I'll assemble tonight's Intelligence Radar brief for 7 October 2026 (Europe/Berlin), checking primary lab, benchmark, and regulator sources from the last day or so.

Aggregator pages disagree on scores and even on which models exist. Checking lab blogs and Artificial Analysis directly.

**7 October 2026 (Europe/Berlin).** Weather: the week’s real moves are an open-weight European preview that lands mid-board, a gated cyber-access product, and EU text-provenance starting to bite — not a new closed frontier model.

## Frontier and markets

Mistral opened a public preview of Mistral Large 4 (“Le Chonk”) on 6 October. The lab’s post describes a natively multimodal mixture-of-experts model, about 1 trillion total parameters and 49 billion active, trained from scratch on roughly 3,800 NVIDIA Grace Blackwell GPUs in European datacenters. Weights are promised by the end of October. The preview API is live on Mistral Studio; docs list a 1M context and promotional prices of about $0.68 / $2.09 per million input/output tokens (list $1.36 / $4.18), with cached input near $0.07. Tag: confirmed, primary. Artificial Analysis, same day, places the preview at 38 on its Intelligence Index — level with GPT-6 Luna (max, 38) and just under DeepSeek V4.1 Flash (max, 39) — and at 50 on its Cyber Index, level with GLM-5.3-Flash. AA calls it the strongest model from outside the US and China on that board. Tag: confirmed, third-party eval of a preview, not open weights yet. Do not mix that 38 with older Index versions. On AA’s current chart as reported by secondary summaries of v4.3.2, closed leaders sit far higher (Claude Opus 5.5 cited near 58; a cluster including Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon near 53). Large 4 is an Efficient/open-sovereignty event, not a Frontier-board displacement.

Google shipped two availability events on 6 October. EmbeddingGemma 2 is a 740M open embedding model (Apache 2.0, Gemma 4 lineage) that maps text, code, image, video, and audio into one space, with an 8K context and a modular text-only path. Google reports ~191 MB active RAM text-only and ~567 MB full multimodal on a Pixel 11 Pro after quantization, and a +9.92 point gain on MTEB Code versus EmbeddingGemma. Tag: confirmed, lab blog; quality claims are self-reported. Nano Banana 2.1 went generally available on the Gemini API, app, Search AI Mode, and Ads, replacing Nano Banana 2. Reported standard-tier image output is about half the prior token price ($30 vs $60 per million image-output tokens; ~$0.034 per 1K image vs ~$0.067). Input and text/thinking output prices rose ($1.50 and $7.50 per million vs $0.50 and $3). Tag: reported from Google’s pricing page and changelog; independent image evals were not public at launch. Google’s own infographic-factuality figure (0.521 vs 0.179) is vendor-only.

No new closed flagship shipped on 7 October. The late-September hold on GPT-6.1 Astra remains the relevant OpenAI constraint: The Register reported on 29 September that OpenAI confirmed the planned October ChatGPT/Codex release was benched after the model scored worse than GPT-6 Astra on alignment evals (persistence without a stop condition). Tag: confirmed by the lab to a reporter; not a fresh 7 October event. GPT-6 Astra, GPT-6.1 Sol, Sol, and Luna remain the available OpenAI workhorses. xAI’s grok-voice-transcribe-1.0 was retired 2 October in favor of 2.0 (reported, model tracker).

SWE-bench Verified is not a ranker here. Epoch’s 3 September 2026 review marks it Flawed: OpenAI’s February audit found a large share of tasks with scoring defects. Aggregator boards for “Verified” disagree with each other and with that verdict. SWE-bench Pro is the less saturated coding board in circulation; BenchLM’s 7 October snapshot has Claude Opus 5.5 at 89.9%, Sonnet 5.5 at 81.3%, Fable 5.1 at 81.2%. Tag: reported, harness unspecified beyond the site’s board. Saturated exams stay off the atlas.

## Science of intelligence

The useful measurement note this week is not a product score. Kepler (arXiv:2610.00834, 30 September, accepted to a non-archival NeurIPS 2026 workshop) treats ARC-AGI-3 hypotheses as executable world models checked by retrospective transitions and conditional predictions. Under one frozen Claude Opus 5 setup it reports a server-verified 100.00 RHAE on all 25 public games, with no per-game model pick. The authors’ own conclusion is the item: public-set score has limited discriminative value; first-attempt, cost-conditioned, and verification-aware reporting is what still separates systems. Tag: confirmed preprint, public set only.

