Gemini 4 Argon Hits 97.5% on ARC-AGI-3 as Meta and Reflection Open-Weight Models Land

Key Takeaways

1. Gemini 4 Argon posts 97.5% on ARC-AGI-3 with a provider adapter harness, 49.8% with the standard harness, and rolls out to paid Google AI Pro subscribers in Antigravity.
2. Meta's Muse Spark Open, an 816B/42B open-weight multimodal model, arrives with day-one transformers and llama.cpp support and an NVFP4 checkpoint for Blackwell Ultra.
3. Reflection's Beam, a 501B MoE open-weight model, ships under Apache 2.0 with weights, tokenizer, reference implementation, and a training harness.
4. NVIDIA expands automotive partnerships with Volvo and Mercedes-Benz while its AI unit details training infrastructure behind the new open models.
5. OpenAI pushes GPT-Live-1.1 voice updates to all users and opens a sponsored academic plan for GPT-Rosalind.
6. Perplexity open-sources its SPACE sandbox platform under Apache 2.0 and adds Gemini 4 Argon and Nemotron 3.5 Super VL to its stack.
7. Anthropic commits $5M plus compute and Claude access to mathematics research, with Simons Foundation managing the funding.

1. Frontier Models and Benchmarks

1. Gemini 4 Argon scored 97.5% on ARC-AGI-3 at $4.8K cost using a provider adapter harness, versus 49.8% at $5.8K with the standard harness; it also reached 91.3% on ARC-AGI-2 at $0.32 per task and 98.5% on ARC-AGI-1 at $0.16 per task. Google DeepMind is the second frontier lab to exceed 95% on ARC-AGI-3. The model is rolling out in Antigravity to paid Google AI Pro subscribers, who get a separate weekly Argon quota alongside their Gemini 3.8 Flash quota, with Google citing heavy demand and deliberate capacity expansion. 1 2
2. A user reported working with Gemini 4 Argon for 11 hours and highlighted its 3D content generation, building and rendering a detailed Asuka-period Japanese pagoda in Blender via Cycles. 1
3. Meta's Muse Spark Open is described as the strongest US open-weight model to date, targeting coding and agent tasks with long-horizon work, complex software engineering, and multi-step reasoning. It has 816B total / 42B active parameters, a 1M token context window, and native text, image, video, and audio support. It was trained on NVIDIA GB300 NVL72, with an NVFP4 checkpoint optimized for Blackwell Ultra delivering over 50K tokens/sec/GPU. 1 2 3
4. Reflection's Beam is a 501B MoE open-weight model released under Apache 2.0 with weights, tokenizer, reference implementation, BF16/FP8/NVFP4 weights, a DFlash2 drafter, and the open-source Mirror CLI training harness. It was trained on NVIDIA GB300 NVL72 and reaches over 70K tokens/sec/GPU out of the box on Blackwell Ultra. 1 2
5. NVIDIA Nemotron 3.5 Super VL is an open model combining reasoning and tool calling with text, image, document, and video understanding for long-running agents, at 120B total / 12B active parameters. It is live on Perplexity for Pro and Max subscribers and the Agent API. 1 2
6. Flux.3 Image Klein shipped with full Diffusers support; the 5.6B model was trained on NVIDIA GB200 NVL72, runs on RTX GPUs with an NVFP4 variant, supports native 4K, bounding boxes, and pixel-precise editing. 1 2
7. Bonsai 2 launched in 2B, 4B, and 9B sizes as ternary models built on Qwen3.5, targeting local assistants with instruction following, loop avoidance, mobile tool calling, and concise output. Sizes are 0.43GB, 0.95GB, and 2.0GB; the 9B uses about 40% fewer tokens on standard instruction-following and everyday math. The family spans 2B to 27B under Apache 2.0. 1
8. NVIDIA Nemotron 3.5 Bolt 24GB is a 74B-A9B model that fits in 24GB of memory, optimized for RTX Spark and able to run alongside applications. A forwarded post says it supports 16 concurrent streams on a single DGX Spark with TensorFold. 1 2

