tegwick
8249c34100
IR-WP-0008 imports those drop-folder exports on the due-job timer and records that website automation runs still have no public feed. Assistant: grok Assistant-Session: 01a106a5-fc07-78c2-b64f-8e04cea25daf
52 lines
No EOL
8.5 KiB
Text
52 lines
No EOL
8.5 KiB
Text
I'll gather primary-source updates from labs, benchmarks, and regulators for the September 30, 2026 brief.**30 September 2026 — Europe/Berlin.** The frontier is not a single model; it is a price ladder. Closed labs shipped mid-tier systems that sit one point off the expensive flagship on composite boards while cutting list price four- to five-fold; open-weight Chinese MoEs now sit on the same cost–capability line; Terminal-Bench 2.1 is saturated and 4.0 is the live agent board; ARC-AGI-1 is saturated and ARC-AGI-3 is the measurement that still moves.
|
||
|
||
## Frontier & markets
|
||
**Confirmed.** Anthropic shipped Claude Sonnet 5.5 on 28 Sep: 1M context, $2 / $10 per million input/output, cache reads $0.20, adaptive thinking. Official claim: 30%+ faster output than Sonnet 5 and up to 30% lower cost per task at the same list price because it uses fewer tokens. Anthropic reports 70.6% on Terminal-Bench 4.0 versus 10.3% for Sonnet 5 (provider harness, not a third-party rebase). Opus 5.5 ($4 / $20, 22 Sep) remains the recommended high-judgment sibling; Fable 5.1 stays the expensive “when Opus is not enough” tier ($10 / $50).
|
||
|
||
**Confirmed.** OpenAI shipped GPT-6.1 Sol on 29 Sep, seven days after GPT-6 Sol. API: 1.05M context, listed around $2 / $10 (some trackers still show Pro at $2 / $10). Artificial Analysis notes GPT-6.1 Sol scoring one Index point below GPT-6 Astra at less than one-quarter Astra’s cost-per-task. Astra remains the premium ($10 / $50). Luna ($0.10 / $0.50) is the high-volume rung. Do not mix AA Intelligence Index versions: v4.2 (4 Sep) had Fable 5.1 then Astra on top; later aggregator snapshots of later composites put Opus 5.5 thinking near 57.6 and Sonnet 5.5 thinking near 56. Those are different harness mixes.
|
||
|
||
**Confirmed.** xAI Grok 4.7 (21 Sep): 500k context, $2 / $6. Xiaomi MiMo-V2.6-Pro (21 Sep, MIT, 1.02T / 42B active MoE, 1M context) is the open-weight move of the month on mid-Index / low output price. DeepSeek V4.1 Flash (10 Sep): $0.30 / $1.20 peak, cheaper off-peak. Google’s Gemini 3.8 Flash / Live / TTS line sits lower on current hard composites than mid-2026 snapshots implied—score drops often reflect Index rebases, not model regression.
|
||
|
||
**Reported.** Artificial Analysis Cyber Index: Grok 4.7 (xhigh) and MiMo-V2.6-Pro tied at 56; GPT-6 Luna at 53 and $0.12/task. High-refusal closed models (Sol, Astra, Opus 5.5, Fable 5.1) refuse 32–38% of vulnerability tasks. That is usable-capability, not coding IQ.
|
||
|
||
**Confirmed (availability).** OpenAI Dots (DevDay, 29 Sep) roll out to Pro / Business Premium / Enterprise in eligible markets **excluding EEA, Switzerland, and the UK**. That is a first-class availability event, not a rumor.
|
||
|
||
Epoch’s latest cost-of-thought work (reported 25–28 Sep): ~47% per quarter, ~13× per year at a fixed performance level since 2023; faster still at newly achieved SOTA. GPU spot floors (30 Sep): H100 on-demand listings from ~$1.30/GPU-h; B200 floors near $3.69. List token prices have not fallen as fast as capability-at-price.
|
||
|
||
## Science of intelligence
|
||
**Confirmed (papers, last 72h).** *What if automating AI R&D triggers an intelligence explosion?* (arXiv:2609.36054, 28 Sep) treats internal-code automation as already majority-complete at frontier labs and asks whether full R&D automation compresses years into months—benefits and loss-of-control both stated, not resolved. *From cacophony to hierarchy* (arXiv:2609.35618, 28 Sep) maps consciousness theories onto a five-level Marr-style stack and shows LLM credence ranging ~0.01–0.8 depending on prior, not evidence. *Failure-Transparent Agents* (arXiv:2609.35732) isolates post-failure reporting: baseline false-success 22.8%, structured evidence contract 0.8%. *Thinking Before Thinking* (arXiv:2609.38147, 29 Sep) reports meta-reasoning controllers beating matched direct-control agents on IMO ProofBench-Adv, ARC-AGI-2, LongCoT-mini, ProgramBench at fixed call budgets. *The Endless Exam* (arXiv:2609.24555) scores construction quality without any model beating published mathematical frontiers.
