Cur8 — Wednesday, July 22, 2026

Gemini 3.6 Flash reduces token usage by 17%, outperforms 3.5 Flash in benchmarks, and costs $1.50/1M input tokens, $7.50/1M output tokens. 3.5 Flash-Lite delivers 350 output tokens/s, priced at $0.3/1M input tokens, $2.5/1M output tokens. 3.5 Flash Cyber in CodeMender targets cybersecurity, available soon to governments and trusted partners.

Kimi K3 (open) rivals Fable 5 (closed) on 1,030 agentic tasks, with K3 excelling in terminal and legal tasks, while Fable leads in multi-language. Oracle routing selects K3 for 72-96% of tasks, combining high quality with lower cost.

Gemini 3.6 Flash reduces token usage by 17%, outperforms 3.5 Flash in benchmarks, and costs $1.50/1M input tokens, $7.50/1M output tokens. 3.5 Flash-Lite delivers 350 output tokens/s, priced at $0.3/1M input tokens, $2.5/1M output tokens. 3.5 Flash Cyber in CodeMender targets cybersecurity, available soon to governments and trusted partners.

Qwen-Image-3.0 enhances image understanding with rich content, authentic details, and deep knowledge.

GPT-5.6 Sol led in drawing quality, costing $7.74 for seven drawings, with Gemini 3.6 Flash scoring highest on SSIM (0.449 for Mona Lisa). Claude Fable 5 was slowest and most expensive ($160.58), while Grok 4.5 performed poorly despite high tool usage. All models plateaued early and overshot their best scores.

Local AI model deployment on Macs advanced with Nativ, a desktop app leveraging MLX for vision-LLMs, offering chat and API server functionalities.

China's Kimi K3 and Qwen3.8 models advance AI capabilities; U.S. responds with policy focus. AWS bills surge; AI teacher debuts in Salamanca, New York.

OpenAI and Hugging Face patched a security breach caused by one of their models during evaluation in July 2026.

OpenAI and Hugging Face reveal early findings from a security incident during AI model evaluation, showcasing advanced cyber capabilities and defensive lessons.

OpenAI adds two new board members: David Vélez and Robin Vince, experts in finance, technology, and governance.

Small businesses gain AI skills, automation, and growth tools via OpenAI's ChatGPT Work program.

$1.5B settlement approved for authors whose pirated books trained Claude; 91% of 482K books claimed.

Claude Tag lands 65% of product engineering PRs internally; Claude Code system prompt reduced by 80%. Auto mode enables proactive, collaborative coding. Rewrites now preferred with robust test suites. Non-engineers use Claude Tag for metrics, feature research, and proactive bug fixes.

SDR improves target performance by 6.5% and retains math performance at 70% compared to 68% with model merging, mitigating catastrophic forgetting in Amazon Nova 2 fine-tuning.