All AI model reports
Collected in Hong Kong time

AI Model Daily - 2026-05-26

2026-05-26 AI model news: no new flagship launch appeared in the last 24 hours; the main signals are Claude Mythos Preview security research, Grok Build agentic coding access, ERNIE 5.1 resurfacing through leaderboard and media signals, Qwen3.7-Max availability, and small video leaderboard changes.

30-Second Summary

  • No new flagship model announcement from OpenAI, Google, Anthropic, Meta, Mistral, or Qwen appeared in the last 24 hours. Today's signals are mostly model capability evidence, availability confirmation, and small leaderboard movement.
  • Anthropic Claude Mythos Preview remains the most important non-public model capability signal: the Project Glasswing initial update shows high-intensity results on open-source vulnerability discovery, but there is still no public access plan.
  • Grok Build 0.1 stays in the agentic coding watchlist: the access ecosystem where X Premium and Grok usage power Openclaw, Opencode, and Hermes is worth validating, but public benchmark coverage is still thin.
  • ERNIE 5.1 should be added to the tracking pool: Baidu officially released it in May, a secondary roundup mentioned it again today, and LMArena Text Overall shows ernie-5.1 at rank 18 with Elo 1473 ±7.
  • Video leaderboards only moved slightly: Dreamina Seedance 2.0 720p and HappyHorse-1.0 continue to lead Artificial Analysis, while Gemini Omni still does not appear on that leaderboard.

Must Watch Today

1. Claude Mythos Preview / Project Glasswing

  • Event type: official research update + model capability signal.
  • Official update time: Anthropic published the Project Glasswing initial update on 2026-05-22.
  • Capability change: Anthropic says Mythos Preview scanned 1,000+ open-source projects and identified an estimated 23,019 vulnerability items, including 6,202 high- or critical-severity items. Anthropic's X account continued to highlight Project Glasswing and Claude Mythos Preview's vulnerability discovery capability on 2026-05-23.
  • Availability: Mythos Preview is still not a public model. It is provided to partners through Project Glasswing for defensive security research.
  • Impact: this is not a normal product launch, but it directly matters for whether models can perform long-horizon, high-value security research tasks. Continue tracking whether Mythos gets broader controlled access, or whether its capabilities migrate into the public Opus/Sonnet lines.
  • Sources: Anthropic Project Glasswing update, Project Glasswing, Anthropic X

2. Grok Build 0.1 / Grok Agentic Coding Access

  • Event type: model availability + X access-ecosystem signal.
  • Observed timing: OpenRouter and model indexes show a mid-May 2026 launch; @xiaohu noted that X Premium members can directly authorize Openclaw, Opencode, and Hermes, using X Premium Grok usage to power those products.
  • Capability change: third-party indexes describe Grok Build 0.1 as a coding model for agentic software engineering workflows, supporting 256K context, tool use, structured outputs, and related agent features.
  • Availability: xAI Console / OpenRouter / third-party Grok coding-agent ecosystem. The X Premium authorization path still needs account-level confirmation.
  • Impact: if you already have X Premium or SuperGrok, the cost structure of powering coding agents through Grok usage is worth evaluating. Public benchmark coverage is still insufficient, so do not assume it beats Codex, Claude Code, or Qwen3.7-Max.
  • Sources: Xiaohu X, LLM Reference Grok Build, KuCoin xAI Grok Build note, OpenRouter xAI models

3. ERNIE 5.1

  • Event type: official release backfill + leaderboard confirmation + media roundup recap.
  • Official release time: Baidu's official blog shows ERNIE 5.1 was released on 2026-05-09.
  • Capability change: Baidu says ERNIE 5.1 inherits the ERNIE 5.0 pretraining base while reducing total parameters to about one third and active parameters to about one half, reaching leading base capabilities with about 6% of the pretraining cost of peer models. Baidu also emphasizes agentic post-training and asynchronous reinforcement learning infrastructure.
  • Availability: Baidu's official site and the Qianfan model platform.
  • Leaderboard status: LMArena Text Overall lists ernie-5.1 at rank 18, Elo 1473 ±7, 9,004 votes, Preliminary.
  • Impact: ERNIE 5.1 is not a new release today, but as a high-ranking Chinese closed model, it should be added to the same comparison group as Qwen3.7-Max, GLM-5.1, and Kimi K2.6.
  • Sources: Baidu ERNIE Blog, LMArena, Codersera ERNIE 5.1 roundup

4. Qwen3.7-Max

  • Event type: availability continuation + leaderboard confirmation.
  • Official release time: Alibaba Cloud Summit announced it on 2026-05-20; Qwen's official X account pinned Qwen3.7-Max on 2026-05-21.
  • Last 24-hour change: no new official upgrade found, but OpenRouter and Qwen Cloud API access remain live, and Qwen X still keeps the OpenRouter availability signal visible.
  • Availability: OpenRouter lists qwen/qwen3.7-max, released May 21, 2026; Qwen Cloud provides OpenAI-compatible API examples.
  • Leaderboard status: LMArena Text Overall lists qwen3.7-max-preview at rank 14, Elo 1475 ±10, 3,741 votes, Preliminary.
  • Impact: still one of the most important closed APIs to test in practice. Focus on tool-call stability, long-context cost, multi-file coding-agent behavior, and comparison with ERNIE 5.1 and GLM-5.1 on complex Chinese tasks.
  • Sources: Qwen X, OpenRouter, Qwen Cloud, Alibaba Cloud, LMArena

