JadidJust-in AI Developments, Incidents & Deadlines
July 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

A quieter day after yesterday's Sonnet 5 / export-control fireworks.

9 stories

The through-line today is quiet deprecation deadlines that bite engineering teams: OpenAI's fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ wind-down hits its next cutoff today, and Windsurf's Cascade local agent reaches end-of-life. On the product side, voice agents and physical AI are where the fresh launches are — xAI ships a no-code voice-agent builder and Meituan open-sources a trillion-parameter coding model trained entirely on Chinese chips.

TLDR

  • OpenAI fine-tuning API hits its next wind-down cutoff today (Jul 2): orgs inactive 60+ days can no longer create new fine-tuning jobs; all self-serve ends Jan 6, 2027 (Jul 2) Pricing
  • xAI launches no-code Voice Agent Builder — build a production voice agent in ~2 min, $0.05/min, 80+ voices + cloning (Jul 1) Agents
  • Meituan open-sources LongCat-2.0: 1.6T-param MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ coding model under MIT, trained end-to-end on domestic Chinese chips (Jun 29) Open Source
  • UBTECH launches UWORLD U1, a mass-produced full-size "ultra-bionic" humanoid line; >13k pre-orders at launch (Jun 30) Robotics/IoT

Action Items

Things your engineering team should look at or do this week.

  • If you rely on OpenAI self-serve fine-tuning, act today. As of Jul 2, orgs inactive for 60+ days lose the ability to create new fine-tuning jobs; all customers lose self-serve fine-tuning by Jan 6, 2027. InferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ on existing fine-tuned models continues until the base model is deprecated — but plan migrations to prompt-based or alternative approaches now. (OpenAI Deprecations)
  • Windsurf users: repoint any Cascade automation before Jul 1's EOL. After Cognition's Jun 2 rebrand to Devin Desktop, the Cascade local agent is end-of-life; CI pipelines, scripts, or rules that explicitly invoke Cascade must move to Devin Local (a Rust rewrite with parallel subagents). Verify local-agent workflows still run. (Cognition/Devin blog)
  • If you run LiteLLM as an AI gateway, confirm you're patched to 1.83.14+. The mid-June privilege-escalation → RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ chain (CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ ~9.9) lets a low-privilege user reach admin and rewrite model responses to steer downstream agents. Patches have been out since mid-June; audit exposed gateways. (Obsidian Security)

Models & Tools

  • xAI Voice Agent Builder — no-code, GA (Jul 1) Agents Customer AI — A no-code platform on Grok Voice: describe call flow in plain language, attach docs/tools/guardrails, and ship a working agent in ~2 minutes. Unified speech-to-speech (vs. stitching STT/LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗/TTS), 80+ preset voices plus 2-minute voice cloning, 25+ languages, sub-second latency claims. Priced at $0.05/min. Aimed squarely at high-volume call-center/booking automation. (xAI)
  • Meituan open-sources LongCat-2.0 (Jun 29) Open Source — A 1.6-trillion-parameter MoE (33B–56B active/tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗) with native 1M context, MIT-licensed, built as an agentic-coding/reasoning engine — and notable for completing full training and inference on a ~50,000-card domestic Chinese compute cluster, a milestone for non-NVIDIA frontier training. Been topping OpenRouter usage; weights are rolling out on GitHub/Hugging Face. (VentureBeat | SCMP)
  • Windsurf is now Devin Desktop; Cascade → Devin Local (rebrand Jun 2, Cascade EOL Jul 1) Agents — Cognition's OTA rebrand ships an Agent Command Center and the open Agent Client Protocol (Apache 2.0), letting Codex, Claude Agent, Gemini CLI, OpenCode and Junie run as first-class agents in one editor. Cascade retires Jul 1 in favor of Devin Local, a Rust rewrite claiming ~30% better token efficiency and parallel subagents. Plans/keybindings carry over. (Cognition/Devin blog)

APIs & Pricing

  • OpenAI: self-serve fine-tuning wind-down — next cutoff today (Jul 2) Pricing — Phased shutdown that began May 7: today, orgs inactive 60+ days can't create new fine-tuning jobs; all self-serve fine-tuning ends Jan 6, 2027. OpenAI's rationale: newer base models (GPT-5.5) follow instructions well enough that prompt-based approaches are cheaper and faster. Inference on existing fine-tunes persists until the base model sunsets. (OpenAI Deprecations)
  • Microsoft 365 Copilot: global pricing update effective Jul 1 (Jul 1) Pricing — New list pricing took effect across purchasing channels: Business Standard with Copilot at $23.50 and Business Premium with Copilot at $32 per user/month. Renewals before Jul 1 could lock in prior rates; budget accordingly for seat expansions. (Microsoft Learn)
  • Google: Imagen 4 deprecation — shutdown Aug 17 (deadline Aug 17) Data Pricing — Imagen 4 models are deprecated and slated for shutdown Aug 17, 2026 (following the Jun 1 shutdown of Gemini 2.0 Flash/Flash-Lite). Migrate image-generation pipelines off Imagen 4 before mid-August. (Gemini API pricing/deprecations)

Security

  • LiteLLM gateway vulnerability chain (CVSS ~9.9) (disclosed mid-Jun; patches available) Security Infra — Obsidian Security's chained CVEsCommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ (CVE-2026-47101/47102/40217) take a default low-privilege user to admin and RCE on the LiteLLM server; once the gateway is compromised, an attacker can rewrite model responses and redirect downstream agents (including coding agents) to attacker-chosen tool calls. Fixed in 1.83.14+. Not new this week, but under-patched and high-impact for teams routing agents through a shared gateway. (Obsidian Security)

Industry & Business

  • UBTECH launches UWORLD U1 humanoid line (Jun 30) Robotics/IoT — Unveiled at UBTECH's Shenzhen Global Launch Event as the first full-size mass-produced "ultra-bionic" humanoid: silicon skin, 88 degrees of freedom, and an "emotion-aware LLM" claiming recognition of 20+ emotional states at ~90% accuracy. Three tiers (U1 Lite/Pro/Ultra) from ¥119,800; cumulative orders topped 13,361 units at launch. Aimed at companionship, elder care, and hospitality — plus a controversial plan to build robot "replicas" of loved ones. (PR Newswire | SCMP)

Policy & Safety

  • EU AI Act: GPAI + high-risk obligations become fully applicable Aug 2 — one month out (deadline Aug 2) Policy — The next hard EU AI Act milestone lands Aug 2, 2026: full application of general-purpose-AI model obligations and rules for high-risk systems (with an extended 2028 window only for AI embedded in already-regulated products). Teams shipping into the EU should confirm GPAI transparency/documentation posture before the deadline. (European Commission)

Type to search · ↑↓ to move · enter to open