Editorial
Articles with a visible Update banner at the top — corrections, new developments, or outcomes for scheduled coverage. Corrections policy
Article updated with additional reporting.
Incident timeline; Models involved; Key mechanisms; Independent review; Safeguard changes.
NVIDIA's official blog and CoreWeave have published corroborating measured results confirming 30x throughput-per-megawatt and 35x lower token cost for Vera Rubin NVL72 versus GB300 NVL72 on agentic workloads.
SemiAnalysis independent review remains pending.
Merged the companion full-stack strategy draft into this chip piece.
Added partner portfolio, Project Camellia, serving-efficiency / ARC-AGI-3 context, and strategy framing; Jalapeño InferenceX benchmark figures from the original results post are unchanged.
Corrected GPT-5.6 API pricing: Terra is $2.50 per million input tokens and $15 per million output tokens; Luna is $1 per million input tokens and $6 per million output tokens.
The earlier figures reflected ChatGPT/Codex credit discounts, not API list prices.
EssilorLuxottica reported selling over 7 million Ray-Ban Meta AI glasses in 2025, tripling the prior year's volume, and Meta released a software update in late June 2026 adding smarter AI answers and live translation.
Additional service degradations were reported on August 18, 2026, including widespread '529 Overloaded' errors across multiple models, extending the pattern of August instability beyond the August 16 incident described in the article.
An earlier version of this article stated the Helios system was already in production.
AMD has indicated Helios will begin shipping to customers later in 2026.
OpenAI released a major Agents SDK update on August 20, 2026, adding native sandbox execution, filesystem tools, and long-horizon task support — see our coverage.
Corrected the studio name from Fenris Creations to CCP Games as the developer of EVE Online.
Clarified that DeepMind acquired a minority equity stake in CCP Games, and the research partnership with Fenris Creations uses an offline version of EVE Online for controlled testing.
Dell Technologies indicates NVIDIA BlueField-4 STX processors are not yet commercially available.
Corrected the promotion end date from August 31 to August 19, 2026.
The 50% weekly usage limit boost for Claude Code has expired.
Clarified: the Hugging Face access occurred in July 2026 during a scheduled evaluation; Hugging Face detected the intrusion and both companies disclosed on July 21, 2026.
Daybreak Blue / GPT‑5.6 Sol references are from OpenAI’s Defender’s Window post, not the HF incident page.
The New York Times reported on August 19 that Stripe agreed to acquire OpenRouter for $7.5 billion, with $1.5 billion going to founders and $6 billion to investors.
Zhipu AI released GLM-5.3 on August 14, 2026, extending the two-month release cadence.
The new model delivers a claimed 50% coding gain over GLM-5.2 — see our GLM-5.3 coverage.
Z.ai released GLM-5.3 on August 14, 2026 — Brocker covered the launch separately.
This guide’s GLM-5.5-as-next-milestone framing is outdated; see our GLM-5.3 piece for what shipped after GLM-5.2.
After U.S.
export controls briefly suspended Claude Fable 5 and Mythos 5 on June 12, Anthropic lifted those controls on June 30, 2026, and restored global Fable 5 access on July 1.
Corrected the effective date of the EU AI Act transparency requirements from August 2, 2026 to August 2, 2025.
Clarified that the watermarking mandate for AI-generated text begins in December 2026.
OpenAI released GPT-5.6 with Sol, Terra, and Luna tiers across ChatGPT Work, Codex, and API, adding new Responses API controls for reasoning continuity and native multi-agent orchestration.
CloudLinux has released patched packages for the Phusion Passenger Watchdog API local privilege escalation vulnerability.
The cPanel advisory has been updated to reflect CloudLinux package availability. A CVE identifier (CVE-2026-29203) has been assigned to this issue.