5. NetEase Youdao Confucius4 / Confucius4-TTS

  • Event type: open model card and repository confirmation.
  • First signal: @xiaohu posted the two-model Confucius4 signal around 2026-05-22.
  • Current status: the Confucius4 Hugging Face page is still visible, based on Qwen3.5-27B, Apache 2.0; the Confucius4-TTS GitHub repository still says code and model weights are in preparation.
  • Quantitative change: in this run, the Confucius4-TTS GitHub repository had 77 stars and 2 issues; the Hugging Face page had 15 likes. Attention is increasing, but engineering availability is still incomplete.
  • Impact: the multimodal model can be sampled on visual math reasoning first. Keep TTS in watch mode for now, and do not treat it as deployable open-source TTS yet.
  • Sources: Confucius4 Hugging Face, Confucius4-TTS GitHub, Xiaohu X

Arena Watch

  • LMArena Text Overall: leaderboard date is still 2026-05-18. gemini-3.5-flash is rank 9, Elo 1480 ±8; qwen3.7-max-preview is rank 14, Elo 1475 ±10; ernie-5.1 is rank 18, Elo 1473 ±7; all three are Preliminary.
  • Artificial Analysis T2V with audio: Dreamina Seedance 2.0 720p ranks 1st with Elo 1214; HappyHorse-1.0 ranks 2nd with Elo 1213; Veo 3.1 ranks 3rd with Elo 1101.
  • Artificial Analysis I2V with audio: Dreamina Seedance 2.0 720p ranks 1st with Elo 1185; HappyHorse-1.0 ranks 2nd with Elo 1169; Veo 3.1 Fast ranks 3rd with Elo 1094.
  • Watch item: HappyHorse-1.0 is still listed as Alibaba-ATH, Released Apr 2026, API Coming soon, with no formal model card or API documentation found; Gemini Omni still does not appear on the leaderboard.
  • Sources: LMArena Text, Artificial Analysis T2V, Artificial Analysis I2V

X Watch

  • @xiaohu: the latest visible high-engagement content is still X Premium authorization for Openclaw, Opencode, and Hermes, using Grok usage to power related products. This is an AI coding-agent access-ecosystem signal, not a new model announcement.
  • @Alibaba_Qwen: continues to pin Qwen3.7-Max and keeps the OpenRouter launch repost visible.
  • @cohere: Command A+ remains pinned; the Microsoft Azure Foundry distribution signal is still visible.
  • @GoogleDeepMind: still pins Gemini Omni; the 2026-05-23 Singapore AI deployment partnership is not a new model release.
  • @OpenAI: the pinned item remains the planar unit distance problem model capability breakthrough; the 2026-05-22 Codex Thursday update is Codex/product capability, not a model launch.
  • @AnthropicAI: Project Glasswing remains the most valuable model capability signal, pointing to Claude Mythos Preview's security research capability.
  • X search: the page loaded in this run, but results were noisy and included many non-model items; this report prioritizes account pages and official sources.

Pending Confirmation

  • Claude Mythos Preview: still no public access plan. Continue tracking broader controlled access, pricing, API availability, or capability migration into the Opus/Sonnet series.
  • Grok Build 0.1: confirm official xAI docs, CLI access path, X Premium authorization boundary, and real benchmarks. Third-party indexes justify monitoring, but not a conclusion.
  • Gemini Omni API: Google still says access is coming in the next few weeks; continue checking Gemini API, AI Studio, Vertex AI, and Flow.
  • Gemini Omni leaderboard: no independent Artificial Analysis video leaderboard entry found yet.
  • HappyHorse-1.0: leaderboard lists Alibaba-ATH, Released Apr 2026, API Coming soon, but there is still no formal model card or API documentation.
  • Confucius4-TTS: GitHub repository is public, but code and weights are still in preparation.
  • Mistral Medium 3.5: today's roundup mentioned it again, but the official model card shows an April release; do not treat it as a new release today.
  • OpenAI/Google/Meta/Mistral/Qwen: no new major flagship model announcement found in the last 24 hours.

Next Steps

  • Put Qwen3.7-Max, ERNIE 5.1, GLM-5.1, and Kimi K2.6 into the same complex Chinese task and agentic coding comparison group.
  • Add Grok Build 0.1 to the coding-agent cost/effectiveness watchlist, but first confirm official access boundaries.
  • Track Claude Mythos Preview as a capability signal only, not an integration candidate; it remains a controlled security research model.
  • Continue watching Gemini Omni API access and HappyHorse-1.0 API availability; either one would change video-model testing priority.