Meta calls them its most capable models. Muse Wonder is on the Meta Model API and in Muse Code; Muse Wonder Full comes only with a Muse Code subscription.
Lead storyMeta launches Muse Wonder and Muse Wonder Full for AI agents
Edited by Ashutosh ShrivastavaStory 1 of 8 · 2 min read
Meta announced both on 12 October and says they are state of the art on agent tasks in personal and professional settings. The model card says both take text, images, video and audio as input, produce text and are not intended for people under 18.
Meta’s benchmark results for Muse Wonder and Muse Wonder Full against three rival modelsImage: Meta
Meta’s chart of Artificial Analysis Intelligence Index scores, including earlier Muse Spark modelsImage: Meta
Meta trained both at Prometheus, its gigawatt-scale data centre, and highlights two behaviours: Active Memory, which carries what matters about a person from one conversation to the next, and Active Focus, which keeps the model on difficult goals and has it check with the user before important steps.
Price and access
On the Meta Model API, Muse Wonder (muse-wonder-1.0) costs $4.00 per million input tokens and $20.00 per million output tokens. Artificial Analysis notes that Muse Spark 1.3 was priced at $1.25 and $4.25.
Both models have a 1M-token context window. Muse Wonder Full (muse-wonder-full-1.0) has no per-token price: it comes with Muse Code, Meta’s multi-agent coding tool for the terminal, on the Power Usage plan at $50 a month or the Maximum Usage plan at $200 a month.
Key numbers
Terminal-Bench 4.0, in Artificial Analysis’s own tests for its Intelligence Index v4.3, at maximum effort: 52% for Muse Wonder and 57% for Muse Wonder Full, up from 33% for Muse Spark 1.3
Muse Wonder Full narrowly ranks first on CritPt and GDP.pdf, at 33% each, and is in the top three on six of the index’s ten evaluations
$4.52 per task in the index for Muse Wonder, against $1.60 for Muse Spark 1.3, because of higher prices and heavy input use
Safety
Meta calls Muse Wonder its most aligned model and says it carries Meta’s strongest safeguards against chemical, biological, radiological, nuclear and cyber misuse. Its model card says Meta cannot rule out that the model, without those safeguards, would reach Meta’s high risk threshold for chemical and biological threats; with them, Meta rates it “moderate or lower” in every domain. Meta hasn’t published the full safety report its announcement mentions.
Microsoft has announced Maia 300, the third generation of its Azure Maia AI accelerator, and says it will scale the Maia family to multi-gigawatt deployments starting in 2027, when Maia 300 production deployments also begin. Satya Nadella described it on X as Microsoft’s “biggest bet yet” on its own silicon.
The current chip, Maia 200, is in production and serves Microsoft Copilot; GitHub Copilot and Microsoft Foundry are due to follow by the end of 2026.
The chip
Microsoft designed Maia 300 first for inference, the work of running trained models: reasoning over long inputs, chat products that need fast responses and multi-step agent tasks. It expects the chip to deliver 40% more tokens per megawatt than leading GPU-based systems.
Microsoft’s peak specifications for Maia 300, beside a labelled view of the chipImage: Microsoft
Key numbers
Against a Maia 200 rack: up to 7 times the token throughput and up to 2.5 times more performance per watt
Per chip: 2.5 times the memory bandwidth of Maia 200 and 33% more capacity
Per rack: 167 kilowatts, double the accelerator density of Maia 200, with 27.6 TB of HBM and 1.7 PB/s of combined bandwidth
Microsoft plans to say more about its AI strategy at Microsoft Ignite 2026, from 17 to 20 November in San Francisco and online.
SpaceXSI, which makes Grok, has introduced The X Subscription, one plan that covers Grok, Cursor and X Premium. New customers can sign up, and SpaceXSI says it is rolling the plan out to everyone globally.
Five tiers
Free, $0: Grok every day, an entry-level allowance for Cursor’s coding agents and the standard version of X
Plus, $8 a month: more Grok and Cursor usage, and on X a checkmark, creator tools and fewer ads
Premium, $30 a month: the benefits of SuperGrok, Cursor Pro and X Premium+, which SpaceXSI says cost $90 a month when bought separately, with four times the Plus usage
Super, $100 a month, and Ultra, $200 a month: 15 times and 50 times the Plus usage
The shared usage pool resets every week. SpaceXSI says each plan includes at least twice the coding usage of the comparable Cursor plan when its own models are used. Third-party models stay available in Cursor on Premium and above, paid for at API list prices through top-ups.
Teams and current subscribers
Team plans cost the same per seat as Premium, Super and Ultra. SpaceXSI says that is less per seat than the older Cursor Teams plans, with three times more coding usage.
Current SuperGrok, Cursor and X Premium subscribers can keep their plans and prices, or switch and get a prorated refund. The old plans are closed to new sign-ups, and SpaceXSI plans to move remaining customers across, with notice first, but has set no timeline.
Anthropic says distillers now get 90 to 99 percent less data out of Claude
Distillers copy a model’s abilities without its maker’s permission. Anthropic credits a new classifier, tighter limits and bans, and names Zhipu (Z.ai) as one example.
Anthropic calls distillation attacks “covert, industrial-scale campaigns”, often run through networks of fraudulent accounts. Its 12 October research post bases the estimate on early evidence.
Anthropic’s daily count of reasoning tokens exposed to distillers, before and after its safeguardsImage: Anthropic
Anthropic says Zhipu (Z.ai) drew on leading US models to build cyber capabilities into GLM-5.3: first Claude Fable 5, then other models once Claude’s cyber safeguards weakened its attacks. Reuters, reporting the post, said Z.ai had not immediately responded when asked for comment.
