The AI Wire ●

In a recent interview, Mark Zuckerberg pitched Muse — Meta's new AI assistant — as "built from the ground up for privacy and security," explicitly borrowing the WhatsApp playbook. He said Meta's 10+ years building WhatsApp taught them that systems designed so even Meta can't see the content are what earn user trust.

The technical core of the pitch: Zuckerberg said he and Nat Friedman personally recruited Moxie Marlinspike — who helped build WhatsApp's end-to-end encryption back in 2014 — to lead a "confidential VM" project. The claim is architectural, not just a privacy policy: your agent's VM runs in a system Meta itself cannot inspect, and Zuckerberg says that commitment is technically verifiable, with more detail promised "in the coming weeks."

For the scariest part — handing an agent your logins — Meta described "Sentinel" agents that monitor what your agent sends and flag anything needing your approval, plus a secure credential store that issues one-time card numbers instead of raw credentials. Sizing it up: the privacy story is the product story for Muse, and Zuckerberg is betting verifiable architecture where competitors offer policies.

On September 13, Elon Musk announced on X that xAI's next model will be Grok 4.8: 2.5 trillion parameters on an in-house C++ software stack rebuilt from scratch, moving straight into reinforcement learning once the main training run completes. The announcement effectively supersedes Grok 4.7 (2.1T), which missed multiple release windows and was publicly downgraded by Musk to 'roughly comparable to Anthropic's Opus 5.0'.

No model card, benchmark table, pricing, API identifier, or release date has been published; every figure attached to Grok 4.8 is a founder claim, not a shipped specification.

Anderson Cooper asked Amodei point-blank whether AI could kill all humans. He didn't deny it — saying he agrees with resigned researcher Jacob Coxon 'much more than I disagree' — refused to give a probability, and warned rogue agent swarms could seize control of the internet within 6–12 months. The interview accompanied his 'Pace the Frontier' essay urging labs to slow capability gains; Altman and Musk publicly backed parts of it.

In an interview with Fortune editor-in-chief Alyson Shontell, Altman ruled out a 2026 IPO ('We've got a lot of stuff to do'), tying the delay to AI safety after the OpenAI–Hugging Face agent incident. He backed Dario Amodei's call to slow frontier development and hinted a cross-industry safety pact 'will happen.'

Sakana AI released Fugu Max v1.0 and Fugu Ultra v2.0 on September 11 — not monolithic models but learned orchestration engines: a language model trained to route a task across a pool of open-weight and specialized models (including NVIDIA Nemotron) and recursively call instances of itself, via an OpenAI-compatible endpoint. Fugu Max is the cost tier at $2 per million input and $6 per million output tokens; Fugu Ultra v2 is the quality tier at $5/$30 ($10/$45 above 272K context), with a 1M-token context window.

Ultra v2 ships without any proprietary frontier model in its pool — its 48.3 on Chartography reportedly beats Opus 5's 27.3 without using Opus. API-only; no downloadable weights; not available in the EU/EEA. (Filed September 21 as backfill — a significant release the sweep missed on September 11.)

DeepSeek launched V4.1 Flash, a 552B mixture-of-experts model (only ~8B parameters active while reading, ~16B while writing) with aggressive KV-cache compression that pushes API prices far below Western frontier models. DeepSeek is routing flagship V4 Pro traffic to it. Independent coding benchmarks put it roughly on par with GPT-5.6 Sol — Matthew Berman covered the launch the same week.

OpenAI temporarily paused new sign-ups for its $200-per-month ChatGPT Pro tier on September 10, saying demand for GPT-6 Astra was putting unprecedented strain on its infrastructure. Product leader Thibault (Tibo) Sottiaux announced the move on X, calling the Pro tier the heaviest load on the company's systems and framing the pause as 'the smallest step that allows us to continue giving the broadest access possible.'

Plus, Go and API access remain open. A day earlier he had warned that demand was 'really unprecedented' and that a pause might be needed to protect service quality for existing subscribers. OpenAI gave no reopen date. (Filed September 21 as backfill.)

OpenAI made GPT-Live-1 generally available in the API on September 10, bringing the full-duplex voice conversations from ChatGPT to developers. The model listens and speaks simultaneously — no turn-taking relay — and delegates deeper reasoning and tool calls to a backend text model of the developer's choice (GPT-6 Astra, Terra/Luna, or third-party).

OpenAI's published numbers: 80.1% on Full Duplex Bench v1.5 interactivity and 0.798s turn-taking latency versus 1.41s for GPT-Realtime-2.1; 83.6% on Tau3 customer-service tasks when paired with GPT-6 Astra. Pricing is $0.05 per minute of voice, billed per second, with backend tokens billed separately. Connects via WebRTC, WebSockets, or SIP/telephony.

NASA and IBM announced the NASA-IBM Lunar Foundation Model on September 10, an open-source AI system that turns more than a decade of lunar observations into detailed maps. Trained on over 30 data layers from nine instruments aboard four NASA missions including the Lunar Reconnaissance Orbiter, it fuses imagery, topographic, thermal and spectroscopic measurements to detect crater rims, volcanic vents and potential ice deposits in permanently shadowed craters — boosting feature-detection accuracy up to 23% over conventional methods.

IBM describes it as one of the first publicly available foundation models purpose-built for lunar science, following the open Prithvi geospatial models for Earth observation.

OpenAI's head of Core Products & Platform — the face behind the usage-reset announcements — talks through how Codex was built and iterated: why the CLI is Rust and open source, how the harness and models evolved, why Codex supports multiple providers, and the ChatGPT + Codex merger.

OpenAI's new flagship model, succeeding GPT-5.6 Sol. State-of-the-art on computer use, browsing, software engineering and science per OpenAI; rolls out to ChatGPT Plus, Pro, Business and Enterprise tiers in waves, plus API access.

Google launched Gemini 3.8 Flash, the latest iteration of its monthly Flash cadence, alongside Gemini 3.8 Flash Cyber — a security-focused variant restricted to vetted defenders under Google's Fairwind program, which spans more than 650 organizations.

The Cyber variant is tuned for defensive security work, reportedly patching vulnerabilities far faster than standard models, with access gated to vetted participants. It mirrors the industry's split between general and security-specialized model access.

As of mid-September 2026, Gemini 3.8 Flash is Google's newest release — and the Pro flagship, once expected in July, still hasn't shipped.

Same underlying model as Fable 5.1 but with reduced safeguards, available only through trusted-access programs for cybersecurity and life-sciences work. Not generally available.

Anthropic released Claude Fable 5.1, a refined version of the premium Fable 5 tier, alongside Claude Mythos 5.1 for vetted institutions. The update addressed reliability complaints from the original June launch, which had been interrupted by export-control halts.

Fable 5.1 kept the premium $10/$50 per-million-token pricing, signaling Anthropic's confidence that a luxury tier had a real market. Early reports focused on steadier long-session behavior.

As of mid-September 2026, Fable 5.1 is the newest shipped Claude release; Fable 5.2 remains rumor only.

Anthropic's most capable generally available model for coding and knowledge work: 1M token context, always-on adaptive thinking, and ~25% lower typical cost than Fable 5 via cheaper cache reads. Now the default in Claude Code.