AI Daily · Nov 19, 2026: Google releases Gemini 4 Neon with better steerability and lower latency

Gemini 4 Neon is live in the Gemini API, AI Studio, Cursor and OpenRouter with a 1M-token context at $3/M input and $15/M output; Tencent open-weights Hy4 flash with 1M context.

Pick a date 11/2026: 19 briefings
11/2026
SunMonTueWedThuFriSat
22232425262728
2930

Dotted dates have a briefing; click one to open it.

  • Google releases Gemini 4 Neon with improved steerability and instruction following

    Gemini 4 Neon targets complex software engineering, knowledge work and science with a 1M-token context; it is live in the Gemini API, AI Studio, Cursor and OpenRouter, and available to AI Pro/Ultra subscribers in the Gemini app and Antigravity, with Tasks runnable on Neon to stretch limits. It scores 63.7% on CursorBench 4.1 at default effort, priced at $3/M input and $15/M output tokens.

    @koraykv 13 duplicate reports merged
  • Tencent open-weights Hy4 flash with 1M context

    Hy4 flash is released as open weights with FP8 and Base variants, 224B total and 12B active parameters, 1M context and native multimodality; Hugging Face hosts 4-bit and 8-bit MLX checkpoints for local Apple Silicon runs. It stays exclusively free on OpenCode for 10 more days with near-unlimited usage.

    @TencentHunyuan 5 duplicate reports merged
  • Qwen4-Max free for five days in Nous Portal

    Alibaba's new flagship Qwen4-Max is free for 5 days in Nous Portal via a partnership with Inferact and Alibaba Qwen; the model targets long-horizon tasks with native vision and a 1M context window.

    @NousResearch 2 duplicate reports merged
  • Perplexity launches Context API for URL content extraction

    Context API reads up to 10 URLs per call and returns clean markdown, HTML, links and snippets, with 92.3% of URLs yielding usable content and 87.9% fact recall at $1 per 1000 URLs.

    @perplexity_ai
  • Mozilla tested Claude Mythos on Firefox codebase, fixing 400+ security bugs

    Mozilla quietly ran Claude Mythos against Firefox's 10-million-line codebase before Fable's release and removal, yielding over 400 security bug fixes including long-hidden flaws.

    @claudeai
  • Datasection slashes revenue guidance 89%, stock hits limit-down

    Datasection (3905.T) cut FY revenue guidance from ¥162B to ¥17B, an 89% reduction, sending the stock limit-down; the FT has reported the firm functions as a Tencent datacenter with Nvidia Blackwell GPUs.

    @SemiAnalysis_ 1 duplicate reports merged
  • Hy4-preview to be discontinued December 7, 2026

    Per Tencent Cloud's schedule, Hy4-preview stops service at 00:00 SGT on December 7, 2026; developers should migrate to Hy4 and update model IDs and prompts.

    @BAI_AGI 1 duplicate reports merged
  • OpenCode harness adapters for Qwen3.8-27B reach 51% on SWE-Bench Pro

    Community harness adapters for Alibaba's Qwen3.8-27B on the OpenCode harness are released for local inference, trained on v2. SWE-Bench Pro (Hard 51 subset) pass@1 rises from 35.3% to 51.0%, and from 47.1% to 54.9% in xhigh thinking mode.

    @huggingface
  • OpenAI models improve AMD GPU power efficiency by up to 14%

    OpenAI says its models found up to 14% PPA improvement on AMD's flagship GPU IP in under five months, and also improved Jalapeño, OpenAI's first custom inference chip, with gains verified on real silicon.

    @OpenAI 1 duplicate reports merged
  • GLM 5.3 on OpenCode Go plan scores high on benchmarks

    A user benchmarked the GLM 5.3 model offered on OpenCode's Go plan at $10 for $500 of usage and recorded GPQA Diamond 92.9%, MMLU-Pro 88.7%, MATH-500 100% and AIME 2025 98.3%, with no sign of aggressive quantization damage.

    @jayleaton 1 duplicate reports merged