GPT-6 Astra Pro update: ChatGPT supports multi-agent parallelism
OpenAI says GPT-6 Astra Pro in ChatGPT now responds faster and handles plugins and tool calls better — and multiple agents can process different parts of a single task in parallel.
GPT-6 Astra Pro in ChatGPT has been upgraded with faster responses and improved use of plugins and tools. In the biggest change, multiple agents can now tackle different parts of a task at the same time, with users able to follow each agent's progress and open completed tasks to review their work. The update points to a broader shift inside ChatGPT from a single assistant toward orchestrated teams of sub-agents that divide labor, then report back through a shared, inspectable timeline.
The release also tightens tool and plugin handling, which OpenAI says reduces the round-trips a model needs before acting on a user's request. For developers and heavy users alike, the practical effect is the same: longer, multi-step jobs that once stalled on sequential reasoning can now fan out across several agents and finish in a fraction of the time.
Meta releases Muse Spark Open: an 816B open-weight model
With 816B total parameters and 42B activated per token, it scores 39 on the Artificial Analysis Intelligence Index.
Meta has released Muse Spark Open, an open-weights model widely described as the strongest US open-weight release to date. The dense-sparse design keeps only 42B parameters live per token while scaling to an 816B body, a configuration aimed at balancing frontier-level quality with the memory and throughput constraints of open deployment.
The launch lands in the middle of a crowded week for open weights, following releases from Tencent Hunyuan, Reflection, and Perplexity. Its 39 on the Artificial Analysis Intelligence Index puts clear distance between Meta and earlier domestic open models, and reinforces the argument that US labs are closing the open-weights gap that Chinese vendors once led.
"Build or surrender — for now, the choice is still ours."
Arthur Mensch · Mistral AI
Claude Mythos 5 designs quantum circuits that beat published bests
Anthropic's STEM Fellows Program reports that, given plain-language instructions, Claude Mythos 5 designed quantum circuits that improve on the best published ones — sometimes by an order of magnitude or more. The result is a striking example of language models doing original scientific design work rather than merely retrieving it.
Four Gemini 4 Argon agents beat Minecraft in 27 hours
Operating in a sandbox with only code and text, four Gemini 4 Argon agents wrote 2,783 programs, divided jobs over chat, shared notes, and debugged errors together — all without human help. The experiment doubles as a demonstration of agent teamwork under a shared goal.
Tencent Hunyuan open-sources Hy Memory 2.0: model, framework, and MemHub
Tencent Hunyuan's Hy Memory 2.0 is out as a single open package: a memory model, a framework, and MemHub, a CLI-plus-Mac client that syncs across more than ten AI tools. The company says the 30B-A3B model ranks first among open source on four benchmarks and beats Opus 5 across mainstream memory frameworks. Positioned as a private "memory steward," it targets developers who want persistent, cross-tool memory without handing their data to a hosted provider.
AMD MI355X approaches B300 on Kimi K3 serving
vLLM reports that 8-card MI355X delivers 24.4% more throughput than HGX B300 (7,567 vs 6,083 tok/s/GPU) on the AgentX agentic-coding benchmark, plus 87.4% more tokens per dollar.
Reflection open-sources Beam 1.0 weights
For coding, reasoning, and agent tasks, models and evals are live on Hugging Face: 501B A23B, DFlash2, NVFP4, and 255B versions, plus a 2B small model, under Apache 2.0.
Anthropic: Claude wrote about 90% of Claude Code's code
Merged PRs doubled from June to September while rollback share fell by over 40%; the team restructured code review to cut manual review time to the lowest level without increasing incidents.
Step 5 Preview rises to No. 2 on OpenRouter in a week
The model consumed 19.5T tokens in its first week and has held No. 2 on the daily chart since October 10.
Perplexity open-sources two SOTA decision models
pplx-decider-flash-v1.2 is 4B at 1 cent per million tokens; pplx-decider-v1.2 is 27B at 2 cents per million, ranking first in 4B and second overall on the HF Decision Index.
vLLM adds Day-0 support for Reflection-Beam-1.0
501B total parameters, 23B active per token, and a 1M-token context. Beam combines hybrid attention with Harmony-format reasoning and tool use, with Day-0 RL support from SkyRL.