|
||
|
||
ARC-AGI-1 is saturated (~98.5% for Astra / Fable 5 / Gemini 3.8 Flash). ARC-AGI-2 still separates. ARC-AGI-3 is the live interactive board: GPT-6 Astra ~99.95% on the official table; GPT-6 Sol 4.6% standard harness vs 23.0% provider-adapter harness—same model, different scaffold. That swing is the measurement lesson. ARC Prize 2026’s final open-source milestone prize closes tonight.
|
||
|
||
SWE-bench Verified is archived-as-saturated in several labs (~96% Claude Opus 5, provider/self-report). Terminal-Bench 2.1 is clustered 88–93% (SWE-2 Cognition 92.8% on one aggregator). Terminal-Bench 4.0 (66 new tasks, Harbor / Laude) is the current agentic computer-use exam: Vals snapshot ~65% Opus 5.5 / ~64% Sonnet 5.5 / ~60% Astra; hardware category mean ~15%. Name the version or do not rank.
|
||
|
||
## Engineering
|
||
**Confirmed.** OpenAI Dots: always-on agents on GPT-6 Astra with a dedicated cloud computer and browser, 4,000+ app connectors, Slack/Teams, consent gates on sensitive actions. Agents API public beta (hosted execution, memory, multi-agent). Ultrafast paid speed tier (claimed up to 8× Codex, 6× API). Pro 500 tier announced. This is productization of long-running tool loops, not a new intelligence theory.
|
||
|
||
**Confirmed.** DeepSeek + Huawei (30 Sep, WeChat primary): open-source compute/comms libraries and TileLang backend for Ascend 950, plus a 128-chip supernode stack. Explicit aim: a CUDA alternative for an independent GPU software ecosystem. This is the software layer of the export-control story.
|
||
|
||
Energy/compute: Epoch and SemiAnalysis still put MW-year ownership in the ~$8.5–10M band; token list revenue per H100-hour remains a large multiple of rental. Capacity is still the constraint on Astra-class and Fable-class inference, not theory.
|
||
|
||
## Political / legal / social
|
||
**Confirmed.** EU Digital Omnibus (Reg. 2026/1744) is in force. Annex III high-risk duties: **2 Dec 2027**. Annex I product-embedded: **2 Aug 2028**. Article 50 transparency for interactive systems is already live (since 2 Aug 2026); machine-readable marking for pre-2 Aug generative systems: **2 Dec 2026**. New prohibited practices (NCII and CSAM generation): **2 Dec 2026**. EN 18286 (QMS for Art. 17) is published; OJEU citation pending. GPAI obligations and prohibited-practice enforcement are not deferred.
|
||
|
||
Dots’ EEA/UK/CH hold is consistent with that stack. Anthropic’s Sonnet 5.5 is the first Sonnet with Opus-class cyber fallbacks—policy product, not just a score.
|
||
|
||
## Watchlist
|
||
- Anthropic: https://www.anthropic.com/claude-sonnet-5-5
|
||
- OpenAI DevDay / Dots (Reuters + community notes on geo-block): https://www.reuters.com/business/openai-takes-on-meta-with-always-on-dots-agent-enterprise-ai-push-2026-09-29/
|
||
- AA Index v4.2 note + later snapshots: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-2
|
||
- Terminal-Bench 4.0: https://www.vals.ai/benchmarks/terminal-bench-4
|
||
- ARC results: https://arcprize.org/results
|
||
- Epoch cost-of-thought: https://epochai.substack.com/p/how-fast-is-ai-getting-cheaper
|
||
- DeepSeek/Huawei tools: WeChat primary; English wrap https://www.straitstimes.com/business/deepseek-partners-with-huawei-to-develop-programming-tools-reducing-reliance-on-nvidia
|
||
- Explosion paper: https://arxiv.org/abs/2609.36054
|
||
- Council AI Act timeline: https://www.consilium.europa.eu/en/policies/artificial-intelligence-act/timeline-artificial-intelligence/
|
||
|
||
## Ledger actions
|
||
Re-run **Terminal-Bench 4.0** (not 2.1) and **ARC-AGI-3** (both harnesses) before changing Frontier / Efficient / Budget labels. A one-point AA Index move after a version bump is not a rebase of the five-benchmark atlas. GPT-6.1 Sol and Sonnet 5.5 are Efficient-class candidates on coding/computer-use if independent 4.0 and SWE-Pro holds match provider claims; Astra / Fable 5.1 / Opus 5.5 stay Frontier on high-refusal, long-horizon, and adaptive-judgment boards. MiMo-V2.6-Pro and DeepSeek V4.1 Flash stay Budget/Efficient open-weight until a held-out 4.0 run. Dots geo-block does not change model labels; it changes availability of the agent product. Cyber Index refusals should be a separate column, not folded into Intelligence Index.
|
||
|
||
## Sources
|
||
Primary and near-primary URLs are listed in Watchlist. Aggregators used only to locate dates and prices: Artificial Analysis, BenchLM, Vals, ARC Prize, Epoch, Anthropic/OpenAI product pages, Council/Kinstellar on the Omnibus, NYT/Bloomberg/Straits Times on DeepSeek–Huawei.
|
||
|
||
Ask Intelligence Radar a task for a Task Router ticket (5 benchmarks → Frontier / Efficient / Budget + optional decomposition). |