Cheaper frontier access, open weights on Macs, and runtime AI controls

This Week in One Line

Google opened Gemini 4 Argon globally at $2/$10 token pricing, Meta published open weights for Muse Spark Open on Macs, NVIDIA pushed local‑AI PCs and new agent models, and Microsoft + NVIDIA issued an open‑weight risk guide — net effect: cheaper, safer, more local AI at work.

Week in Numbers

• $2 per 1M input tokens — Gemini 4 Argon’s introductory API price lowers trial costs for complex work. ¹
• 1M tokens — Context window in Meta’s Muse Spark Open, enabling very long, multimodal tasks. ²
• 5x — NVIDIA says Nemotron 3.5 Super VL delivers up to 5× higher single‑GPU throughput for agent tasks vs. open peers in its class. ³
• 40% — Microsoft’s Maia 300 targets 40% more tokens per megawatt vs. leading GPU systems, signaling efficiency focus. ⁴
• $0.75 per 1,000 calls — DeepSeek opens Search/Fetch APIs at metered rates for web retrieval pilots. ⁵
• 501B parameters — Reflection’s Beam releases open weights at frontier scale with a 1M‑token context. ⁶
• 24% — Share of banks ServiceNow cites as having governance in place despite 100% using AI, motivating runtime controls. ⁷

Top Stories

Google’s Gemini 4 Argon opens globally with aggressive launch pricing

Google made its newest flagship model broadly available through its developer tools and application programming interface (API), and in the Gemini app and AI Mode in Search. Introductory API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens, with a 95% discount on cached inputs — a direct cost lever for long, context‑heavy work. ¹

Google reports strong internal developer productivity signals and early enterprise testers across legal, healthcare, and R&D. External coverage notes rates are set to double after Nov 30, creating a time‑boxed window to run comparative pilots without a full platform switch. ⁸

Meta releases Muse Spark Open weights for desktop‑class multimodal agents

Meta published open weights for Muse Spark Open — a multimodal Mixture of Experts (MoE) system that reads text, images, audio, and video and generates text — with a 1M‑token context window and about 42B active parameters per token. The release includes runnable recipes and notes a bfloat16 (BF16) checkpoint around 1.5 TB, with guidance for popular inference engines. ²

Meta highlights local execution tuned for Apple’s MLX framework on Mac Studio (M5 Ultra), showing end‑to‑end on‑device agent workflows for coding and media tasks. Hosted access is available via multiple providers, and the Muse Model License allows limited commercial use rather than open‑source terms — relevant for teams planning distribution. ⁹

Microsoft details Maia 300 and plans multi‑gigawatt Azure Maia expansion

Microsoft introduced Maia 300 as its next in‑house accelerator focused on long‑context reasoning, latency‑sensitive conversations, and sustained multi‑step agent tasks — positioning Maia as its primary inference platform. The company also disclosed rack‑scale design details and efficiency goals that point to steadier capacity for everyday cloud AI features. ¹⁰

Specs and targets include eight HBM4 stacks, a 950‑watt TDP, and claims of 40% more tokens per megawatt versus leading GPU setups and up to 7× Maia 200’s rack‑level token throughput (96 accelerators). For users of Microsoft 365, Teams, and Copilot, the emphasis on tokens‑per‑watt implies more responsive assistants as backends scale. ⁴

NVIDIA ships Nemotron 3.5 Super VL for agents; Bolt variants target edge

NVIDIA announced Nemotron 3.5 Super VL, an open 120B‑parameter model that activates 12B parameters per token, combining reasoning, tool use, and vision (OCR, tables, charts, video) for specialized agents. The company claims top task completion and up to 5× higher single‑GPU throughput than other open models in its class. ¹¹

A technical deep‑dive outlines the MoE design and packaging via NVIDIA NIM for production use, with open datasets and recipes to customize. For document‑heavy workflows in finance, legal, or ops, the text‑plus‑vision design aims to reduce tool handoffs and improve throughput. ³

Reflection releases Beam: a 501B open‑weight model with 1M‑token context

Reflection published Beam under Apache 2.0 as a sparse MoE Large Language Model (LLM) built for coding and agentic tasks, totaling 501B parameters with ~23B active per token and a 1M‑token context window. Checkpoints ship in BF16, FP8, and NVFP4 with day‑0 support in popular inference stacks and a Rust harness. ⁶

Independent analysis places Beam among the stronger U.S. open‑weight options on agent workflows while flagging weaknesses on Terminal‑Bench and a higher hallucination rate — a cue to pair long‑context power with strong evaluation and guardrails in production. ¹²

UK grants its AI Safety Institute early access to test frontier models

The UK’s AI Safety Institute (AISI) secured pre‑deployment access to test the newest models under voluntary agreements with major developers, including Google’s Gemini 4 Argon. The government frames the program as a way to surface risks earlier, in coordination with U.S. counterparts. ¹³

For teams operating in or selling to the UK, this implies tighter pre‑launch safety documentation and alignments — expect procurement and legal to ask for concrete safeguards and evaluation evidence ahead of deployment. ¹³

ServiceNow debuts AI Passport for runtime agent authorization

ServiceNow introduced AI Passport, a cryptographically signed credential that agents check at runtime to verify what actions they are authorized to take, enforced via NVIDIA’s Open Agent Safety Platform (OpenShell, Sentry). It adds graduated responses up to a kill switch and full audit trails under its Super Intelligence Control Tower. ⁷

Built on open specifications, the push is from after‑the‑fact audits to prevention. ServiceNow highlights a governance gap — 100% of banks report using AI while only 24% report governance in place — positioning runtime credentials and unified logs as baseline controls. ⁷

