LLM News: Grok Bots, GLM-5.3, Muse Glimmer & More

> **TL;DR:** This week's biggest LLM news: xAI shipped autonomous Grok Bots that learn tasks by observation, Meta released a free 30B-parameter offline model called Muse Glimmer, and OpenAI split GPT-5.6 into a security-hunting Cyber variant and a 14x-faster Ultrafast mode. Z.AI, DeepSeek, and xAI also pushed new agentic and coding models that undercut Western pricing.

Key Takeaways

- xAI's Grok Bots watch a task once, then repeat it autonomously and hand off work between bots - Meta's Muse Glimmer runs a 30B-parameter model fully offline in under 20GB of memory - OpenAI split GPT-5.6 into a vulnerability-hunting Cyber model and a 14x-faster Ultrafast mode - GLM-5.3, DeepSeek V4 Pro, and Grok 4.6 are squeezing agentic-coding prices well below Opus 4.8 - Anthropic added invisible watermarking to Claude outputs while deepening its Chrome side panel

Agents that watch, learn, and hand off work

The clearest signal from this week's [LLM updates](https://speka.info/llm-updates/) is that the industry has moved past chat-based assistants and into standing, autonomous teammates. xAI's new **Grok Bots** are cloud-based AI workers that learn a task by watching a human do it once — sending an email, checking a calendar, filing a report — and then repeat it on their own. Each bot runs on its own dedicated cloud computer, and multiple bots can hand off pieces of a workflow to one another, effectively forming a small automated team rather than a single model answering prompts.

That pattern — observe once, execute forever — is a meaningful shift from prompt-and-response usage. It also raises the stakes on oversight: a bot that has "learned" a task by imitation still needs guardrails on what it's allowed to touch, and xAI hasn't published details on how permissions or audit trails work across handoffs. Teams evaluating this kind of tooling should treat it the way they'd treat any new agent-ready repo — worth testing in a sandboxed account before granting real inbox or calendar access, a caution we've echoed when covering agent-ready projects like [Block Buzz & Wasp](https://speka.info/blog/block-buzz-wasp-two-agent-ready-github-repos).


Meta pushes a capable model fully offline

Meta's answer to the privacy side of the agent race is **Muse Glimmer**, a free, 30-billion-parameter model designed to run entirely on a personal computer, offline, in under 20GB of memory. In practice that means multi-step tasks — Meta cites building a dashboard as an example — can execute locally without any data leaving the machine. For a model of that size to fit comfortably on consumer hardware is a genuine compression achievement, and it puts real agentic capability within reach of anyone without a cloud subscription or an internet connection.

The offline angle matters beyond convenience. Enterprises with data-residency requirements, and individual users wary of sending drafts, calendars, or financial spreadsheets to a third-party server, now have a credible local option instead of a cloud-only trade-off. Expect more releases in this vein as parameter-efficient training keeps closing the gap between local and hosted models.

The coding and agent model race gets cheaper and faster

Three separate releases this week all target the same battleground: agentic coding. Z.AI's **GLM-5.3** improves on GLM-5.2 with fewer output tokens per task and benchmark results the company positions as competitive with Kimi K3, Claude Fable 5, and GPT-5.6 — and it's rolling out free after staged safety evaluations. DeepSeek's **V4 Pro** claims to match or beat GLM-5.2, Kimi K3, and Opus 4.8 on agent benchmarks, adds adjustable thinking time for harder problems, and cut its API price by 50% starting August 16.

xAI's **Grok 4.6** rounds out the trio with better codebase understanding, longer autonomous task runs, and deeper hooks into GitHub, Slack, and Figma. Its pricing — $2/$6 per million input/output tokens — undercuts Opus 4.8's $5/$25 by a wide margin. For developers picking between models on cost alone, that gap is significant enough to change which model ends up wired into a CI pipeline or coding agent by default. If you're weighing access to premium models against tighter budgets, it's worth comparing against router options like [OmniRoute](https://speka.info/blog/omniroute-free-ai-model-router-unlocks-claude-opus-4-6), which has been unlocking free-tier access to models like Opus 4.8.

