TabLOID

Google announced Gemini 4 Argon on September 30. Its first users are trusted cyber defenders in the Fairwind Program; developers, enterprises and consumers are due to follow. There is no broad-access date in the announcement.

The details:

• Introductory API prices are $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper.
• Google says the maximum output is now 1 million tokens, up from 64,000.
• Its internal examples include memory savings of more than 300 TiB and C/C++ to Rust migrations. These are Google's reported results, not independent findings.

Why it matters: a launch can be real without being something you can use yet. Before building around it, check the actual access and pricing.
Google's announcement

Quartz reports Meta claimed $3.9 billion in 2025 research tax credits by classifying AI facilities as experimental. The position faces potential IRS challenges.
Read Quartz

OpenAI describes a coordinated effort to extract protected reasoning through manipulated model interactions. It says operators did not break encryption or directly access stored conversations.

Its July 24–25 spike involved 16,000 attempted extraction requests from more than 4,000 users. Related patterns extended to over 15,000 users, and OpenAI says it disrupted the activity by July 28.

The attribution needs precision: OpenAI's own post names individuals associated with Moonshot AI as a core cluster. It does not claim every operator belonged to one actor.

Why it matters: model outputs and tool results need controls too. A locked database is not the whole security plan.
OpenAI's findings

Anthropic's new mods are TypeScript functions that alter Claude Code's behaviour and interface. They ship inside plugins and work in its CLI and desktop app.

A mod can rewrite prompts, change tool calls, redact secrets or add interface controls. Mods can run around events and stack in load order. The built-in diff feature now ships as one.

The catch: these functions run with Claude Code's machine access. They are not sandboxed. Treat a new mod like software you are installing, because that is exactly what it is.

Why it matters: small, useful team tools could live inside the agent people already use. Start with one dull problem you can reliably solve.
Anthropic's announcement