Hanzhi's BLOG

记录、思考、生活

[市场·2026-05-16] Frontier Radar

发表于 更新于

← 返回主报告：[市场·2026-05-16] Politics

Frontier Radar — 2026-05-16

Window: 2026-05-14 → 2026-05-16 (UTC)
Coverage: OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, Qwen, DeepSeek

Summary

The 48-hour window is dominated by an OpenAI vs. Anthropic skirmish over agentic coding economics: Anthropic put Agent SDK / third-party harness usage on a separate metered credit pool (effective June 15), and OpenAI countered same week by extending Codex to mobile and reportedly offering two months of free Codex to new business customers. Google ramped up to I/O (keynote May 19) with The Android Show pre-event teeing up “Gemini Intelligence”. Meta, xAI, Mistral, Qwen and DeepSeek were quiet inside the strict 48h window after major April / early-May launches (Muse Spark, Grok voice APIs, Medium 3.5, DeepSeek V4, Qwen 3.5).

Labs

OpenAI

  • 2026-05-15 — “Work with Codex from anywhere”: Codex preview lands in the ChatGPT mobile app, with live state sync to laptops / devboxes / remote environments, mobile approvals, model switching, screenshots, terminal output, diffs and test results streaming back to the phone. (OpenAI release page)
  • 2026-05-14 — “Helping ChatGPT better recognize context in sensitive conversations”: Safety / model-behavior update from OpenAI. (OpenAI News)
  • 2026-05-14 (read-out) — Codex business pricing push: Sam Altman said on X that OpenAI is giving new business customers two months of free Codex usage, widely read as a counter to Anthropic’s same-week metering move. (Axios)

Anthropic

  • 2026-05-14 — Claude Agent SDK / programmatic credit pool: Anthropic announced that starting June 15, Agent SDK, claude-p, GitHub Actions and third-party harnesses (e.g. OpenClaw) will draw from a dedicated monthly programmatic credit, separate from interactive subscription limits. Pro tier gets $20 in credits, Max 5x $100, Max 20x $200, billed at API-style rates. This reinstates third-party agent access that had been banned in April. (SiliconANGLE, VentureBeat, InfoWorld, Axios)
  • 2026-05-15 (commercial signal) — Salesforce CEO Marc Benioff said publicly Salesforce expects to spend ~$300M in Anthropic tokens this year on coding/product work, an outsized enterprise-adoption datapoint. (Let’s Data Science)

Google DeepMind

  • 2026-05-12 — The Android Show: I/O Edition (pre-I/O, just outside strict 48h but bridging into the window): Google unveiled “Gemini Intelligence” as the intelligence layer underneath Android, plus Googlebooks (premium Android-powered laptops) and Android Auto upgrades. (Tom’s Guide live blog, Engadget)
  • Pre-I/O posture (May 14–16) — Google I/O keynote scheduled for May 19; a new Gemini version (rumored major overhaul, unified text/image/audio/video/code) is expected; Veo and Lyria to get dedicated stage time. No frontier-model drop inside the strict 48h window. (Tom’s Guide preview, Android Authority)

Meta

  • No verified 2026-05-14 → 2026-05-16 frontier signal. Most recent material event remains the April 2026 debut of Muse Spark (first major proprietary model from Meta Superintelligence Labs, replacing Llama as the consumer Meta AI engine). (CNBC, Miraflow analysis)

xAI

  • No clean 2026-05-14 → 2026-05-16 frontier signal. Most recent items remain early-May: standalone Grok Speech-to-Text / Text-to-Speech APIs, grok-voice-think-fast-1.0 voice agent model, Grok Web Connectors (SharePoint, Outlook, OneDrive, Google Workspace, Notion, GitHub, Linear, custom MCP), and Grok Build beta TUI for SuperGrok Heavy users. (xAI News, Releasebot xAI updates)

Mistral

  • No verified 48h signal. Most recent is 2026-05-01 Mistral Medium 3.5 (128B dense, toggleable reasoning, $1.50/$7.50 per M tokens, weights on HF under a Modified MIT license), plus Vibe cloud agents and Le Chat “work mode”. (the-decoder, Mistral news)

Qwen (Alibaba)

  • No verified 2026-05-14 → 2026-05-16 release. Active context remains Qwen 3.5 (397B MoE, Apache 2.0, 256K context, 201 languages, vision) shipped earlier and Qwen3.5-Omni / Qwen3.6-Plus in April 2026. (Particula, Qwen Wikipedia)

DeepSeek

  • No verified 48h release. Most recent frontier event remains 2026-04-24 DeepSeek V4 Preview (V4-Pro + V4-Flash, MoE, 1M-token context, Hybrid Attention Architecture, open-source). (Bloomberg, TechCrunch, Al Jazeera)

Coverage gaps

  • DeepMind_no_48h_signal — pre-I/O posture only; Android Show was May 12 (outside strict 48h).
  • Meta_no_48h_signal
  • xAI_no_48h_signal
  • Mistral_no_48h_signal
  • Qwen_no_48h_signal
  • DeepSeek_no_48h_signal

Unverified / rumour

  • Google I/O (May 19) Gemini “4.0 vs 3.x” naming — pre-event speculation only; no primary confirmation in the 48h window.