Google's Argon lands (and your first Signal)

Hello, and welcome to the very first Signal Tech! 👋

You're one of our founding readers, and we're chuffed to have you. Here's the deal: every week we pick out the tech stories that actually matter in AI, chips and dev tools, explain them in plain English, and skip the hype. Five minutes, tops. Let's get into it.

📡 The Big Signal: Google finally ships Gemini 4 Argon

On Tuesday (13 October), Google released Gemini 4 Argon, its new flagship AI model. It's available now to developers and businesses through the Gemini API and AI Studio, and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. Google says Argon is built for "deep work": long, complex jobs in coding, science, cybersecurity and agent-style tasks, with a 1-million-token limit via the API.

Why it matters: this is Google's first major frontier release since Gemini 3 in late 2025, after months of watching OpenAI and Anthropic pull ahead. Reuters notes Google had already scrapped plans for Gemini 3.5 Pro. Pricing is the bit to watch if you build with it: $2 per million input tokens and $10 per million output tokens at launch, rising to $4 and $20 after 30 November. Try it now while it's cheap, and do your sums before you commit.

Sources: Google · The Verge · Reuters via Yahoo

⚡ Quick Hits

AI chip designers can write the code, not finish the job. ChipAgents and NVIDIA launched Build-A-Chip-Bench, which tests whether AI agents can take a chip from spec all the way to a finished physical layout. Tested agents passed functional checks on about 40% of tasks, but only around 10% hit their performance, power and area targets. ChipAgents

NVIDIA wants AI running the fab, too. At SEMICON West on 14 October, NVIDIA announced cuPotential, an early-access library for simulating how atoms and molecules interact, which it says sped up already GPU-optimised workflows by a further 2-3x in testing. It also said the open-source NanoVDB library helped Lam Research run chipmaking simulations 30x faster. NVIDIA

A 24B model built to run on your laptop. Liquid AI released LFM2.5-24B-A2B, an open-weight model designed for on-device agents, with a 131,072-token context window and downloads on Hugging Face. Its demo showed the model calling tools on a MacBook Pro without any cloud API calls. AI.info

Meta's crowded week. Argon landed just a day after Meta announced its own frontier model, Muse Wonder Full. The takeaway: the AI race is very much not a two-horse affair any more. The Verge

🛠️ Tool of the Week: ML Drift

Google has open-sourced ML Drift (Apache 2.0), the GPU engine behind its LiteRT on-device AI runtime. It hides the messy differences between OpenGL ES, OpenCL, Metal and WebGPU, so one model can run fast on Android, iOS, the web and desktop. If you still use the old TFLite GPU delegate, note that it won't get new features any more. Google says ML Drift is fully backwards compatible, so switching should be painless.
👉 Read the announcement

💡 One Takeaway

Watch the "introductory" price tags. Argon's API price doubles after 30 November. Before you build anything on a new model, put a rough cost estimate in your README (tokens per request × requests per day × price), then work it out again at the full price. If the numbers only add up during the discount, you want to know now, not at invoice time.

That's issue #1! What did you think? Just hit reply: we read every message, and we'd love to know what you want more (or less) of.

And if Signal Tech saved you some scrolling this week, forward it to a friend who'd enjoy it. That's how we grow.

Until next Friday,
The Signal Tech team 📡
The tech that matters, minus the noise.