tech-update

What changed in AI this month, and what you can try today.

Updated

6 Oct – 4 Nov 2026

Open-weight models closed in on the frontier, while AI-found bugs and agent plugins showed new security risks.

  1. Open-weight models are close to the frontier on agentic coding

    Moonshot's Kimi K3.1 Pro, a 2.8T mixture-of-experts model, reports 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5, and its weights are now on Hugging Face. Mistral released weights for the 1T-parameter Mistral Large 4. Ricursive's dense Ripple-65B Preview is small enough to self-host. All scores are self-reported, and the licenses differ.

    Based on Ripple-65B Preview: a dense open-weight model aimed at agentic coding; Mistral releases open weights for Mistral Large 4, a 1T-parameter model

  2. The big labs competed on price and speed, not only raw power

    Anthropic's Claude Haiku 5.5 costs 90% less than Haiku 4.5 for prompts up to 100k tokens. OpenAI's GPT-6.1 Terra costs $1 in and $5 out per million tokens and targets computer use. Google brought Gemini 4 Argon to the Gemini app for paying subscribers.

    Based on Claude Haiku 5.5: Anthropic's cheapest small model, now with an effort setting; OpenAI ships GPT-6.1 Terra, a cheaper model for computer use and vision; Gemini 4 Argon arrives in the Gemini app for AI Pro and Ultra subscribers

  3. AI is now finding serious bugs in old code

    PHP2Shell turned an old PHP out-of-memory crash into unauthenticated remote code execution, found by Calif, OpenAI and Wiz with AI agents helping along the way. PHP shipped fixes on Oct 21. If you run PHP, upgrade. Expect more old code to get audited this way.

    Based on Ripple-65B Preview: a dense open-weight model aimed at agentic coding

  4. Agent configs are a new trust boundary

    A GitHub issue with 1478 HN points reports that Postman's desktop app installed Codex and Claude Code plugins and a startup hook without asking. Agent hooks and plugins can steer every session, so check what is installed and who put it there.

    Based on Postman reportedly auto-installs AI agent plugins for Codex and Claude Code

  5. Skills and multi-agent apps are where developers are building

    Anthropic gave the Agent Skills standard to the Linux Foundation's Agentic AI Foundation, next to MCP. Skill repos filled GitHub trending, and T3 Code, an app for running several coding agents side by side, shipped a large v0.1.7 release.

    Based on T3 Code v0.1.7 adds Kiro support, managed policies and shareable thread links

  6. Agent workloads strain infrastructure and raise trust questions

    Anthropic replaced git in its monorepo after merged changes per day passed 30,000. It also released FlashProve, an early prototype that uses cryptographic proofs to show which model produced a response. FlashProve is still far too slow and limited for production use.

    Based on Anthropic replaced git in its monorepo with an in-house system called Scion; Ripple-65B Preview: a dense open-weight model aimed at agentic coding

Last 30 days

Try it

Mistral releases open weights for Mistral Large 4, a 1T-parameter model

On Oct 29 Mistral published weights for an improved Mistral Large 4, after a public preview earlier in October. The 1T-parameter multimodal model got about three more weeks of RL on coding and defensive security.

Why it matters

Mistral reports 41% on Terminal-Bench 4.0, up 13 points from the preview, and 84.5% on BrowseComp. It is a large open-weight option from Europe, but companies selling it as a hosted service need a separate commercial license.

How to try

Download mistralai/Mistral-Large-4-1T-A52B from Hugging Face and serve it with vLLM or SGLang, or call it through the Mistral API or OpenRouter.

Talked about by

Try it

OpenAI ships GPT-6.1 Terra, a cheaper model for computer use and vision

OpenAI released GPT-6.1 Terra on Oct 26. The API model gpt-6.1-terra costs $1 input and $5 output per million tokens and has a 1,050,000-token context window.

Why it matters

Terra targets visual reasoning, precise computer use and writing at a mid-tier price. For agents that click through UIs or read screenshots, it is a cheaper option to test against larger models.

How to try

Call the OpenAI API with model "gpt-6.1-terra".

Talked about by

Try it

Ripple-65B Preview: a dense open-weight model aimed at agentic coding

Ricursive Intelligence released Ripple-65B Preview on Nov 4 under Apache 2.0. It is a dense 65B model with up to 1M context. The company reports 80.52 on TerminalBench 2.1.

Why it matters

A 65B dense model is small enough to self-host; the model card says it fits on an NVIDIA DGX Spark. The score is self-reported and the model is a preview, so test it on your own tasks first.

How to try

Download ricursiveAI/Ripple-65B-Preview from Hugging Face, or try it hosted on Together AI, GMI Cloud or OpenRouter.

Talked about by

Try it

T3 Code v0.1.7 adds Kiro support, managed policies and shareable thread links

T3 Code, the open-source app for running coding agents in parallel, shipped v0.1.7 on Nov 4 with 659 changes from 160 contributors. New: Kiro (beta), managed policies, one-time schedules and shareable thread links.

Why it matters

T3 Code was one of the fastest-growing repos on GitHub this week. It puts several agent harnesses in one UI, and recent builds work with Claude, OpenCode, Pi and Cursor, which helps if you switch between tools.

How to try

Download v0.1.7 for your platform from github.com/pingdotgg/t3code/releases.

Talked about by

Try it

Gemini 4 Argon arrives in the Gemini app for AI Pro and Ultra subscribers

On Oct 13 Google made Gemini 4 Argon available in the Gemini app for Google AI Pro and Ultra subscribers. Google says it reasons better, hedges less and gives clearer progress updates on long tasks.

Why it matters

Google's newest frontier model is now in a mainstream product, with a chat-to-tasks toggle that keeps tasks running while you are away and connections to more than 40 apps. Worth testing against Claude and GPT on your own work.

How to try

Open the Gemini app on a Google AI Pro or Ultra plan. Gemini 4 Argon is available there.

Talked about by

Try it

Claude Haiku 5.5: Anthropic's cheapest small model, now with an effort setting

Anthropic released Claude Haiku 5.5 on Oct 7. It is the first Haiku with adjustable effort. For prompts up to 100k tokens it costs $0.10 input and $0.50 output per million tokens, 90% below Haiku 4.5.

Why it matters

A cheap, capable small model changes the math for subagents, classification, summaries and compaction. You can run many more calls for the same budget and pair it with Opus 5.5 or Sonnet 5.5 as a worker model.

How to try

Call the Claude API with model "claude-haiku-5-5". It is also on AWS, Google Cloud and Microsoft Azure.

Talked about by