A companion limit on latent world models (arXiv:2609.39235, 30 September) finds five-step rollouts rank actions reliably only about 5–10 control steps ahead, while task goals sit 16–53 steps out. Scaling the predictor did not extend the range; even a perfect simulator loses success as the horizon grows. Planning needs longer imagined trajectories or nearer subgoals. Tag: confirmed preprint. Together the two papers say interactive-intelligence scores without an audit trail, and imagination without a horizon budget, are both easy to over-read.

## Engineering

Anthropic expanded the Cyber Verification Program on 6 October into three tiers — Defense, Red Team, Specialized — covering Opus 5.5, Sonnet 5.5, and Mythos 5.1, with lighter blocking classifiers for vetted security teams. On CyScenarioBench (10 challenges × 5 attempts, Opus 5.5), default safeguards blocked all 50 trials at the first prompt; Defense Access blocked 46 and let 4 succeed; Red Team Access matched the unsafeguarded 67.6% (34/50). Specialized Access is reserved for organizations authorized to test critical infrastructure. Tag: confirmed, lab post. This is containment as a product surface: capability is not the scarce object, authorized tool-use is.

Mistral’s training footprint is the hardware footnote: 3,800 Grace Blackwell GPUs, European sites, same stack serving the preview. NVIDIA’s Vera Rubin efficiency claims (about 5× peak inference vs Grace Blackwell at 2× system power; 45°C coolant, no chiller, per Jensen Huang at CES) are vendor architecture marketing recirculated this week, not a measured token-price change. Tag: reported, CEO remarks.

## Political, legal, social

Article 50 transparency is already in force (2 August 2026). High-risk Annex III duties were pushed to 2 December 2027 by Regulation (EU) 2026/1744. Next hard date is 2 December 2026: machine-readable marking for generative systems already on the market, plus new Article 5 bans on non-consensual intimate imagery and CSAM generation. Tag: confirmed, Official Journal timeline as restated in EU-focused explainers this week.

OpenAI’s 5 October post is the operational response: textGrain, a statistical watermark in token choices, opt-in globally on the API for selected models, and an EU-only rollout over the coming weeks for eligible ChatGPT and Codex text. The watermark does not identify the user. OpenAI’s own tests: synonym replacement of 10% of words drops detection from about 92% to 66%; 25% drops it to 17%. Tag: confirmed, lab post. A provenance signal this brittle is a compliance artifact, not an attribution system.

The June episode in which a nonpublic Commerce directive briefly treated Anthropic frontier access as a deemed export — and Anthropic suspended the models for everyone — remains the legal precedent commentators are still unpacking. Controls were lifted 30 June. Tag: reported, not a new order.

## Watchlist

- Mistral Large 4 weights, end of October: https://mistral.ai/news/mistral-large-4/ — if the drop matches the preview, open-weight Efficient moves; if it regresses, the preview was a serving-time artifact.
- AA model page for Large 4: https://artificialanalysis.ai/articles/mistral-large-4-france-ai — re-score after weights, same Index version only.
- Anthropic CVP tiers: https://www.anthropic.com/news/cyber-verification-program — whether Specialized Access stays a handful of orgs.
- OpenAI textGrain: https://openai.com/index/eu-text-provenance/ — detector access for researchers, and whether paraphrase attacks stay this cheap.
- EmbeddingGemma 2: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
- Kepler: https://arxiv.org/abs/2610.00834
- Epoch on SWE-bench Verified: https://epoch.ai/benchmarks/swe-bench-verified/review

## Ledger actions

No Frontier label change. On Artificial Analysis Intelligence Index v4.3.2 as cited this week, the closed cluster (Opus 5.5, Fable 5.1, GPT-6 Astra, Gemini 4 Argon) still defines the top of that board; Large 4 at 38 does not enter it. Efficient: add Mistral Large 4 Preview as a European open-weight candidate once weights land and AA re-runs the same Index; until then it is API-only and proprietary on AA’s listing. Budget: Nano Banana 2.1 cuts image-output price roughly in half at 1K/2K but raises input and text-output prices — net Budget only for image-heavy, low-input jobs. GPT-6 Luna remains the cheap text configuration on cost-per-suite boards; do not rebase that from image prices. Atlas: keep SWE-bench Verified off the five-benchmark set (Epoch: Flawed). Prefer Terminal-Bench Hard or SWE-bench Pro with harness named, AA Intelligence Index version pinned, a cyber board (AA Cyber or CyScenarioBench with tier), an interactive board that reports first-attempt and cost (ARC-AGI-3 / Kepler-style), and a retrieval or embedding check that is not a saturated knowledge exam. A public-set 100 on ARC-AGI-3 does not move the interactive slot by itself.

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