AI News August 16, 2026: Z.ai's GLM-5.3 Turns a Coding Model Into a Cyber Defender, Grok 4.6 Matches the Frontier at Half Price, Anthropic Banks Its First Profit, DeepSeek Flips the Pricing Switch Today

🚀 Top 5 AI Stories — August 16, 2026

The weekend arrives, but the AI industry doesn’t. This Saturday’s landscape: a Chinese lab taught an existing model brand-new tricks — including cybersecurity skills it says it never planned for — an American rival matched the frontier’s intelligence scores at half the price, Anthropic quietly proved that frontier AI can be profitable, and DeepSeek flipped on a pricing model no Western lab has tried. Plus: two deadlines now just 15 days away that every engineering team should have on their calendar.

1. 🇨🇳 Z.ai’s GLM-5.3 Learned Cybersecurity On Its Own — and It’s Already Finding Bugs

Z.ai shipped GLM-5.3 on August 14, and the release is a strange one: the base model is unchanged from GLM-5.2. Everything new comes from extended post-training — and the gains are anything but incremental:

• DeepSWE v1.1: 66.9%, up from 46.2 — over twenty points in one release
• Terminal-Bench 3.0: 28.3, up from 4.6
• CyberGym: 84.5%, the best score on the benchmark — ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%)

That last number matters most. Z.ai says the model’s vulnerability-discovery skill was emergent, not planned — and per VentureBeat, GLM-5.3 has reportedly already found a serious vulnerability in Cursor. It’s not universal dominance: on ExploitBench it scores 54.4%, well behind GPT-5.6 Sol (76.5%) and Mythos 5 (78%), and Claude Fable 5 still leads Z.ai’s internal Code Bench at 39.5%.

The efficiency story is the quiet headline: GLM-5.3 hits 31.4% on Z.ai’s Code Bench using ~50,000 output tokens per task, versus Claude Opus 4.8’s 29.5% at 120,000 tokens — more work from far less compute. Access is currently limited to the GLM Coding Plan and ZCode environment; API access and open weights arrive once safety hardening completes, roughly two weeks after launch — which is when independent verification finally begins.

2. ⚡ Grok 4.6 Ties GPT-5.6 Sol on Intelligence — at Roughly Half the Price

xAI released Grok 4.6 on August 12, five weeks after Grok 4.5, and it lands at 61 on the Artificial Analysis Intelligence Index — matching GPT-5.6 Sol and gaining five points over its predecessor. The pricing is the story: $2 per million input tokens and $6 per million output tokens, about half of what equivalent frontier models charge.

• 500,000-token context window, text and image input, no output limit
• Reasoning effort configurable from low up to a new “xhigh” tier
• Default model in Grok Build, available in Cursor, OpenRouter, Vercel, and Cloudflare gateways
• Built specifically for long-running agents and agentic coding

The caveat: as of mid-August, no independent third party has replicated the benchmark figures, and on xAI’s own ten-row comparison table, Claude Fable 5 still wins the most rows. Treat the numbers as directionally strong, not settled.

3. 💰 Anthropic Books Its First-Ever Operating Profit: $559M on $10.9B Revenue

The most consequential business news of the week: Anthropic’s Q2 2026 results show $10.9 billion in revenue — up 127% quarter-over-quarter from $4.8B — and its first operating profit ever, at $559 million (a ~5.1% margin, including model training costs but excluding stock-based compensation).

Two numbers explain how the profit became possible: revenue more than doubled in a quarter, and compute cost per dollar of revenue fell 21% in the same quarter. The primary growth driver is enterprise adoption of Claude Code, the agentic coding tool. The market response is already priced in: investors now eye a $2 trillion valuation for a listing expected as early as October 2026, after a confidential IPO filing in June — potentially the largest IPO on record.

4. 🐋 DeepSeek Flips the Pricing Switch — Off-Peak Rates Start Today

Two DeepSeek moves landed within 48 hours:

• V4 Flash input price rose 93% on August 14, from $0.14 to $0.27 per million tokens — still cheap, but the gap to Western models has narrowed
• Peak/off-peak pricing goes live today, August 16 at 16:00 UTC, with off-peak rates at 50% of peak — a time-based incentive no Western frontier lab currently offers

Add the reminder that deepseek-chat and deepseek-reasoner are deprecated on October 24, and the message is clear: DeepSeek is trading raw cheapness for yield management. Batch workloads that can run overnight just got a 50% discount.

5. ⏳ 15 Days to the August 31 Deadline Cluster

If your stack touches Claude or Kimi, circle August 31:

• Claude Sonnet 5 input pricing rises from $2 to $3 per million tokens — front-load batch workloads now
• Kimi’s API sunset (K2.5 / Moonshot v1) completes the same day — migrations should already be underway

Combined with DeepSeek’s October 24 deprecation, Q3’s back half is a window for cost-model housekeeping that won’t reopen at these prices.

📊 The Day in Numbers

🔮 What to Watch Next

• GLM-5.3 open weights (~end of August): the moment independent labs can finally verify Z.ai’s numbers
• Pricing responses: whether OpenAI or Google answer DeepSeek’s time-based pricing and xAI’s $2 frontier tier
• Anthropic’s IPO paperwork: a $2T October listing would reset every private AI valuation
• Grok 4.6 replication: watch for third-party benchmark runs as traffic ramps
• The Cursor vulnerability disclosure: whether GLM-5.3’s first find gets a CVE and a patch timeline

This article was compiled from multiple sources on August 16, 2026. All figures and claims are attributed to the original publications linked in the sources list above.