2. Product Rollouts and Platform Moves

1. OpenAI's GPT-Live-1.1 updates voice in ChatGPT, Codex, and dots with French, German, Hindi, Mexican Spanish, Korean, and Japanese, plus improved accents and fewer "checking"-type pauses. It is rolling out to all users now, with API availability later. 1
2. OpenAI expanded academic access to GPT-Rosalind with the Rosalind Academic Plan, giving eligible US university labs sponsored access to use AI for understanding biology, exploring ideas, and deciding next tests. 1
3. Perplexity open-sourced SPACE, the sandbox platform powering all Perplexity Computer tasks, under Apache 2.0. Each agent task gets its own Firecracker microVM, with credentials and egress controlled outside the VM. The author argues sandboxes are commoditized but serving millions concurrently at production traffic requires substantial engineering, and that neoclouds need a full agent inference platform spanning open-weight models, fine-tuned models, and secure sandboxes. 1 2 3
4. Perplexity added Gemini 4 Argon, which it says set a new SoTA on the Wide-And-Deep-Research benchmark, for all users across fast search and Computer modes, and integrated a Reverse-Engineer-Anything skill into Computer that analyzes apps without source code and builds working versions, free through October 31 up to 10,000 credits per user. 1 2
5. Anthropic committed $5M in funding, compute, and Claude access to help the mathematics community navigate AI-driven change, with the Simons Foundation managing the grant through workshops, conferences, and expert lectures. A cited post says the Simons Foundation is putting $15M into workshops and initiatives with Anthropic and OpenAI and providing free AI compute to its supported researchers. 1
6. Anthropic said it has committed to third-party review of its safety practices, is in talks with multiple organizations, and has given METR Evals staff-like access to explain its approach to external evaluators. 1