DeepSeek opens Search/Fetch APIs and updates its model

DeepSeek released an updated model (DeepSeek‑V4.1‑Flash‑1017) and opened its Search and Fetch APIs to all users at $0.75 per 1,000 calls across web, mobile, and API access. The company cites faster task completion, shorter outputs, and reported gains on DeepSWE and Automation‑Bench, alongside optimizations for its harness and partner tools. ⁵

For growth and research teams, the endpoints provide a budgeted path to add web retrieval to prototypes without switching your primary model — useful for testing Retrieval‑Augmented Generation (RAG) alternatives piece by piece. ⁵

NVIDIA primes RTX Spark PCs for local‑AI agents

Ahead of new laptop availability, NVIDIA and partners rolled out software updates to make local agents easier: sandboxed agent apps, PAIR to pool multiple PCs, Broadcast 2.3 upgrades, and creative app optimizations. The focus is “plug‑and‑play” local AI for coding, research, and media. ¹⁴

For privacy‑sensitive or latency‑critical work, the stack reduces cloud dependence and setup overhead — like running a mini AI studio on your desk. The ecosystem framing also signals a path for teams to mix local inference with cloud services pragmatically. ¹⁴

Microsoft and NVIDIA publish a shared playbook for open‑weight model risk

A joint paper from Microsoft and NVIDIA offers enterprises a five‑category taxonomy — Cybersecurity, Data Protection, Content, Human‑AI Interaction, and Bias — to assess and mitigate risks when deploying open‑weight AI models. The guidance maps risks to evaluations and clarifies how control ownership shifts with deployment choices (managed vs. self‑hosted). ¹⁵

For buyers, this supplies a common language across engineering, security, legal, and product, and a reminder to pair model‑layer controls with runtime, harness, and application‑layer testing and monitoring. Future installments and a shared repo are planned to expand the library of evaluations. ¹⁵

Trend Analysis

Cheaper frontier access met practical guardrails. Google’s Gemini 4 Argon arrived globally at $2 per 1M input and $10 per 1M output tokens with a 95% cached‑input discount, making it straightforward to A/B stronger reasoning on real work without full migration — a cost window external coverage says narrows after Nov 30. That pricing move coincided with internal speedups and enterprise pilots, pointing to near‑term, measurable trials over broad platform bets. ^{1 8}

At the same time, open‑weight options pushed into frontier territory and onto desks. Meta’s Muse Spark Open brought a 1M‑token multimodal agent to Macs with MLX optimizations, NVIDIA added Nemotron 3.5 Super VL for agentic text‑plus‑vision tasks, and Reflection’s 501B‑parameter Beam offered a long‑context, Apache‑licensed path to customization. NVIDIA’s RTX Spark updates stitched local hardware, sandboxes, and apps into a coherent “personal AI workstation,” signaling a viable local‑plus‑cloud mix for privacy or latency. ^{2 11 6 14}

Governance moved from documents to runtime. The UK’s AI Safety Institute secured pre‑launch model access, while Microsoft and NVIDIA published a shared risk taxonomy for open weights, and ServiceNow’s AI Passport introduced signed, revocable credentials for agent actions with full audit trails. Together these steps imply earlier safety expectations and live‑system controls becoming table stakes for enterprise AI — not just post‑hoc reports. ^{13 15 7}

Capacity and efficiency formed the backdrop. Microsoft’s Maia 300 emphasized tokens‑per‑watt gains and long‑context throughput, complementing local compute pushes and hinting at steadier user‑facing responsiveness. DeepSeek’s metered Search/Fetch filled a pragmatic gap: add web retrieval where it pays off, instead of wholesale stack swaps. The week’s arc: do more per dollar, keep more on trusted machines, and prove it with evaluable controls. ^{4 5}

Watch Points

• “Argon pricing window” — External reports say launch rates double after Nov 30; if you see “pricing change,” plan bake‑offs now. ⁸
• “Runtime passports” — If vendors tout credentials or kill switches, it’s about signed, enforceable agent permissions with audit trails. ⁷
• “1M‑context memory budget” — Long‑context open weights are here; watch KV cache growth and hardware needs before piloting. ²

Open Source Spotlight

• LiteLLM — One gateway to 100+ model APIs with cost tracking, guardrails, load balancing, and logging; good for teams standardizing the OpenAI‑style interface across vendors. BerriAI/litellm
• Mastra — TypeScript framework for agents; new Agent Scopes route state safely in multi‑tenant apps. Useful for SaaS builders adding agent features. mastra-ai/mastra
• Ollama — Local runner for popular open models on macOS/Windows/Linux/Docker; recent prerelease reduces per‑request overhead and updates llama.cpp. Great for private, fast iteration. ollama/ollama
• OpenHands — Self‑hosted “developer control center” to orchestrate coding agents across clouds and local backends; centralizes runs and logs. OpenHands/OpenHands
• LEASH (paper) — Training‑free routing that escalates from a “weak” to a stronger model only when uncertainty demands, cutting costs in computer‑use agents. arxiv:2610.22280

What Can I Try?

1. Pilot Gemini 4 Argon in AI Studio: run a real 10–20 page review or code refactor, capture quality, latency, and token spend at $2 per 1M input / $10 per 1M output, and design prompts to reuse cached inputs (95% discount). ¹
2. Draft an “AI Passport” scope card: list one agent’s allowed data, tools, actions, approvers, and a kill‑switch owner to mirror ServiceNow’s runtime‑authorization model before any production use. ⁷
3. Add retrieval with DeepSeek Search/Fetch: run 100–500 calls at $0.75 per 1,000 to compare web‑assisted answers against your current workflow, then decide where retrieval actually pays off. ⁵
4. Spin up a local model via Ollama: install the prerelease and test two open‑weight models on a private document to gauge latency, privacy, and fit for desk‑side agents. ollama/ollama