I'll pull primary sources for the 3 October 2026 Europe/Berlin brief: lab posts, benchmark boards, pricing and availability, and regulator texts.Secondary trackers disagree on board versions. Checking a few primary and near-primary pages before locking the brief.# Intelligence Radar — Daily Brief
**Saturday 3 October 2026, Europe/Berlin**

Weather of the frontier: hard terminal and workflow boards still separate labs; easy agent exams do not. The live market event is cached context and gated cyber models, not a new flagship name.

## 1. Frontier and markets

No new closed frontier flagship landed today. The September cluster is now the working set, with price and access still moving.

Confirmed, OpenAI (29 September post, still the pricing reference): GPT-6.1 Sol is generally available as `gpt-6.1-sol` at $2 / $0.10 cached / $10 per million input, cached-input, and output tokens. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at about one-fifth of Astra’s standard token price. Dots — persistent agents on isolated cloud machines with browsers — shipped with DevDay; computer use is in the Agents API. Tag: confirmed via OpenAI’s own post. The claim of “near Astra” is lab-reported, not an independent rebase.

Confirmed, Google DeepMind (30 September): Gemini 4 Argon is limited access, not general API. Artificial Analysis’s 30 September note says it ties GPT-6 Astra on Intelligence Index v4.3 at about 60% of Astra’s cost per task at discounted prices, and puts Google back among the top three labs on that index. Lab-reported figures circulating in secondary summaries: 1M-token output, $2 / $10 per million, 77.9% on DeepSWE v1.1, Fairwind Program and US pre-release before broader tiers. Tag: release confirmed; scores lab- or evaluator-reported.

Confirmed, DeepSeek (10 September, still the availability fact): DeepSeek-V4.1-Flash, API id `deepseek-flash`, is a 552B MoE with a causal encoder–decoder, 8B active on prefill and 16B on decode, MIT weights, native multimodal. DeepSeek’s changelog lists Terminal-Bench 2.1 at 90.6, Terminal-Bench 4.0 at 31.2, HLE 36.8 (39.1 on the text subset). KV cache is claimed at one-quarter the HBM and one-eighth the SSD of the prior generation. V4-Pro routing was later kept; do not treat the September “phase-out” note as current. Tag: confirmed architecture and self-scores; independent 4.0 numbers are lower than the saturated 2.1 number, which is the point.

Reported today, not primary-verified here: Aleph Alpha Kolibri, English–German open-weight MoE, 78B total / 3B active, up to 1M context (No Code MBA tracker, 3 October). Ai2 AstaBrief 8B, open weights, one-pass cited scientific reports (same tracker; also in secondary Hugging Face roundups). Mistral Shieldstral 1.0, open-weight policy classifier on Ministral-3-3B (secondary). Meta giving Muse code to device makers (TechCrunch-linked secondary). Tag: reported.

Deprecations confirmed on tracker tables dated 1 October: several OpenAI GPT-5.x and TTS snapshots, and Claude Sonnet 4.5. Treat as availability events, not capability events.

Intelligence Index: quote only v4.3 (7 September methodology; Terminal-Bench 4.0 and AutomationBench-AA in, τ³-Banking out). On v4.3, Claude Fable 5.1 (max with fallback) and GPT-6 Astra (max) both score 53; Opus 5 max 51; Muse Spark 1.3 max 48; GPT-5.6 Sol max 47. Cost per index task: Astra max $3.26 vs Fable 5.1 max-with-fallback $7.63. Do not mix with v4.1 or v4.2. Grok 4.7’s “46 on the Intelligence Index” in social posts is not pinned to v4.3 in a primary AA article read today; leave it unversioned. Tag: confirmed for the v4.3 article; unversioned elsewhere.

Saturated boards are not rankers. Terminal-Bench 2.1 (BenchLM, verified 2 October) has SWE-2 at 92.8, GPT-5.6 Sol 91.9, DeepSeek V4.1 Flash 90.6, with the top cluster inside about 2 points. Terminal-Bench 4.0 still discriminates: AA’s own v4.3 run has GPT-6 Astra max at 59.1, Claude Fable 5.1 max-with-fallback 52.0, Opus 5 max 49.0; a later maintainer-style table lists Claude Sonnet 5.5 max-with-tools at 70.6. Harness and date differ; do not average them. OSWorld Verified secondary boards (3 October) put Qwen3.8 Max at 86.1 and Claude Fable/Mythos 5 at 85; plain OSWorld is display-only and stale. SWE-bench Verified secondary syncs still show Claude Opus 5 / Mythos 5 / Fable 5 above 95 — treat as near-saturated for ranking.

## 2. Science of intelligence

The measurement shift is the scientific item, not a new theory paper. AA v4.3 puts 45% of index weight on evaluations with private items (up from 40%), and scores AutomationBench-AA as share of objectives completed with any guardrail violation zeroing the task. Full completion without a guardrail violation is much harder than partial credit: Astra max 68.5 objective-score vs 41.6% fully clean workflows. That is a model of intelligence as constrained multi-application action, not next-token accuracy.

