This run's scan window was [2026-08-15T12:26:05Z, 2026-08-16T00:00:26Z] (Saturday afternoon to early Sunday UTC, with the retrieval overlap rolled back to 2026-08-15T09:26:05Z). A quiet weekend: arXiv paused announcements, major labs shipped nothing, and no new events met the threshold. Today's output is two items: one significant update to an existing event, and stronger evidence for one trend candidate.

Daily Executive Summary

  • DeepSeek peak/off-peak pricing details confirmed (update to ev-20260813-02): the official pricing documentation now provides the full rate table, effective 2026-08-16T16:00Z. This is not a price cut. During peak hours (01:00–04:00 and 06:00–10:00 UTC, i.e. Beijing business hours) V4-Pro costs $1.32 per Mtok input and $3.96 per Mtok output, roughly 3x/4.6x the current flat rate. Off-peak is half of peak but still 1.5–2.4x above current rates. Teams running heavy agent workloads need to redo their cost modeling, or move batch jobs to off-peak hours.
  • No new model, API, or tool releases in the window. Claude Code's latest version is still 2.1.233 (8/14, already recorded). GLM-5.3 has no new developments; the #3 spot on Product Hunt's 8/15 leaderboard is only a community-popularity signal. Kimi K3 (7/16), Meta Muse Glimmer / Muse Spark 1.2 weights (8/10), and Grok 4.6 / Grok Bot (8/11–12) are all pre-window old news and were not re-recorded.
  • Trend observation: the "MCP entering the enterprise security and governance phase" candidate gained several new verifiable pieces of evidence (Workday's official MCP server, a report tallying 40+ MCP CVEs) and was upgraded to emerging / Medium. See Trend Signals.
  • Papers submitted to arXiv over the weekend should appear in the Sunday/Monday announcement batch. The next scan will prioritize them.

Updates to Existing Events

Event Update Attribution
DeepSeek V4-Pro GA + Responses API (ev-20260813-02) The official pricing page confirms the full peak/off-peak rate table, effective 2026-08-16T16:00Z. Peak-hour definitions, cache hit/miss rates, and increase multiples are recorded in the event Confirmed in the 2026-08-16T00:00Z run

Models

No new releases. Still pending: GLM-5.3 open weights, expected ~2026-08-28 after the security review.

Agent & AI Engineering

No new events. One piece of background evidence (counts toward trends, not recorded as a standalone event): the Workday Adaptive Planning 2026R2 release notes, updated 8/14, add an official MCP Server that lets customer AI agents access Adaptive Planning data. Another first-party MCP server from an enterprise SaaS vendor.

Open Source

No new events. No significant new GitHub hotspots over the weekend. DeepSeek Harness (dsh) star growth continues; not re-sampled this run, so the 8/14 baseline carries over.

Research

arXiv announcements are paused over the weekend. Papers submitted 8/14–8/15 should appear in the Sunday/Monday batch. The next scan will prioritize them.

Developer Tools

  • Claude Code: the changelog confirms no new version in the 8/15–8/16 window; the latest is still 2.1.233 (8/14).
  • Cursor Builds turns on by default for all new and existing environments on 8/17, at no extra charge. See Watch Items.

Infrastructure

No new events. Cerebras SUPERNOVA 2026 (8/18, livestream) is coming up, and new inference products are expected.

Business & Policy

No qualifying events. Funding news such as xAI's $20B Series E is filtered out per the default rules.

Trend Signals

MCP entering the enterprise security and governance phase → upgraded to emerging (Medium confidence)

New verifiable evidence this run, independent of the 8/14 Cloudflare Gateway enforcement:

  • Workday Adaptive Planning official MCP Server (2026R2 release notes, updated 8/14). A supply-side signal of first-party MCP from enterprise SaaS.
  • Practical DevSecOps "MCP Security Statistics 2026": 82% of implementations carry path-traversal risk, and more than 40 MCP-related CVEs have been disclosed (as of early August). A demand-side signal: pressure from the security community.
  • Background: since the MCP 2026-07-28 spec revision, the ecosystem has passed 10,000 servers.

Upgrading to strengthening still requires the outstanding criterion: a second cybersecurity vendor (Zscaler / Netskope / Palo Alto class) shipping MCP identification capability. See trends/2026-08-16.md and trends/current.md.

The other candidates (Chinese labs' open-weight agentic coding models; coding agents converging on multi-agent runtimes) saw no new independent signals this cycle and remain unchanged.

Tech Radar

No changes. No new events this cycle, so the 8/13–8/14 radar carries over: Claude Code multi-agent ADOPT / Qwen3.8-27B, Gemini 3.7 Flash, Copilot Agent Plugins, Cloudflare MCP enforcement, DeepSeek V4-Pro TRIAL / GLM-5.3, DeepSeek Harness WATCH, etc.

Worth Trying

  • OpenAI published a GPT-5.6 builder's guide for startups. This was reported in the 8/15 AI Intelligence Briefing; the official primary link is pending verification in the next scan. Worth a look for teams building products on GPT-5.6.

Watch Items

  • Today 16:00Z: DeepSeek peak/off-peak rates take effect. The next run should confirm actual billing matches the documentation.
  • 8/17: Cursor Builds default rollout for everyone. Watch cloud agent startup behavior and failure-fallback performance.
  • 8/18: Cerebras SUPERNOVA 2026 (livestream). New inference products expected beyond Ultrafast.
  • ~8/28: GLM-5.3 open weights land (after the security review).
  • The arXiv weekend submission batch should appear in the Sunday/Monday announcement. Cover it in the next scan.

Sources