Grok 4.6, open weights from Qwen to Meta and Nvidia, security models with lowered refusals, and the EU AI Act gets concrete.
August so far: open weights reach frontier scale, and in the same breath Meta and Nvidia follow up with models that fit on a single consumer GPU. xAI ships an agentic coding model at a fighting price, OpenAI lowers refusals for authorized security work. On top of that, the EU AI Act now applies to everyone whose product talks to users in Europe, and Anthropic starts marking whatever Claude writes.
The first Max-tier Qwen with open weights: a mixture-of-experts model with 2.4 trillion parameters, 95B of them active. The open variant is text only, drops the hosted 1M context, and ships under a custom license; the smaller Qwen3.8-27B followed under Apache 2.0.
Frontier scale you can host yourself, as long as the custom license fits the use.
A frontier agentic-coding model at $2/$6 per million tokens with 500k context, trained in agentic RL environments. Available in the API, Grok Build, and Cursor at launch.
Agentic coding at a fraction of the usual frontier pricing.
30B dense, distilled from the closed Muse Spark, under Apache 2.0, with 131k context and image input. Built for agents that run locally and permanently on a single consumer GPU. The weights for Muse Spark 1.2 are meant to follow.
An agent model that runs on hardware you already own, with no cloud in the loop.
An open 30B mixture-of-experts model with 3B active parameters that mixes Mamba-2 layers with attention and, by Nvidia's numbers, generates tokens up to four times faster than open models in its class. Alongside it comes NeMo Switchyard, an open source library that hands each step of an agent workflow to the model that fits it.
Few active parameters, high throughput: built for agents that never stop running.
The 1.6T flagship is now generally available on the DeepSeek API, topping off the V4-Flash weights opened in July.
Open Flash at the bottom, hosted Pro at the top: both ends are covered now.
Google's new workhorse for coding and agents, with introductory pricing through the end of 2026. It follows the GA of 3.6 Flash by only three weeks.
The introductory pricing runs out at the end of the year, so budget past it.
Plus and Pro now steer how long Sol thinks, Free and Go get Luna as their default model. OpenAI puts the drop in answers carrying at least one factual error at 68 percent against GPT-5.5 Instant, and rates the August versions High capability in cybersecurity and bio-chemistry.
If your team works in ChatGPT, the answers change shape starting now.
A Sol-based model with lowered refusals for authorized security work, gated behind a vetted Daybreak Red tier; Daybreak Blue gives defenders frontier models with defensive guardrails.
Relevant even without access: it shifts what attacker tooling can do.
Plugins can be searched and installed across local, workspace, and remote catalogs, plus scoped auto-approval via --approve-for-me. A Linux desktop preview covers Ubuntu, Debian, and Fedora.
Plugins turn the CLI into a platform a team maintains together.
0.148 exports conversations as Markdown and lets you fork a session instead of restarting it. 0.149 puts an interactive dashboard over the running tasks.
Forking instead of restarting rescues the context of a long session.
With 2.1.234, sessions can message each other by name, continue automatically when a usage limit resets, and GitLab merge requests reach parity with GitHub. Subagent forking is now on by default, and 2.1.237 adds a built-in Concise output style that leads with the result and skips the preamble.
Concise changes the tone, session messaging changes how work is cut up.
Selectable in the model picker across VS Code, Visual Studio, JetBrains, Xcode, and the Copilot CLI.
Off by default for Business and Enterprise until an admin enables the policy.
Regional endpoints are generally available and keep inference in the EU or the US, joined by a Priority Tier with a committed service level. Multi-year commitments from Amadeus, ASML, Capgemini, Caisse des Dépôts, and CMA CGM are meant to underwrite a gigawatt of European compute by 2030. Open models from other labs will run on the same infrastructure, starting with GLM-5.2 from Z.ai.
EU data residency without running it yourself, and a European place to serve other labs' open weights.
The Frontier Red Team catalogs failure modes in emerging multiagent deployments.
Directly applicable if you orchestrate subagents in production.
Since August 2, AI content labeling, chatbot disclosure, and the penalty regime apply; the heavy high-risk obligations moved to 2027 and 2028.
If your product talks to EU users, disclosure and labeling are live requirements now.
Text gets an invisible watermark, images and SVGs signed C2PA metadata; models from August 2 onward mark from day one, older ones follow.
Satisfies Article 50(2) with no work on your side, but it does not survive heavy editing or a format change.
