On August 12, models moved toward opposite ends of the scale. xAI sells a 500,000-token context with a price cliff after 200,000. Qwen opens the weights of a 2.4-trillion-parameter MoE. Liquid and Cohere move the other way, putting vision and document work into 2–3 billion parameters for the edge.
A distribution layer is forming between them: GitHub ships skills and MCP as one plugin, while infrastructure companies show how much power, networking, and future contracting the whole system requires. Model scale can no longer be discussed without machine scale and the price of actual use.
Models
Grok 4.6 sells long context on a stepped price curve
Grok 4.6 launched in the API, Grok Build, and Cursor with a 500,000-token context window. Up to 200K tokens, input costs $2 per million, cached input $0.50, and output $6; above the threshold, prices double. xAI claims 69.9 on CursorBench and 65.9 on DeepSWE.
The 500K is a maximum length, not a guarantee of effective recall. The benchmarks mix xAI's results with figures published by competitors and provide no independent cost-normalized comparison.
The price cliff makes context's architectural cost visible. An application that indiscriminately accumulates history and tool output can cross into another tariff with one message. Long context does not eliminate retrieval and memory policy; it makes their financial mistakes more expensive.
Sources: Grok 4.6, model documentation.
Open 2.4-trillion-parameter weights still require an enormous serving system
Qwen published the Qwen3.8-2.4T-A95B weights: an MoE with 2.4 trillion total, 95 billion active parameters and 92 layers. The repository dates the weight release to August 12, though a separate cloud endpoint existed earlier.
Parameter count does not equal memory or compute per request: precision, sharding, KV cache, and runtime all matter. Vendor benchmarks are not independent.
Active 95B substantially reduces cost relative to a dense 2.4T model, but does not make it local in the everyday sense. Open weights grant ownership of the checkpoint, not a free machine to run it. At this scale, freedom from an API provider means dependence on one's own distributed serving.
Sources: Qwen model card, repository.
Edge vision splits between interfaces and documents
LFM2.5-VL-3B combines a roughly 3.1B language backbone with a SigLIP2 NaFlex encoder. Liquid targets screenshots, UI localization, OCR, grounding, and tool use. In its own ablations, ToolSandbox rose from 26.4 to 59.5, BFCL from 20.5 to 32.5, and RefCOCO from 57.1 to 87.9.
Cohere's North Micro Vision uses a 400M native-resolution encoder and a 2B language model, preserves document aspect ratios, and supports 11 languages. Its curriculum covered tens of millions of examples up to A4 resolution; the model shipped under Apache 2.0.
Both sets of numbers come from their authors. Liquid supplies no hardware latency to support the word "edge," and its license is not Apache or MIT. Cohere still only promises vLLM support, while some competitors perform better on STEM and hallucination. The day's signal is specialization even among small VLMs: a UI agent and document parser need different resolution, grounding, and runtime tradeoffs.
Sources: LFM2.5-VL-3B, North Micro Vision.
Distribution layer
Agent Plugins 1.0 turns skills and MCP into a portable package
GitHub enabled Agent Plugins 1.0 in VS Code, Copilot CLI, the SDK, and the Copilot App. A package combines Agent Skills and MCP-server configuration; Business and Enterprise receive managed policy. The specification appeared on August 6, and GA distribution across products began on August 12.
Version 1.0 says nothing about the number of compatible vendors. "Portable" does not mean identical permissions or UX in every harness, while retaining old plugins may preserve earlier unsafe agreements.
Even so, agent capability gains a recognizable delivery unit. Procedure and tools arrive together, and an organization can control their availability. Portability must now be joined by provenance, versioning, and separate consent for each MCP server's permissions.
Source: Agent Plugins 1.0 GA.
Cybersecurity
A four-day campaign showed near-autonomous agent operation, not a "fully autonomous attack"
Dream Security described a recovered 160 MB workspace with 1,395 files tied to a July 1–4 campaign using Hermes and OpenClaw. Up to eight subagents explored paths in parallel and changed tactics against 21 systems; researchers report 85 accounts and more than 2,500 personnel records.
All operational numbers come from Dream's report and contemporaneous coverage. The primary report URL was not found, the model is unknown, and the China link rests on indirect indicators. Taiwan's government confirmation arrived only on the 13th and is not used to strengthen this day's account.
The operator supplied goals and infrastructure, so "fully autonomous" exceeds the evidence. Near-autonomous execution is more defensible: a human defines the campaign, while agents scale reconnaissance and adaptation within it.
Source: contemporaneous report on the research.
Infrastructure economics
Nebius grows faster than it can bring promised capacity online
Nebius reported quarterly revenue of about $582 million, up 454% year over year, and positive adjusted EBITDA for its AI cloud. Its investor deck shows more than $8 billion in cash, roughly $2.3 billion in operating cash flow, and contracted power guidance raised from more than 3 GW to more than 4 GW.
Revenue covers the whole group, adjusted metrics are not GAAP, and 4 GW is a contracted pipeline rather than operating capacity. Four large deals with average TCV above $1 billion are spread across future years.
Demand is already measured in contracts and money, but physical delivery trails the headline. Energy becomes a promise of future compute supply that still has to be financed and built.
Sources: Nebius results, SEC presentation.
Cerebras nearly quadruples cloud revenue and accumulates $25.4 billion in obligations
Cerebras reported GAAP cloud revenue of $126 million, up 281%, and total revenue of $180.1 million, up 74%. Remaining performance obligations reached $25.4 billion, while live and contracted power is expected to exceed 600 MW by the end of 2027.
Core cloud revenue is non-GAAP, RPO is future obligation rather than cash, and 600 MW combines live and contracted capacity; the manufacturing expansion does not disclose its baseline.
As with Nebius, the main question is not demand but conversion of backlog into a working service. Concentration and execution risk are especially high with specialized hardware: a contract cannot be served from someone else's interchangeable fleet if the company's own machine is late.
Sources: Cerebras release, 10-Q.
Cisco breaks networking out as its own line of AI capex
Cisco ended its fiscal year with $9.3 billion in AI-infrastructure orders from hyperscalers, $4 billion of them in the fourth quarter. The company says it recognized about $4 billion of related revenue in FY2026 and expects $7.5 billion in FY2027.
Orders, recognized revenue, and guidance are three different quantities. Cisco does not disclose customer concentration or product mix, so the entire amount cannot be attributed to Ethernet fabrics or one silicon generation.
What matters is the network's scale as a product of the AI boom in its own right. GPUs do not form a cluster without a fabric, and hyperscaler orders are making this hidden portion of capital spending visible in financial reports.
Source: Cisco FY2026 results.
The issue's main technological shift
On August 12, "model" stopped denoting comparable products.
One product sells 500K context through an API; another opens 2.4T weights for distributed serving; a third puts vision at the edge. A plugin binds a model to skills and tools. Beneath them, cloud providers sell capacity not yet built, while networking becomes a billion-dollar line item.
Comparing these systems with one benchmark is meaningless. We need the complete unit of deployment: capability, context, price, runtime, hardware, data boundary, and infrastructure available today rather than in a future contract.
What to discuss with the technical team
- Where does long context cross a pricing threshold, and which data need not be sent again?
- What exactly do open weights provide when serving remains distributed infrastructure?
- Does a small edge model's license meet the product's requirements?
- Are MCP-server permissions granted separately from plugin-package installation?
- Which power and networking figures are live capacity, and which are merely contracted guidance?