I will pull primary-source updates from the last day or so and assemble tonight’s Europe/Berlin brief from those, not from secondary rumor.Aggregator scores disagree. I am checking lab pages, ARC Prize, and Artificial Analysis directly before writing the brief.**Intelligence Radar — Daily Brief, Tuesday 6 October 2026 (Europe/Berlin)**

Weather: Europe ships an open-weight preview trained on its own Blackwell cluster; the measurement story is still harness, not weights. Cyber access is the policy event of the day.

### Frontier and markets

Mistral Large 4 entered public API preview today. Confirmed. The lab post describes a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 49 billion active (docs page lists 1.05T total and 52B active, plus a 1.6B vision encoder). Context is 1M tokens. Preview prices on the docs page are about $0.68 per million input tokens, $0.07 cached input, and $2.09 output. Weights are promised by the end of October, not today. Training was on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters; the preview is served from the same stack, with a European deployment under European law. Self-reported coding: 61.7% DeepSWE v1.1, 28.3% Terminal-Bench 4, Coding Agent Index 49.8, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max on that composite. On visual grounding, Mistral claims Dense 200 at 42% versus GPT-6 Astra at 41%. Refusal-sensitive cyber scores are not comparable to closed models that decline the task. Tag: confirmed launch; self-reported scores until independent re-runs.

Google released EmbeddingGemma 2 under Apache 2.0. Confirmed. A modular sub-1B embedder (270M text/code up to 740M full multimodal) maps text, code, images, video, and audio into a shared 768-dimensional space, aimed at on-device retrieval rather than the agent frontier.

Amazon Nova 2.5 Sonic went generally available on 5 October. Confirmed. Speech-to-speech voice agents, 256K context, tool use, seven languages, lower latency. Not a general intelligence board.

No new flagship from OpenAI, Anthropic, DeepMind, xAI, Meta, DeepSeek, Qwen, Moonshot, or Z.ai today. Standing September board, named by harness, not a single rank:

- ARC Prize verified results: GPT-6.1 Sol (29 Sep) 98.5% / 94.2% / 96.37% on ARC-AGI-1 / 2 / 3 (v3 via provider-adapter harness; standard harness was 52.7% at higher cost). GPT-6 Astra (2 Sep) 98.5% / 95.0% / 99.95%. Claude Opus 5.5 (22 Sep) 98.5% / 93.3% on v1 and v2. Grok 4.7 (21 Sep) 90.2% / 61.4% / 10.05% adapter (1.8% standard). DeepSeek V4.1 Flash (10 Sep) 94.5% / 72.9% on v1 and v2, at $0.13 per ARC-AGI-2 task in ARC Prize’s 6 October note.
- Pricing still in force from primary posts: GPT-6.1 Sol $2 / $0.10 cached / $10 per million tokens, positioned as near-Astra agentic coding at about one-fifth Astra’s token price. Opus 5.5 $4 / $0.20 cache reads / $20, Fast mode $8 / $40. Gemini 4 Argon (30 Sep, Fairwind first, not general availability) introductory $2 / $10 with 95% cache discount and a 1M output limit; Google’s DeepSWE v1.1 claim is 77.9%.

Artificial Analysis versions must not be mixed. Secondary scrapes dated 6 October disagree: one cites Intelligence Index v4.3.2 with Gemini 4 Argon at 63.8 in an Antigravity CLI harness and GPT-6.1 Sol at 62.9 in Codex xhigh; another displays older self-report tables with GPT-5.6 Sol at 58.9. Neither is a ranker for this brief. Saturated ARC-AGI-1 (many models at 98.5%) is not a ranker.

### Science of intelligence

The 6 October measurement fact is scaffolding. ARC Prize’s community track for ARC-AGI-3 reached 55.89 RHAE on 4 October (Tufa Labs), nearly double the 27.9 Milestone 2 mark of 30 September, with little model change. The $700,000 human-level prize is unclaimed; final deadline is 2 November. Reported, from the competition write-up.