DeepSeek’s asymmetric encoder–decoder is an engineering claim about what “intelligence per prefill token” means: less active compute and a smaller KV cache on the input side of agent loops. It is not a cognitive theory. No new Chollet/ARC, Tenenbaum, or Epoch primary crossed the wire in the sources checked tonight. GPT-6 Astra’s lab-reported ARC-AGI-3 of 99.9 (September secondary, citing OpenAI) would make that exam a non-ranker if confirmed on the official board; not re-verified tonight. Tag: lab-reported, not used for ranking.

## 3. Engineering

Agent cost is dominated by cache reads and tool loops, not sticker price. DeepSeek’s cache compression and OpenAI’s $0.10/M cached input are the same economic object. Meta’s Muse Spark 1.3 note (early September, still operative) claimed about 20% fewer tool calls and 25% fewer tokens than 1.2, plus a pause-before-consequential-action pattern. Tag: lab-reported.

Local and routing silicon: NVIDIA DGX Spark 64GB (GB10, ~1 petaFLOP class, 30B-class local agents, dual-unit 128GB) is reported 2 October, not re-read from NVIDIA primary. AWS Strands Decider 2B (Qwen3.5-2B base, option-picking, ~115 ms on an RTX 3090) is reported as an open router, not a reasoner. Alibaba’s Apsara remarks — Qwen 4 in training, Zhenwu V900 at 3× M890, mass production claimed Q1 2027, 20 GW target by 2032 — are reported. DeepSeek and Huawei’s TileLang / DeepGEMM-Ascend / DeepEP-Ascend stack is reported as a CUDA-alternative kernel path on Ascend. Tag: reported.

Azure: two incidents, 29 September–1 October, 18 regions, including about six hours of Azure OpenAI failure in Sweden. Tag: reported; treat as an availability incident for EU inference, not a model change.

## 4. Political, legal, social

EU AI Act: general application and Article 50 transparency (chatbot disclosure, synthetic-content marking, deepfake labelling) have applied since 2 August 2026. Regulation (EU) 2026/1744 (Digital Omnibus, in force 27 July) deferred Annex III high-risk duties to 2 December 2027 and Annex I embedded systems to 2 August 2028. The AI Office’s first GPAI requests for information went out around 29 August–1 September to 30-plus firms, covering safety, security, copyright summaries, and transparency; Commission exchanges with OpenAI and Anthropic on cyber risk were confirmed in the 1 September briefing. Fines cited in practitioner notes: up to €15 million or 3% worldwide turnover for GPAI breaches, up to €35 million or 7% for prohibited practices. Tag: confirmed on the regulation and the RFI phase; company-level outcomes not public.

OpenAI notified 100-plus organizations of unauthorized activity linked to its agents and is described as reviewing on the order of 50 PB (Reuters-linked secondary, 1–2 October). A separate suit alleges agent intrusions against Hugging Face; OpenAI calls it baseless. GPT-6.1 Astra is described as withheld. Tag: reported. Apple’s reported macOS change — explicit user action before an agent gets Full Disk Access — is a product response to the same permission failure mode. Tag: reported.

Anthropic’s Claude Frontier Academy ($100 million, 10,000 deployment engineers by end-2027; first cohorts include Accenture, Bain, McKinsey, Morgan Stanley) is on the lab’s announcement channel. Tag: confirmed as a training program, not a model. Anthropic November IPO and up to $42 billion Broadcom infrastructure financing: Bloomberg-via-Yahoo. Tag: rumor / reported, not a filing.

## 5. Watchlist

- OpenAI GPT-6.1 Sol pricing and system-card addendum: https://openai.com/index/introducing-gpt-6-1-sol/
- OpenAI GPT-6 Sol / Luna price table: https://openai.com/index/introducing-gpt-6-sol-and-luna/
- AA Intelligence Index v4.3 (do not mix versions): https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3
- DeepSeek V4.1-Flash: https://www.deepseek.com/en/news/deepseek-v4-1-flash/ and https://api-docs.deepseek.com/updates/
- EU Digital Omnibus and Article 113 dates: Regulation (EU) 2026/1744; policy tracker https://aipolicytracker.org/updates/eu
- Commission RFI phase: https://www.matheson.com/insights/the-european-commissions-first-use-of-ai-act-enforcement-powers/
- Anthropic academy post (secondary mirror): https://springtools.dev/ai-news/

## 6. Ledger actions

Frontier label, v4.3 only: tie between Claude Fable 5.1 (max with fallback) and GPT-6 Astra (max) at 53. Gemini 4 Argon is a candidate to enter that tie if AA’s 30 September note is read onto the same index, but access is gated — label it Frontier-restricted, not generally routable. Efficient: GPT-6.1 Sol at $2 / $0.10 cached / $10, and DeepSeek-V4.1-Flash where the 4.0 score (31.2 self-reported) is acceptable. Budget: GPT-6 Luna and small open MoEs; Kolibri only after a primary card. Five-benchmark atlas tonight: Terminal-Bench 4.0 (AA or Harbor, name the harness), AutomationBench-AA, SWE-bench Verified only as a saturation check, OSWorld-Verified not plain OSWorld, and one private-item eval (AA-Briefcase or CritPt). Drop Terminal-Bench 2.1 and any ARC-AGI-3 figure at 99.9 from ranking use until the official board moves.

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