Meta joins the frontier race as Grok, Cursor and X become one subscription

Meta launches Muse Wonder and Muse Wonder Full, its first frontier models

This is the day Meta stopped being the open-source alternative and started competing head on with OpenAI and Anthropic for paid agent work. Muse Wonder is available now through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, which puts it in the same price bracket as GPT-6 Sol and Claude Opus 5.5. The larger Muse Wonder Full is reserved for Muse Code subscribers, and both will soon power Muse, Meta's personal agent, starting with paying users.

Meta is leaning on two features. Active Memory carries relevant details across conversations, and Active Focus keeps the model working towards a goal when a task gets hard, trying new approaches and checking in before important actions. Meta says the models are state of the art for personal and professional agents and claims Wonder beats Claude Opus and GPT-6 Astra on its MemoryBench test; those are Meta's own figures, and a full report with safety evaluations is still to come. Early independent scoring from Artificial Analysis puts Wonder and Wonder Full at 52 and 56 on its Intelligence Index. It is a big step up from Muse Spark, and because Wonder is already on OpenRouter, a side-by-side test against your current agent model takes minutes.

Grok, Cursor and X Premium merge into one plan called The X Subscription

If you pay separately for SuperGrok, Cursor Pro and X Premium, this could save you real money. The new plans start with a free tier and run through Plus at $8 a month, Premium at $30, Super at $100 and Ultra at $200, with Grok and Cursor drawing on one shared usage pool that refreshes weekly. xAI says the $30 Premium tier covers what used to cost around $90 across the three products.

Read the small print before you switch. Third-party models inside Cursor, such as Claude or GPT, are only available from Premium upwards and are billed separately at API list prices, so heavy Cursor users should check their current usage first. Existing subscribers can switch from their billing settings with a prorated refund, and annual subscribers keep their term until renewal. It is the clearest sign yet of how tightly Cursor now sits inside the Musk stack, which last week also gave Grok Bot its own email inbox.

SubQ Snap is a very cheap long-context model for an agent's routine decisions

Most agent pipelines waste expensive frontier calls on simple routing, filtering and yes or no decisions, and Snap is designed to take those off their hands. It reads up to 384K tokens of multimodal input and returns structured choices, scores or labels, so your main model only gets called for the hard parts. Pricing is $0.05 per million input tokens with output free, and the first 10,000 users get 100 million free tokens through the SubQ console.

It joins a crowded new category: Microsoft's Decision-1 landed on Friday and Inception's Mercury Decide the day before. Snap's pitch is the long context window. Subquadratic's earlier launches drew debate about its benchmark claims, so treat its performance figures as unproven until independent tests appear.

QwenGyre is Qwen's new system for training agents that work for hours

This is a behind-the-scenes release, but it hints at where the next Qwen models are heading. QwenGyre shifts GPUs between running agents and training on their results as demand changes, and turns messy branching runs (sub-agents, retries, context compaction) into tidy training data. Alibaba says it scaled the system to a 2.4 trillion parameter Qwen model with around 700,000 tokens per run, lifting its NL2RepoBench score from 52.5% to 58.5% while cutting training time by about 44%. For everyday users the takeaway is that the open Qwen family is being trained specifically for long, repository-scale coding jobs, where it has trailed the closed labs.

Snowflake previews Horizon Context and shows off a much faster write path

Snowflake's latest release notes add Horizon Context in preview, extending its Horizon governance layer, alongside general availability of better listing access history. The company also shared that the new write path in Adaptive Warehouses made one MERGE job write 545 times less data and finish in 28 seconds instead of 6 minutes 28 seconds, by its own measurement. If your team runs Cortex agents or heavy pipelines on Snowflake, both are worth a look: one for keeping AI answers grounded in governed data, the other for cost.

NVIDIA and Cisco launch new pluggable optics, and NVIDIA opens its Vera Rubin rack designs

Nobody reading this will buy an optical module, but this is the plumbing that decides how fast and cheaply the next generation of models gets trained and served. NVIDIA and Cisco introduced Ultra-Dense Small Form Factor Pluggable (USFP) optics to pack more bandwidth into AI data centres, and at the OCP Global Summit NVIDIA contributed its Vera Rubin rack architecture to the open hardware community. Open designs let more cloud providers build Rubin-class capacity, which should eventually mean more availability and lower prices for the models you use.