ARC Prize’s own verified card for Grok 4.7, recirculated today, shows reasoning budget as a causal variable: low effort averaged about 10k tokens per ARC-AGI-2 attempt and scored 25%; medium through xhigh used 86k–120k tokens and scored 57.5–60%, with a best of 61.4% at $2.01 per task — worse and more expensive than Grok 4.6’s 67.1% at $0.76 per task at the same token prices. Confirmed on the results page. The September OpenAI harness note remains the cleanest controlled result: retaining chain-of-thought between actions and compacting rather than truncating context moved GPT-5.6 Sol from 7.8% to 38.3% on ARC’s infrastructure. Provider-adapter versus standard harness still swings GPT-6.1 Sol on ARC-AGI-3 from 52.7% to 96.4%. Intelligence here is a system: model, memory policy, action budget, and verifier.

A 30 September arXiv report (Kepler, 2610.00834) claims a server-verified 100.00 RHAE on all 25 public ARC-AGI-3 games under one frozen Claude Opus 5 configuration by treating hypotheses as executable world models. The authors state replay scores are not first-exposure skill and may not meet the action-budget protocol. Reported, not a leaderboard entry.

### Engineering

Mistral’s post is the engineering document of the day. RL is asynchronous: tens of thousands of parallel rollouts, multi-compaction trajectories with million-token budgets, shared sandboxes and APIs, rewards mixed from unit tests, judges, and static checks. That is the agent stack, not a chat model. Cyber red-teaming with reduced moderation runs in parallel with the public preview — capability and policy are gated separately from the weights drop.

Anthropic expanded the Cyber Verification Program today. Confirmed. Three tiers (Defense, Red Team, Specialized) open Claude Opus 5.5, Sonnet 5.5, Mythos 5.1, and later models with reduced blocking classifiers to verified security teams. Reuters reports the prior Glasswing track helped surface more than 100,000 vulnerabilities this year. That is an availability event: the same weights, different refusal surface.

EmbeddingGemma 2 is the inference-economics item for agents that should not send every retrieval to a frontier API: modular encoders, on-device, Apache 2.0. Nova 2.5 Sonic is the voice-agent inference path, with Strands bidirectional agents now GA.

No new chip or energy primary today. The relevant hardware fact is already in the Mistral post: a European lab can train a 1T-class MoE on 3,800 Grace Blackwell GPUs it operates.

### Political, legal, social

The live EU regime is the August application, not a new text today. Official Commission pages: AI Act generally applicable since 2 August 2026, including Article 50 transparency (chatbot disclosure, synthetic-content marking). High-risk Chapter III was deferred by Regulation (EU) 2026/1744 (in force 27 July 2026) to 2 December 2027 for standalone Annex III systems and 2 August 2028 for AI embedded in Annex I products. GPAI obligations and AI Office enforcement powers are already live. Confirmed.

Today’s social fact is the split cyber market. Closed labs (Anthropic CVP, Google Fairwind for Argon) sell reduced-refusal access to vetted defenders. Mistral argues provider refusals are themselves an incident risk and that open weights plus self-hosting are the sovereignty response. Both are access-control designs, not capability claims.

### Watchlist

- Mistral Large 4 post and docs: https://mistral.ai/news/mistral-large-4/ — https://docs.mistral.ai/models/mistral-large-4
- Anthropic CVP: https://www.anthropic.com/news/cyber-verification-program
- EmbeddingGemma 2: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
- ARC Prize results and Grok 4.7 card: https://arcprize.org/results — https://arcprize.org/results/xai-grok-4-7
- Gemini 4 Argon: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- GPT-6.1 Sol: https://openai.com/index/introducing-gpt-6-1-sol/
- Opus 5.5: https://www.anthropic.com/claude/opus
- Nova 2.5 Sonic: https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-nova-2.5-sonic/
- EU AI Act timeline: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- Kepler note: https://arxiv.org/html/2610.00834

### Ledger actions

Do not move Frontier on Mistral Large 4 until weights ship and an independent harness re-runs DeepSWE v1.1, Terminal-Bench 4, and ARC-AGI-2. Preview API can enter Efficient only after cost-per-task is measured on a named board, not from list price. EmbeddingGemma 2 belongs on a retrieval atlas, not the five-benchmark agent atlas. ARC-AGI-3 must be stored as two columns (standard vs provider-adapter); a single number is not a label. Do not rebase the atlas to any Artificial Analysis index until the version string is locked (v4.1 weights from June are not v4.3.2). Saturated ARC-AGI-1 stays off the ranker. A Budget label changes only if cached-input economics on a long agent trajectory beat DeepSeek V4.1 Flash’s verified ARC-AGI-2 cost ($0.13/task), which no Western preview today has shown.

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