Taken together, these three releases confirm a trend that's been building for months: agentic coding is now a genuinely competitive, multi-vendor market, and the models worth benchmarking against each other change on a roughly monthly cycle. Anyone building tooling on top of a single model provider should budget time to re-test against these newcomers.

OpenAI splits GPT-5.6 into specialist modes

Rather than one general-purpose update, OpenAI shipped GPT-5.6 in two distinct flavors. **GPT-5.6 Cyber** is a security-focused variant built to find vulnerabilities before attackers do — OpenAI reports it completing 95% of advanced security tasks in testing, versus 1.5% for the base GPT-5.6. Access is currently restricted to approved security researchers, which suggests OpenAI is being deliberate about who gets a model this good at finding exploits.

Separately, **GPT-5.6 Ultrafast** trades some flexibility for raw speed, running up to 14x faster than standard GPT-5.6. On one reasoning benchmark, it reportedly finished in 11 hours versus 78 hours for Claude Fable 5 on the same task — a striking gap if it holds up across other workloads, though independent benchmarking will matter more than any single vendor-reported number.


Deeper integrations: ChatGPT, Claude, and Hermes

On the productivity side, ChatGPT now integrates directly with **Google Workspace** — users can tag a Drive folder and work with Docs, Sheets, and Slides without leaving the chat window, including creating and editing Docs inline. A separate desktop feature, **"computer history" memory**, reviews activity across apps like Slack, email, and calendar to summarize what's been done and surface likely next tasks or blockers.

Anthropic moved in a similar direction, upgrading Claude's **Chrome side panel** to read and act across multiple open tabs — for example, pulling data from several invoices open at once — and automatically pushing updates into connected Google Sheets. Anthropic also introduced **invisible watermarking** for Claude-generated text, images, and files, embedded in a way that survives copying and editing so origin and edit history can be verified later. That's a meaningful step for provenance as AI-generated content spreads across the web, and it pairs with the kind of workflow automation we've tracked in other agent tooling roundups, including [GitHub Weekly Wins](https://speka.info/blog/github-weekly-wins-12-repos-worth-starring-now).

Image tools, infrastructure money, and what to watch

On the creative side, **Grok Imagine Image 2.0** adds selective and localized editing, smart resizing across aspect ratios, and better handling of complex instructions like generating books with embedded diagrams and text — alongside templates for e-commerce, UGC, and game assets. Google also open-sourced a DIY project pairing a small **Gemma** model with a Raspberry Pi 5, battery, touchscreen, and speaker to build an offline speech translator with no internet dependency.

Behind all of this sits a much larger number: **Nvidia** partnered with BlackRock, Goldman Sachs, and other investors to raise over $500 billion for AI infrastructure, structured so investors can earn from hardware usage revenue. That scale of capital commitment is a strong signal that the compute buildout behind this week's model releases isn't slowing down anytime soon — if anything, the pace of releases from xAI, Meta, Z.AI, DeepSeek, and OpenAI in a single week suggests the infrastructure is already catching up to demand, not waiting on it.

Frequently Asked Questions

What are xAI's Grok Bots?

Cloud-based AI teammates that learn a task by watching a person perform it once, then repeat it autonomously — such as sending emails or checking schedules — with each bot running on its own cloud computer and able to hand off work to other bots.

What makes Meta's Muse Glimmer different from other AI models?

It's a free, 30-billion-parameter model that runs entirely offline on a personal computer in under 20GB of memory, so multi-step tasks execute locally without sending data to the cloud.

What's the difference between GPT-5.6 Cyber and GPT-5.6 Ultrafast?

GPT-5.6 Cyber is a security-focused variant for finding vulnerabilities, restricted to approved researchers. GPT-5.6 Ultrafast is a speed-optimized mode that runs up to 14x faster than the base model.

How does Grok 4.6 pricing compare to Claude Opus 4.8?

Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens, compared to $5/$25 for Opus 4.8.

What does Claude's new invisible watermarking do?

It embeds hidden watermarks in Claude-generated text, images, and files that persist through copying and editing, allowing origin and edit history to be verified later.