Cursor cloud agents can now build interactive canvases
Ask for a dashboard or report while an agent works, then share the finished canvas with your team as a link. The feature moves Cursor's agents from code editing toward lightweight, shareable artifacts.
Gemini 4 Argon sets a WANDR record, lands on Perplexity
The model set a new record on the WANDR benchmark and is now available to Perplexity Pro and Max subscribers.
OpenAI engineers use Codex for the maintenance backlog
Internal teams use Codex to search the codebase, find what needs attention, and chip away at maintenance technical debt.
OpenAI SDKs simplify browser and computer use
The latest Python and Node.js SDKs make integrating browser-use and computer-operation capabilities far easier.
GPT-Live-1 supports prompt-customized voices
Shape a voice with a prompt or choose from newly added international preset voices.
Mistral Large 4 shines on STEMBench
The Ulam private benchmark scores answers 0–4 by how far they advance the original goal, letting partial solutions count.
PrismML releases ternary Bonsai 2 (2B/4B/9B)
NVIDIA deepens Mercedes-Benz self-driving pact
Runway launches Mirelo SFX 2.0 sound design
Replit launches BYOC across three clouds
Run development and production in your own Azure, Google Cloud, or AWS for stronger isolation and control.
Reflection Beam free on Ollama for now
A 1M-token context, software-engineering focus, US hosting, and zero data retention (ZDR) — free for a limited time.
Pika embeds creative tools in ChatGPT
The Pika X ChatGPT plugin puts all of Pika's creative tools directly in chat and the side panel.
First TPUv7 submission enters AgentX via vLLM
Google TPU, Red Hat, and inferact teamed up to get it running; more models are expected to land on TPU through AgentX.
SGLang adds Day-0 support for RTX Spark
An N1X chip with up to 128GB shared CPU-GPU memory brings frontier models closer to local deployment.
Claude Code gzips conversations for speed
Conversations are compressed before each turn so the model receives them faster; cloud sessions start as you type.
ASPIRE: an open agentic robotics framework V1.0
Researchers have open-sourced ASPIRE, an agentic robotics framework with real-robot skill discovery and self-improvement. On the day a new backend model arrived, the lab swapped it in and saturated benchmarks to roughly 99% overnight — a signal of how fast the agentic stack is compounding.
DeepMind models speed disease detection and drug discovery
Working with scientists worldwide, DeepMind's models and tools are helping pinpoint rare genetic variants and accelerate drug R&D, moving lab breakthroughs toward the clinic.
Cohere launches North Admin for token budget control
Control model use by the who, what, when, and where, and track and manage tokens in real time — aimed at teams that just tightened the token budget.
Luma integrates Claude Motion for continued animation
Animations made in Claude Motion can be sent to Luma with one click to keep building from there.
Replit integrates Runway for asset generation
Generate images, video, and assets directly inside projects for ads, launch videos, and landing pages.
Perplexity Computer adds Plaid as a personal CFO
Manage personal financial data and tasks inside Computer through the Plaid connector.
Perplexity adds reverse-engineering skill and credits
Up to 10,000 bootstrapping credits to reverse-engineer any app of your choice inside Computer.
Transformers offloads activation checkpoints to CPU
HuggingFace Transformers can now offload activation checkpoints to CPU, easing GPU memory pressure with spare CPU memory.
The RL flywheel behind Muse Spark Open
RL makes models better, LLMs optimize inference, and faster inference enables more RL — a recursive loop that keeps compounding.
Open Math OS: infrastructure for Math 2.0
A new open initiative argues the math community deserves its own AI system, built on open models and open infrastructure.
A mindset shift for how we deploy AI
As capabilities advance, safety, control, and skill erosion must move from risks to be managed into gating criteria.
Few-Step Generation via Data-Space Iteration
A shared generator refines predictions directly in data space instead of iterating along a fixed probability flow, beating standard discretization baselines on class-conditional ImageNet.
OmniCompass measures omni-modal video grounding
The benchmark spans 820 videos and about 1,800 hours of annotation across temporal grounding, spatial grounding, and QA. Across 50 multimodal LLMs, open models still lag closed ones, especially on audio-visual understanding.
Hybrid Reasoning Agents fuse latent and text reasoning
HybridAgents combine latent internal reasoning with text-based interaction, using a lightweight HybridLink to switch modes while keeping continuous working memory across turns.