AI news this week: Gemini 4 Argon, Haiku 5.5 and Muse Spark Open

AI news this week: Google ships Gemini 4 Argon, Anthropic opens up Haiku 5.5 and fast mode with no waitlist, Meta releases the Muse Spark weights, and Krea and Liquid AI put out new models you can use right now. On top of that, a bugging scandal at OpenAI and $15 million for mathematicians. Below is everything that topped the charts in the AI world between October 10 and 16, 2026.

AI news this week: models

[October 15] Muse Spark Open. Meta released the weights of its 816B model. I covered what the license allows and what hardware you need in a separate post. Details on Meta’s blog.

[October 15] Claude fast mode for Opus 5.5, no waitlist. The mode that generates responses up to 2.5 times faster is out of research preview, according to Anthropic’s Release Notes. Same model, faster, higher price per token.

[October 15] Krea 3. An image editing model: most edits at 2K in under 8 seconds via the API, $0.06 per image, up to six reference images. It’s in beta, including on the free plan. More in my post on Krea 3 and on Krea’s site.

[October 13] Gemini 4 Argon. The biggest launch of the week and the top-voted post on Hacker News this week. Google describes Argon as a model “for complex workflows” in software engineering, office work and cyber defense. The model page lists plenty of results, including 91.9% on Vibe Code Bench and 77.9% on DeepSWE v1.1, but it doesn’t lead everywhere: on Terminal-Bench 4.0 it scores 57.4% versus 66.4% for Claude Opus 5.5, and on FrontierSWE v2 55.0% versus 65.5% for GPT-6 Astra. The model is available in the Gemini app for Google AI Pro and Ultra subscribers, and for developers in the Gemini API and AI Studio on the paid tier. Google is rolling out the cyber defense features gradually to trusted partners through the Fairwind program. In my own tests, though, it unfortunately doesn’t stand out at all. :(

[October 13] LFM2.5-24B-A2B from Liquid AI. Open weights, 24B parameters with about 2B active per token, 131K-token context. The model is built for agents running locally: according to Liquid AI, it decodes 184 tokens per second on an Apple M5 Max. The weights are on Hugging Face.

[October 10] Claude Haiku 5.5. Anthropic calls it the “cheapest, fastest and most capable small model” it has released. Pricing: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, which Anthropic says is about 75% cheaper than Haiku 4.5. It’s the first Haiku with effort control across five levels. The jump shows up in agents: 72.4% on OSWorld 2.1 versus 15.7% for its predecessor. Details on the model page.

Tools

[October 14] OpenCode 2.0. A major rebuild of the open-source coding agent for the terminal: a shared background process for the terminal, web and desktop clients, a rewritten API, “Code Mode”, where the agent chains tool calls with a small JavaScript script, background subagents, and more than 80 internal plugins. To upgrade: opencode upgrade --major. Heads up: plugins from 1.0 don’t work in 2.0. Details in the Release Notes.

If you prefer Claude Code, though, this week I published a from-scratch tutorial. ;)

Curiosities from the labs

Anthropic ditched git. In an engineering post, the company describes how, with daily merges up more than tenfold, version control became its biggest bottleneck, and how it moved its monorepo to its own system, Scion, in five months. Scion isn’t released or open, so Anthropic only shares the design principles. There’s a twist, though: the author of the sunshowers.io blog writes that the project is heavily inspired by her system Mononoke, which hit the HN front page twice.

“Neural networks are clean”. An essay by Nick Cammarata of Goodfire hypothesizes that the famous examples of polysemanticity and superposition were misread, and that most neurons have a single job. The author cites preliminary audits of Qwen3-32B in which about 90% of neuron “poles” have one main function, and notes himself that the numbers depend on thresholds and small samples. Recommended for interpretability research fans.

Events

Bugging at OpenAI. On October 13, prosecutors announced charges against an OpenAI engineer who, according to investigators, planted a recording device in the executive conference room. According to the SF Standard, the company says employees used its own models to identify the device, then handed the case to the FBI, and that there are no signs of leaked weights, code or customer data. The story is all the funnier because the would-be “secret agent” reportedly used very amateur methods (like asking his wife’s ChatGPT how to plant a bug). :)

$15 million for mathematicians. On October 15, the Simons Foundation, Anthropic and OpenAI announced a joint program to help researchers evaluate results obtained with AI and develop standards for its use. The labs are contributing tools and compute.

A deadline for developers on older OpenAI models. On January 14, 2027, GPT-4o, GPT-4o mini and GPT-4.1 disappear from the API, and as early as October 23, 2026, the older models from the previous deprecation wave go. I wrote up what to check and what to migrate to here.

Regulation. The AI Office sent OpenAI and Anthropic requests for information after incidents involving agents. More in my post on AI Act enforcement.

FAQ

Where can I try Gemini 4 Argon?

In the Gemini app if you have a Google AI Pro or Ultra subscription, and in Google AI Studio and the Gemini API on the paid tier. Google hasn’t announced free access yet.

How much does Claude Haiku 5.5 cost?

$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens; above that threshold, $0.50 and $2.50 respectively.

Which of these models can I run locally?

LFM2.5-24B-A2B has open weights and is built for consumer hardware. Muse Spark Open also has open weights, but at 816B parameters it’s server-class.