Three layers of defence
Anthropic first summarised Claude’s reasoning, which makes its transcripts less useful for distillation.
Anthropic’s test of a model trained on full or summarised Claude transcriptsImage: Anthropic
Distillers then tried to make Claude reveal its reasoning in full. The newer measures:
Preserved thinking: with Claude Fable 5.1 and Claude Opus 5.5, API accounts created from 31 August 2026 can’t edit Claude’s prior context, which makes repeated extraction harder.
A two-stage classifier that detects distillation attempts, which Anthropic says is roughly as robust as its safeguards against helping with bioweapons and wrongly blocks fewer than 0.1% of legitimate requests.
Limits and bans: accounts that keep trying face stricter limits, persistent ones are banned, and since 24 September Anthropic again charges for requests its safeguards block.
Results and trade-offs
On its most capable generally available models, Anthropic says, the classifier detects every distillation technique it described in a September report, and Anthropic has shut down tens of thousands of distillation accounts. Refusing distillers’ paid traffic has cost it meaningful revenue, and it can’t tell how long the improvement will hold: its defences have not faced a highly sophisticated, sustained attempt to evade them.
OpenAI has raised GPT-6.1 Sol’s default context, the amount of text the model can keep in view at once, from 272k to 500k when it is used through Codex or through Work, another OpenAI product. Sottiaux posted the change at 05:15 IST on 12 October.
He wrote that the larger default reduces how often compactions happen and helps with tasks that need a lot of context kept in view. A compaction condenses the earlier part of a long session to make room. He added that OpenAI intends to go higher, without giving a figure or a date.
OpenAI hasn’t published release notes or documentation for the change.
Z.ai pitches GLM-Legal, which is in beta, at everyday legal work such as reviewing contracts, analysing evidence and compliance. It says the model matches frontier closed-source models on LegalBench and on PLawBench, a test of Chinese legal practice, while being much smaller.
Z.ai’s legal benchmark results for GLM-Legal and four other modelsImage: Z.ai
Vals, which tested GLM-Legal independently, ranks it seventh of 79 models on Harvey’s benchmark and 14th of 150 on LegalBench.
Z.ai built the model with extra training on legal material, followed by reinforcement learning, in which the model is rewarded for good results, across tens of thousands of lifelike legal tasks. It says GLM-Legal is coming to AutoClaw, its desktop agent platform, without giving a date. Its examples include a contract review delivered as tracked changes and comments in Word, an analysis of 40 documents from a software project dispute, and a privacy compliance plan for a website launching in Singapore, Malaysia and Indonesia.
Correction, 13 Oct: an earlier version of the benchmark chart’s description said GLM-Legal was close to GPT-6 Astra on PLawBench; it is ahead, at 77.3 against 69.1.
Qwen’s QwenGyre framework speeds up training AI agents on hours-long tasks
Alibaba’s Qwen team moves GPUs between running agent tasks and training as the load shifts, and reports training up to 1.85 times faster than when both share the same GPUs.
Alibaba’s Qwen team has published QwenGyre, a framework for training agents with reinforcement learning, in which a model improves by being rewarded for good results. It targets what the paper calls extreme-long-horizon tasks: jobs in which one run may take hours, with hundreds of exchanges between model and environment and close to 1M tokens. The paper, on arXiv, has 12 authors from Alibaba Token Hub at Alibaba Group, the University of Science and Technology of China and Tsinghua University.
QwenGyre has two core parts. Its Elastic Scheduler shifts GPUs back and forth between training and rollouts as the load changes, without stopping runs already in progress. Rollouts are the runs in which the agent attempts tasks to produce training data. Its Trajectory Processor pieces together agent histories that branch, gives credit for partial progress and strips out repeated paths so training costs stay in check. The framework works with existing agent tools without changing them, and every experiment in the paper used Claude Code 2.1.220.
On Qwen 3.8 2.4T, Qwen’s flagship model, with 700k tokens per rollout, the same number of training steps took 75.42 hours with QwenGyre, against 134.55 hours for an asynchronous baseline and 91.47 hours for a shared-GPU baseline.
Qwen’s training runs on Qwen 3.8 2.4T; Async and Colocate are the asynchronous and shared-GPU baselinesImage: Qwen
Across all the paper’s setups, QwenGyre ran 1.38 to 1.78 times faster than the asynchronous baseline and 1.21 to 1.85 times faster than the shared-GPU one on equal GPU budgets, while matching their training scores. The paper doesn’t link to a code release.
NVIDIA open-sources USD BRep, a library for CAD solid models in OpenUSD
The C++ and Python library builds and edits the precise solid shapes that CAD software works with, and reads and writes them in OpenUSD. NVIDIA published its first release on 9 October.
USD BRep represents each solid as a boundary representation, or BRep, which describes the solid by the surfaces that enclose it. Those surfaces are built from NURBS, a mathematical way of describing smooth curves and surfaces that CAD software uses. Version 1.0.0 is on GitHub under the NVIDIA-Omniverse organisation and is licensed under Apache 2.0.
The Python package installs with pip install usd-brep and runs on CPython 3.12 on Windows or Linux. OpenUSD comes bundled with it, so it doesn’t need to be installed separately. A heal function repairs BRep definitions in USD files, and the repository includes STEP, IGES and JT models, common CAD file formats, for testing.
The repository’s tools carry the SMLIB name, in files such as smlib_gui.py. A CAD developer posting as @0xca0a, who says it has worked with SMLIB and its team for more than 20 years, described the code on X on 12 October as SMLIB, NVIDIA’s NURBS-based geometry engine, becoming open source. NVIDIA isn’t taking code contributions to the project, though it accepts issues and bug reports.
