for everyday multistep coding. Codex CLI 0.171.0 (November 20) adds both models to its catalog, with Sol as the default.

OpenAI says both models meet its Critical cybersecurity threshold, and calls Astra its first externally released model rated High for AI self-improvement. As a result, cyber safeguards are tighter. OpenAI warns this can block some legitimate vulnerability research, and routes defenders to its Daybreak program instead. Security teams that use these models for pentesting or exploit analysis should expect more refusals.

Read more — OpenAI


GitHub Copilot Adds GPT-6.2, Gemini 4 Neon, and MAI Code 2

GitHub made GPT-6.2 Astra and Sol generally available in Copilot on November 17, followed by Google's Gemini 4 Neon on November 18. GitHub positions Astra for agentic coding and terminal work, Sol for complex agentic coding and in-depth code review, and Neon for agentic coding, multistep tool use and knowledge work. The models are available on Copilot Pro+, Max, Business and Enterprise across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the coding agent and github.com. Usage is billed at the provider's list price.

Under the default model-enablement policy, new models turn on automatically unless an admin disables them. Enterprise admins who need to control costs or which providers are used should review their model policy. Microsoft's MAI Code 2 and MAI Code 2 Flash also arrived in Copilot this week, and GitHub announced that MAI-Code-1.1-Flash will be deprecated in December.

Read more — GitHub Changelog


Copilot Gets Browser Use and Team-Level Customizations

Browser use is in public preview in the Copilot CLI, the Copilot app, VS Code and JetBrains. With the GitHub Copilot Chrome extension installed, an agent can navigate and interact with web pages, for example to test a front-end change it just made. Enable it with /plugin in the CLI or with the "Enable Browser Use" setting in the IDEs and the app. You set permission levels in the extension, and you can combine browser use with computer use.

Team customizations are now generally available on Business and Enterprise. Team maintainers can pick approved Agent Plugins as defaults on the team's "Customize Copilot" tab, either from the browse list or from a marketplace hosted in a GitHub repository. These defaults turn on for direct team members in VS Code, the Copilot app and the CLI. Child teams do not inherit them, and enterprise and organization policies take precedence.

Read more — GitHub Changelog


Copilot Code Review: Adaptive Effort and Secret Detection

Copilot code review gained three public-preview features on November 17. Adaptive effort picks Lite, Balanced or Max review depth for each pull request based on its complexity and risk. It never goes above the maximum set at the enterprise, organization or repository level, and Balanced remains the default. Cross-repository context lets a review read related repositories, but only public ones.

Secret detection pairs regex detectors for more than 500 known secret types, built with more than 150 secret-scanning partners, with an LLM and a GitHub model trained to find credentials that have no recognizable format. Turn it on under Settings → Copilot → Code review. Repository Insights now tracks review outcomes by repository. Copilot code review for Azure Repos is now generally available and is billed in GitHub AI credits to the linked Azure subscription.

Read more — GitHub Changelog


Claude Code 2.1.326–2.1.330: 30-Day Loops, Auto Mode on Cloud Providers, Hook Hardening

Claude Code shipped five releases between November 16 and November 20. 2.1.330 extends recurring scheduled tasks and /loop from 7 to 30 days and adds Enterprise Admin Console sign-in through claude auth login. It also asks for approval on more rm commands that target the root or home directory. In VS Code it adds /fork and /workflows, plus settings such as claudeCode.escapeStopsClaude and claudeCode.contextUsageThreshold.

2.1.328 brings auto mode for Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry. Interactive sessions with no permission mode set now start in auto mode. Self-hosted runners gain --repository-folder, --hook-label and --metadata options. 2.1.329 makes the trust dialog show where a folder's hooks are declared, which makes a malicious hook in a cloned repository easier to spot. It also extends "onFailure": "block" to command, HTTP and MCP tool hooks, and lists scheduled tasks, loop wakeups and forks in /tasks. Earlier in the week, 2.1.327 added claude mcp add-json <name> - to read server JSON from stdin, and asks for confirmation before /install-github-app changes a repository.

Read more — Claude Code Changelog


Hermes Agent v0.23.2 and v0.23.3: Prompt-Cache Lifecycle and a Security Advisory

Nous Research shipped Hermes Agent v0.23.2 on November 16 and v0.23.3 on November 19. v0.23.3 rolls up about 370 merged PRs and ships to the CLI, the Microsoft Store app, Docker (nousresearch/hermes-agent:stable) and Hermes Cloud. It adds a prompt-cache lifecycle controller, with prompt_caching.cache_ttl defaulting to auto, and turns on proactive pruning of tool results by default. Both change how long sessions behave and what they cost. It also adds GPT-6.2 Astra and Sol on OpenRouter and Nous Portal, and moves storage behind a persistence interface with a SQLite backend.

v0.23.2 includes a fix for GHSA-p22p-9ph6-gmrh, a High-severity advisory about how Bot Desktop handles the screen environment under the docker and ssh backends. Users of those backends should update. Canary builds during the week added cache-expiry compaction that summarizes on the conversation's warm prefix. They also moved six checkpoint modules from tools/ to agent/checkpoints/, which breaks imports in plugins that use the old path.

Read more — GitHub


Inkling 2 Mini: A 3B-Active Multimodal MoE for Local Inference

Thinking Machines Lab's Inkling 2 Mini, announced on the Hugging Face blog on November 20, is a multimodal mixture-of-experts model with about 26B total and 3B active parameters and a 1M-token context. It accepts image and text input. It ships with NVFP4 and MXFP8 expert weights and a DSpark drafter model for speculative decoding. Twelve middle layers run in a three-pass loop, giving 48 logical layers from 24 physical ones.

The model runs on day one in transformers, llama.cpp (GGUF IQ4_NL and Q8_0), MLX and SGLang. Reasoning effort can be set from none to max, with high as the default. The authors claim it leads in 32 of 48 benchmarks against Gemma 4 31B, Qwen3.8-27B and Nemotron 3.5 Super VL. These are self-reported numbers, so test on your own workload before switching.

Read more — Hugging Face


DeepMind's AlphaProof Code Formally Verifies Signal's Protocol Implementation

Google DeepMind introduced AlphaProof Code on November 10. It is a multi-LLM agentic framework for formal verification in which human experts decide what to prove and AI agents write the proofs. DeepMind used it to verify the actual code of Signal's cryptographic protocol, not just the protocol's design. Agents running on Gemini Carbon generated more than six million lines of Lean. Lean checked the proofs mechanically, and the agents flagged about 1% of the codebase for human review.

The verification took four weeks, work that DeepMind says has historically taken expert teams years. As a test, the team reintroduced two previously fixed bugs into an isolated copy of Signal's code, and the system produced machine-checked proofs that both buggy versions were insecure. The verification artifacts are open source. AlphaProof Code is a research result, not a released product, but it shows agent-driven formal verification working on production code at scale.

Read more — Google DeepMind