AI News Roundup — Week of August 17, 2026

It was a fast week: a fresh wave of open-weight models, an intensifying price war, a shake-up in AI browsers, and big money moving into the plumbing behind AI. Here is what mattered for the tools people actually use.

New models

• Meta released Muse Glimmer, a roughly 30-billion-parameter open-weight model under an Apache-2.0 license, built to run on a single consumer GPU. It marks a notable swing back toward open weights for Meta and lands squarely in the "run a capable agent on your own hardware" trend that has defined much of 2026.
• GLM (Z.ai) shipped GLM-5.3, an open-weight coding and agentic flagship. It is the latest sign that Chinese labs are keeping pace with — and in places matching — Western frontier models on real coding tasks, at a fraction of the cost.
• DeepSeek put out a new, aggressively cheap V4 model, and Alibaba's Qwen refreshed its flagship. Together with Kimi from Moonshot, the open-weight field is now genuinely crowded at the top.
• Google added a faster, lower-cost Gemini Flash tier, while Gemma remains the easiest way to run Google's models locally via tools like Ollama and LM Studio.

Pricing & product updates

• The AI price war escalated on every side. OpenAI reportedly cut the price of its cheapest model by roughly 80%, Google trimmed Gemini Flash rates, and — in a striking reversal of the race to zero — demand for DeepSeek surged so sharply that the company warned developers of a coming price increase (reported). For anyone building on these APIs, the takeaway is that "cheapest model" is now a moving target that is worth re-checking monthly.
• The AI browser story reshuffled. Perplexity's agentic browser Comet became free worldwide, while OpenAI retired its standalone Atlas browser — it stopped working on Aug 9 — and folded agentic browsing into ChatGPT and a Chrome extension. Google, meanwhile, continued pushing Gemini directly inside Chrome. The "browser as the AI agent's front door" is quickly becoming a three-way race.
• On the creative side, Suno shipped Studio 2.0, adding DAW-style capabilities such as MIDI support and stem splitting — a step from "type a prompt, get a song" toward a genuine production workspace.

Money

• Capital kept flowing toward AI inference infrastructure rather than model training. Fireworks AI and Together AI both closed large new rounds this month (reported), and additional funding landed in specialized inference hardware. The pattern is clear: with frontier model quality converging, investors are betting on who can serve those models fastest and cheapest — and on vertical, industry-specific AI built on top.

Tool shake-ups worth knowing

• OpenAI's Atlas browser was discontinued as a standalone app; if you relied on it, agentic browsing now lives inside ChatGPT and Chrome rather than a separate product.
• OpenAI's consumer Sora app is also reported to be winding down in its standalone form, pushing creators toward alternatives like Google Veo, Kling, and ByteDance's Seedance for AI video.
• At Black Hat, security researchers demonstrated code-execution vulnerabilities in several popular AI coding agents. If your team uses agentic coding tools like Cursor or Claude Code, it is a timely nudge to sandbox agent permissions, restrict CI access, and keep secrets out of the model's context.

Figures here are third-party and fast-moving; verify current pricing and availability directly with each vendor before relying on them.