Friday, July 31, 2026

OpenAI cut GPT-5.6 Terra 20% and Luna 80% three weeks after launch — the clearest sign yet that frontier pricing power is eroding under competition. Google DeepMind's Gemini Robotics 2 pushed vision-language-action control from tabletop arms to whole-body humanoids. And the agent-security money kept flowing: Onyx Security raised $113M for an enterprise "control plane" that supervises other agents.

TLDR

OpenAI slashes GPT-5.6 prices — Terra −20% ($2/$12) and Luna −80% ($0.20/$1.20) per MTok, plus a new "Fast Mode" for Sol replacing Priority Processing (Jul 30) Pricing Trending
Google DeepMind unveils Gemini Robotics 2 — whole-body humanoid control plus a Gemini Robotics ER 2 reasoning/planning model, now on AI Studio (Jul 30) Robotics/IoT
Onyx Security raises $113M Series B at ~$640M to govern enterprise AI agents in real time via a "Guardian Agent" (Jul 29) Security Agents
Anthropic hit by two network failures on Jul 29–30 — elevated errors across claude.ai, the API, Claude Code, and Cowork (Jul 29–30) Production
EU AI Act GPAI + synthetic-media transparency rules become enforceable Aug 2 — fines up to €15M or 3% of global turnover (deadline Aug 2) Policy

Action Items

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

Re-benchmark high-volume workloads against the new GPT-5.6 Terra/Luna prices. Luna at $0.20/$1.20 per MTok changes the economics for classification, extraction, and bulk pipelines; also check whether "Fast Mode" on Sol (2.5x speed at 2x price) fits any latency-bound path. (OpenAI | CNBC)
Confirm EU AI Act readiness before Aug 2. If you serve EU users with general-purpose models or generate synthetic media, transparency obligations become enforceable Aug 2 with fines up to €15M or 3% of turnover — verify disclosure/labeling now. (Lumenova summary)
Review agent governance if you're deploying autonomous agents. With prompt-injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ still the top production failure mode and control-plane vendors (e.g., Onyx) scaling fast, make sure agent actions are observable and interruptible before you widen deployment. (SecurityWeek)

Models & Tools

Google DeepMind — Gemini Robotics 2 (Jul 30) Robotics/IoT — A new family: the VLA model converts vision + language into motor control for full humanoids ("feet to fingertips") and bi-arm robots; Gemini Robotics ER 2 is an embodied-reasoning VLM that plans multi-step, multi-minute tasks and is available now on Google AI Studio (private preview on the Gemini Enterprise Agent Platform); On-Device 2 runs locally with fast adaptation to new robot bodies. The headline jump is from tabletop upper-body tasks to intelligent whole-body control and multi-robot collaboration. (Google DeepMind | ER 2 post)

APIs & Pricing

OpenAI: GPT-5.6 Terra −20%, Luna −80%, plus "Fast Mode" (Jul 30) Pricing Trending — Terra drops to $2/$12 and Luna to $0.20/$1.20 per MTok (input/output), ~3 weeks after the family launched on Jul 9; Sol's list price is unchanged but gains a "Fast Mode" (up to 2.5x speed at 2x price) that replaces Priority Processing. OpenAI says it cut end-to-end serving costs ~20% and frames the move as answering cost-sensitive enterprise buyers and Chinese/Google/Microsoft price competition. (OpenAI | CNBC)

Industry & Business

Onyx Security raises $113M Series B at a ~$640M valuation (Jul 29) Security Agents — Bessemer-led round (Cyberstarts, TCV, Conviction, FirstMark and others) for an "AI control plane" that discovers, monitors, and remediates agent risk, with a "Guardian Agent" that watches other agents and intervenes in real time. The company says it secures 1.1M+ agents across 1.8M employees and quadrupled revenue since leaving stealth four months ago; Anthropic announced an integration in June. (SecurityWeek | Axios)
AMD secures $14B / 529MW of AI capacity from Core Scientific (Jul 28) Infra — 15-year leases to deploy AMD's newly unveiled MI455X GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ and Helios rack platform (announced Jul 22–23 at Advancing AI 2026), a direct challenge to Nvidia's data-center lead and another neocloud-style capacity land grab. (TechTimes)
Physical-AI megarounds: K2 Space ($500M Series D) and Antora Energy ($550M Series C) (Jul 30) Infra — Two nine-figure raises the same day underscore investor appetite for the hardware and power layers underneath AI (space infrastructure and thermal energy storage for data-center-adjacent demand). (Tech Startups roundup)
DataBahn raises $40M Series B for AI-era security data pipelines (Jul 30) Data Security — Insight Partners-led round (ForgePoint, S3 Ventures) for a platform that filters and routes telemetry before it hits SIEMs/data lakes — a relevant reference point if you're wrestling with log/observability cost as agent traffic grows. (Tech Startups roundup)
AI-cited engineering layoffs continue (updated Jul 25) Hiring — TechCrunch's running tracker added Monday.com to a 20+ company list citing AI in cuts this year; recent examples include Thomson Reuters (up to 500 engineering roles) and Sprout Social (260, ~20%). Counter-signal: several CEOs report a quiet rebound in entry-level AI hiring. (TechCrunch)

AI in Production

Anthropic: two network failures degrade Claude on Jul 29–30 (Jul 29–30) Production — Back-to-back incidents caused elevated errors and reduced availability across claude.ai, the API, Claude Code, and Cowork. Takeaway for teams building on Claude: this is the second provider-side reliability wobble in as many weeks (OpenAI had four disruptions Jul 22–25), so budget for multi-provider fallback and ret/idempotency on agent workflows. (Anthropic status)

Policy & Safety

EU AI Act: GPAI and synthetic-media transparency obligations enforceable Aug 2 (deadline Aug 2) Policy — Providers of general-purpose models and generators of synthetic media face binding transparency/labeling duties, with penalties up to €15M or 3% of global turnover for non-compliance. (Lumenova)
China's rules for "intelligent agents" take effect Jul 15 (Jul 15) Policy Agents — Implementation Opinions establish what's described as the first dedicated regulatory category for AI agents, with a three-tier decision-authorization framework and mandatory filing for high-risk sectors — worth noting if you ship agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ products into the China market. (Cubbbix summary)

Agent-tooling repos dominate late July Agents — Notable movers by momentum: Strix (open-source AI penetration-testing agent that behaves like a researcher rather than a static scanner), OpenWiki (a LangChain CLI that auto-generates and maintains AI-friendly codebase docs), and CubeSandbox (isolated execution for agents). The through-line: the trending frontier has shifted from models to agent infrastructure — sandboxing, orchestration, and security. (OSSInsight trending AI)

A price-and-physical-AI day.