The economics moved into the architecture this week. Grok 4.6 shipped tuned for long-running agents, with a 500,000-token window and a price that doubles once you pass 200K. Anthropic is reported to be in talks to buy a chip-efficiency and video startup for around $6 billion, destined for its inference team. Meta put a genuinely capable 30B agent model under Apache 2.0 and squeezed it onto a single consumer GPU. A new MCP gateway claims to cut tokens by 60% mostly by not calling the model. And an industrial engineering firm shipped an agent platform whose selling point is that it plans before it spawns. Five items, dates marked on each.
A frontier model tuned for long trajectories — and told to check its own work
12 August. SpaceXAI released Grok 4.6 with a 500,000-token context, priced at $2 input, $0.50 cached and $6 output per million, doubling above 200K. It scores 61 on the Artificial Analysis Intelligence Index and trails on coding, 65.9% on DeepSWE v1.1 against GPT-5.6 Sol Max’s 73%. The training is the interesting part: reinforcement learning inside agentic environments, with data regenerated across reasoning-effort levels and agent harnesses. The claim to read carefully is that on longer trajectories the model does more self-testing, checking its own work before moving on. That is vendor observation, not independent validation — and Module 4 is blunt about the pattern: do not ask one agent to generate, approve and publish inside a single context. A model that verifies itself more is welcome. It is not a verifier node, and it does not remove the need for one.
Anthropic is reportedly buying compute efficiency, not capability
13 August. Bloomberg reports Anthropic in talks to acquire Israeli startup Decart for roughly $6 billion, which would be its largest known acquisition. The reporting rests on people familiar with the matter; neither company commented, and the article is explicit that the deal is not finalised and the talks could fall through. Treat it as a report, not an event. What makes it worth noting anyway is the shape: Decart builds world models and, more relevantly, software that optimises chip efficiency during training, plus real-time video generation. The reported destination is Anthropic’s inference team, to absorb more demand and manage computing costs. When a frontier lab spends acquisition money on efficiency rather than on a better model, the binding constraint has moved from what the system can do to what it costs to run.
A capable agent model that fits on one consumer GPU, under Apache 2.0
10 August. Meta released Muse Glimmer, 30 billion parameters including a 1.8B vision tower, dense causal transformer with grouped-query attention and a 131,072-token context, under an Apache 2.0 open-weights licence with BF16, GGUF and ExecuTorch builds on Hugging Face. Four-bit compression brings it under 20GB, so it runs on one 24GB GPU or an M4/M5 Max Mac; two quantised variants trade 0.2% and 1.0% degradation for size. Meta positions it for desktop agents that read screenshots, coding agents, and schema-based function calling. The licence matters as much as the weights — Apache 2.0 means a node in your graph can be one you host, version and inspect, rather than an endpoint whose behaviour changes when someone else ships.
Someone is selling the edge that does not call the model
13 August. SelectHub launched DataGrout, an MCP gateway sitting between agents and enterprise applications, claiming a 60% token reduction on ERP- and CRM-heavy work through dynamic context pruning, symbolic inference layers, and automatic routing of deterministic workloads away from the LLM entirely. The percentage is a vendor claim and should be read as one. The architecture is not novel and that is precisely the point — Module 2’s lesson is that flattening, filtering, joining and status checks are deterministic operations that need no model call, and that a graph paying tokens for its own wiring is paying for nothing. What is new is that this is now a product with a price rather than a design principle in a course. When the plumbing becomes a purchase, the people who understood it as architecture were right early.
An engineering services firm shipped a planning-first agent platform
11 August. L&T Technology Services announced AgenticIQ, an end-to-end platform for autonomous multi-agent workflows across engineering, product development, manufacturing, industrial operations and customer experience, built on what it calls a planning-first architecture with enterprise governance. Details are thin and it is a vendor announcement, so weight it accordingly. The phrase worth keeping is “planning-first”: decide the shape of the work before spawning anything, which is the difference between a graph and a swarm. Industrial engineering is an unforgiving place to run agents — the outputs touch physical systems, and the tolerance for a plausible wrong answer is close to zero. If agentic workflows hold up there, it will be because the planning node did its job, not because the models got better.