3. Infrastructure, Automotive, and Industry Signals

1. NVIDIA expanded its automotive partnerships: with Volvo, combining Volvo safety expertise with NVIDIA Hyperion, Halos OS, and the Alpamayo Stack for safe autonomous driving, first appearing in next-generation Volvo models; and with Mercedes-Benz, jointly developing next-generation driver assistance and autonomous driving systems while maintaining Mercedes' safety standards. 1 2
2. Jensen Huang appeared with Wallenberg Investments' Marcus Wallenberg, Spotify's Daniel Ek, and Sferical AI's Jenny Nordlöw at the launch of Sferical One, a shared AI supercomputer powered by NVIDIA for researchers and enterprises. NVIDIA also highlighted Nordic developers applying its technology to manufacturing and healthcare challenges, with Berlin as the next stop. 1 2
3. SpaceX and NVIDIA's jointly designed Star Mind computer is claimed to perform far beyond expectations and enter mass production next year; Musk said working with NVIDIA's engineering team was "pure joy," and Jensen Huang said NVIDIA has moved Musk's AI business from the enterprise category to cloud service provider, calling its scale "huge," with both engineering teams designing new server architectures for space and Earth. 1 2 3 4
4. Tesla's Roadster was unveiled with 0-60 mph in 0.81 seconds using thrusters and suction-enhanced grip, a top speed of Mach 0.5, and flight capability; the new high-voltage battery and drive unit deliver one-third more power than Model S Plaid at half the mass, outputting 1MW. Musk said it "flies like a rocket" with attitude-control and rear main thrusters, can tumble, and self-inflates air tanks without special propellant, adding it is not safety-optimized and may not be legal in all jurisdictions. Track Mode aero provides the best passive downforce of any road car, near F1 levels, with active surfaces driven by Tesla Optimus actuators. 1 2 3 4 5
5. Tesla's supervised self-driving was approved in Slovakia and will begin rollout soon, while Cybercab Robotaxi reported zero collisions or incidents in its first two weeks of public passenger service based on the latest NHTSA data. 1 2
6. Starlink in aviation expanded: United said over 5.4 million passengers, about a third of its customers, used Starlink on flights last month, with equipped aircraft scoring over 40 points higher in satisfaction than other Wi-Fi services. Southwest will extend Starlink to its entire fleet by fall 2027, expected to be the first major US carrier with full-fleet deployment. A seven-year Delta top-tier loyalty member wrote that after experiencing United's Starlink, which felt like home internet, they decided to switch, citing Delta's choice of Amazon Leo in 2028 as insufficiently reliable. 1 2 3
7. X's advertising and creator programs: Musk announced all ads on X will be subject to "reasonable aesthetic standards," with advertisers running ugly, jarring, or exaggerated product promotions banned. X is piloting an invite-only Video Kickstart Bonus requiring vertical, under-3-minute original videos, allowing watermark-free cross-platform content, with at least 15 videos on 15 different days within 30 days and applications closing October 30. 1 2 3
8. Grok Bot saw Musk claim unprecedented growth with no free tier, citing Huang's remark that usage is "surging." Users described it monitoring calendars to precondition a Tesla before landing and summon the car, and Musk said it improves significantly daily. The Grok Bot team released 10 official templates covering design bots, code review, competitor monitoring, SaaS spend negotiation, prototype generation, video editing, Figma-to-spec, marketing distribution, product QA, and CRM buying-signal mining. 1 2 3 4 5
9. Interpretability debate: Sam Altman welcomed Nick Cammarata's view that interpretability may be the most important scientific problem today, since alignment cannot be solved without understanding built systems. Cammarata said early mechanistic interpretability at OpenAI was pessimistic, with only about 2% of neurons readable using primitive tools, but modern agent tools make the vast majority of neurons in trillion-parameter models readable with elegant circuits, and many famous "entangled neuron" cases were misunderstandings. He now expects ambitious interpretability to be solved when most needed, possibly soon, as automated scientists improve, calling it a possible "bitter lesson" where simple methods work at absurd scale. Deedy cited this as one of AI's most important results, noting 90% of neurons in Qwen-32B serve a single function. 1 2
10. Ethan Mollick argued that harnessed intelligent models are ending multiple well-known hard benchmarks, citing Gemini 4 Argon's ARC-AGI results and saying models can work for long stretches, produce mathematically correct professional documents, and solve complex novel puzzles. He also cited a paper on why frontier models cannot be cheaply replaced, advocating the strongest model as orchestrator because small gains in accuracy and judgment compound over long-horizon tasks. 1 2
11. John Carmack argued performance improvements should be measured against the "speed of light" rather than bad competitors, suggesting UI should display locally cached media within a few vblanks of process start, servers should fire a single query in parallel with Gfx API initialization to return everything needed for dynamic UI, and large media should use modern transport protocols, targeting a response before muscle-memory click after double-clicking a launcher. 1
12. Ben Tossell described being "non-technical" and terminal-averse yet shipping code to production at FactoryAI, storing agent context in folders, not using Claude Code, and finding AI makes him want to spend more time at the computer. His episode covered learning AI collaboration through boring small projects, a mascot animator for his personal site, a visual dictionary for coding agent terms, splitting tasks among Pi, Factory, and Codex, preferring faster models and steering agents mid-work, using a shared glossary to avoid humans and agents talking past each other, and questioning whether calling AI adoption a "skills problem" lets AI labs off the hook. He also raised whether personal agents converge on a single thread or multiple agents. 1 2
13. AI Engineer New York attendees summarized the industry as maturing, with engineers from Monzo to Anthropic holding clear opinions on what works and systems no longer experimental. The "Agent Factory" was the main topic, seen as the most innovative theme at roughly 5% adoption in the technical community, and Evan Conrad's "How compute markers work" was called the most interesting talk, covering compute backlogs and GPU financing. 1
14. Mistral Large 4 scored well on Ulam's private STEMBench benchmark in engineering and astronomy projects. 1
15. Google Quantum AI demonstrated "magic state cultivation" on physical hardware for the first time, achieving a 40x error rate reduction on the Willow processor versus prior methods and a logical fidelity of 99.99%, published in Science. Google DeepMind also held its first Bench Summit on advancing benchmarks and model evaluation. 1 2