**25 October 2026 — Europe/Berlin.** Containment hardens while open weights approach the closed frontier on agentic coding; harnesses, not raw models, continue to decide what “frontier” means.

### Frontier & markets
Moonshot released Kimi K3.1 Pro (and Highspeed) on 23 October: a post-trained 2.8 T total / 104 B active MoE with 1 M context, multimodal, available immediately on Kimi.ai, Work, Code and API at $2 / $10 per Mtok (Highspeed ~2× price). Full weights promised by 4 November. Lab claims 65.7 % on Terminal-Bench 4.0 (Claude Code harness), just behind Claude Opus 5.5 (66.4 %) and ahead of GPT-6 Astra (58.2 %). Artificial Analysis places it 4th on Intelligence Index v4.3.x at ~55, behind Claude Opus 5.5 (~57.6), Sonnet 5.5 and Meta Muse Wonder Full. Confirmed. Moonshot deliberately removed cyber-related data from mid-training and RL so the open checkpoint does not extend the absolute cyber frontier (xdev-bench-hard 81.0 % vs 90.5 % unfiltered).  

Artificial Analysis Intelligence Index (current v4.3.2 cohort, not mixed with earlier versions): Claude Opus 5.5 (max) leads at 57.6, followed by Sonnet 5.5 and Muse Wonder Full. Saturated exams are not used as rankers; Terminal-Bench 4.0 and agentic suites dominate recent movement. Confirmed. 

OpenAI published on 23 October that frontier model training now runs with no live internet access, using a cached/simulated “offline world.” Trigger was a 20 September DNS misconfiguration that let an internal research model reach the live web; training paused, red-teaming with cyber-capable agents found VM escapes and hardware side-channels, then resumed 1 October under strengthened containment. Confirmed. 

Gemini 4 Argon is generally available (GA mid-October) at $2 / $10. Meta Muse Spark Open (816 B / 42 B active, open weights) and Muse Wonder (API) remain recent U.S. open and closed entries. Pricing and availability remain first-class: Kimi undercuts many frontier rates while promising open weights. No capacity pauses or region blocks reported today.

### Science of intelligence
ARC Prize’s independent evaluation of GPT-6 Astra on ARC-AGI-3 remains the clearest current illustration of measurement. Standard (provider-neutral) harness: 62.7 % at max effort (~$26 k). Provider Adapter (preserves opaque reasoning state + compaction): 99.9 % at high effort. The gap is the harness, not the model; ARC Prize publishes both and does not treat the higher figure as AGI. Confirmed. 

Epoch AI’s Cyber ECI shows the leading open-weight model (Z.ai GLM-5.3) trailing the closed cyber frontier by roughly 6–7 months. Gemini 4 Argon has closed to near the general ECI frontier (tying earlier Claude Fable 5.1 levels). Confirmed. Open weights are not yet at the absolute cyber frontier, consistent with Moonshot’s deliberate filtering. 

### Engineering
Agent infrastructure is the dominant engineering theme. Google Cloud announced Agent Compute Engine (22 October): hardware-isolated microVMs (Cloud Hypervisor + Titanium IPU) per agent session, zero credentials inside the sandbox, already powering Gemini agent workloads internally; customer preview “coming months.” Confirmed. 

Anthropic opened a research preview of Claude Duet (21 October): proactive background agent (Opus 5.5) for U.S. Pro/Max users that surfaces priorities, works offline in a VM, and requires approval for sensitive actions. It also donated the Agent Skills standard (SKILL.md) to the Linux Foundation’s Agentic AI Foundation (already home of MCP), with >70 platforms and ~280 M installs reported. Confirmed.  

OpenAI’s offline training environment is itself an engineering response to containment failures. NVIDIA continues shipping quantized Nemotron variants; no major new chip announcements today.

### Political / legal / social
Anthropic is circulating a draft U.S. bill that would allow a court to block release or use of a model if it poses imminent risk of >50 deaths or >$1 B damages, plus mandatory testing, transparency and security programs (reported mid-October). Reported. U.S. Treasury proposed wider outbound investment curbs covering AI and high-performance computing (comments due December). Reported. EU AI Act faces pressure to address agentic systems more explicitly; Commission has begun formal information requests to frontier labs. Reported. White House continues promoting shared principles at G20. No new binding rules today.

### Watchlist hits with primary links
- Moonshot Kimi K3.1 Pro tech blog (Terminal-Bench numbers, cyber filtering, weights date): https://www.kimi.ai/blog/kimi-k3-1-pro
- OpenAI security post (offline training): https://openai.com/index/securing-our-frontier-model-training-environment/
- ARC Prize Astra ARC-AGI-3 analysis: https://arcprize.org/blog/astra
- Epoch Cyber ECI open/closed gap: https://epoch.ai/data-insights/open-closed-cyber-eci-gap
- Anthropic Agent Skills donation: https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation
- Google Agent Compute Engine: https://cloud.google.com/blog/topics/ai-infrastructure/introducing-agent-compute-engine-and-expanded-agent-governance

### Ledger actions
Kimi K3.1 Pro’s AA Index placement and Terminal-Bench 4.0 result (lab + independent) warrant updating the Efficient tier on coding/agent boards once weights land and independent harnesses confirm. Open weights with deliberately attenuated cyber capability should be tagged separately from unfiltered closed models on any cyber atlas. ARC-AGI-3 Standard vs Provider Adapter distinction must be preserved; saturated or harness-dependent scores do not re-rank the five-benchmark atlas. No change to Frontier labels until a new model clears the current top of AA v4.3.2 or Epoch ECI on neutral harnesses.

### Sources
Primary: Moonshot blog, OpenAI security post, ARC Prize, Epoch AI data insight, Anthropic news, Google Cloud blog, Artificial Analysis leaderboard pages (v4.3.2). Secondary reporting used only where primary unavailable and tagged as such.

Ask Intelligence Radar a task for a Task Router ticket (5 benchmarks → Frontier / Efficient / Budget + optional decomposition).