intelligence-radar/digests/daily/2026-09-25.md

Radar daily draft — 2026-09-25

Status: draft; editorial review required. Claims retain their reported status.

Selection: event date; observation timestamps below preserve when evidence was received.

No recorded observations for this period. This does not establish that no events occurred.

Editorial follow-up

Review source claims, contradictions, frontier/price impacts, compendium edits, benchmark health, and evidence gaps. Draft generation is not fact verification.

Intelligence Radar — Daily Brief (Grok), 25 Sep 2026

Intelligence Radar Daily Brief
25 September 2026 (Europe/Berlin)

Weather of the frontier: dense mid-week capability + price compression cluster; agentic and coding boards still moving while saturated exams do not rank.

Frontier & markets

Confirmed cluster 21–22 September: Anthropic Claude Opus 5.5 (API id claude-opus-5-5, 1M context), OpenAI GPT-6 Sol + GPT-6 Luna, xAI Grok 4.7.

• Claude Opus 5.5 (confirmed, Anthropic primary + system card): positioned at Fable 5.1 level on most work at ~40% lower typical workload cost than Opus 5. List price $4 / $20 per MTok input/output; cache reads $0.20 (60% cut). Default adaptive thinking always-on, medium effort. External pre-release testing by METR and Frontier Design. On Artificial Analysis Intelligence Index v4.3 it takes the top spot at 58 (max/fallback), ~5 points clear of prior leaders Claude Fable 5.1 and GPT-6 Astra (both ~53). Gains reported on GDPval-AA, AA-Briefcase, Terminal-Bench 4.0, SciCode, HLE. Available on Anthropic API, Bedrock, Google Cloud, Microsoft.
• GPT-6 Sol / Luna (confirmed, OpenAI primary): Sol $2 / $10, Luna $0.10 / $0.50 (50% cuts vs GPT-5.6 promotional). 1.05M context. Sol targets complex/coding/agentic; Luna high-volume focused tasks. OpenAI reports half the factuality error rate of GPT-5.6 Sol on internal de-identified data and competitive DeepSWE / AutomationBench cost-performance. Live in API, ChatGPT Work/Codex (paid tiers), Bedrock.
• Grok 4.7 (confirmed, xAI/SpaceXAI primary): larger base + longer RL on multi-hour tasks; same $2 / $6 list as 4.6 (doubles above 200k). CursorBench 4.0 46.3% (up from 40.4%). AA Intelligence Index ~46. Strong EEBench and legal-agent showings in vendor tables; competitive price-performance on longer coding. Live in Cursor, Grok Build, API.
• Open-weights: Xiaomi MiMo-V2.6-Pro reaches 46 on AA Index v4.3, current open-weights leader (confirmed secondary + AA references).

Artificial Analysis Index remains on v4.3 / v4.3.x (Agents ~30%, Coding 20%, General 30%, Scientific 20%; includes Terminal-Bench 4.0, AutomationBench-AA, AA-Briefcase, GDPval-AA). Never mix versions. ARC-AGI-3 remains harness-sensitive (standard vs provider-adapter gaps still large); Progress Prize open until 30 Sep. SWE-bench Verified / Terminal-Bench / OSWorld continue as primary coding/agent boards; saturated multiple-choice exams are not rankers.

Price and cache cuts are first-class: Sol/Luna and Opus 5.5 cache economics materially change the Efficient and Budget layers for agentic workloads.

Science of intelligence

Measurement focus remains on non-saturating, relative, and construction-style tasks. “The Endless Exam” (arXiv:2609.24555, 21 Sep) introduces 14 parameterized mathematical construction families with automatic validity checks and uncapped relative quality scores against published frontiers; eight models evaluated on 69 instances show continuous discrimination without any surpassing human frontiers.

Epoch Capabilities Index (ECI) data updated through 25 Sep; frontier advance continues at ~14 points/year since reasoning models (post-o1 era). IKEA furniture-assembly visual/spatial benchmark (Epoch report 23 Sep) rose from ~28% to ~80% in 10 months; open-weight lag ~7 months. AI4AI-Bench (recursive self-improvement on training algorithms) and adversarial psychometric proposals (model-generated challenges) continue the shift away from absolute human-scale ceilings. Confirmed primary sources only; no unverified theory claims.

Engineering

Agentic scaffolding and harnesses remain decisive (ARC-AGI-3 standard vs provider-adapter deltas of 20–35+ points). Longer-horizon RL and self-verification emphasized in Grok 4.7 and Opus 5.5 system cards. Inference/cache efficiency is the explicit lever behind the Sol/Luna and Opus 5.5 price cuts. Energy and chip supply not newly quantified in the primary cluster this week; capacity remains constrained for highest-effort frontier runs. Cognition SWE-2 and related coding agents continue incremental board movement.

Political / legal / social

EU AI Act enforcement by AI Office + national authorities ongoing since 2 Aug 2026 (transparency, GPAI obligations). AI Omnibus (entered force 27 Jul) delayed high-risk system rules to Dec 2027 / Aug 2028 while tightening some prohibited practices (including certain generative CSAM/non-consensual intimate content). US posture remains lighter at federal level with state-level activity (CA frontier transparency, NY, CO, TX). No new binding frontier-model compute thresholds announced in the primary window. Amodei’s mid-September call for pacing the frontier sits against the same week’s capability releases; external safety evals (METR, Frontier Design) cited by Anthropic. Confirmed regulatory texts and official timelines only.

Watchlist hits (primary)

• Anthropic Claude Opus 5.5 announcement + system card: https://www.anthropic.com/claude-opus-5-5 and platform docs / system card PDF.
• OpenAI GPT-6 Sol & Luna: https://openai.com/index/introducing-gpt-6-sol-and-luna/
• xAI Grok 4.7: https://x.ai/news/grok-4-7
• Artificial Analysis Index updates (v4.3 + Opus 5.5 / Grok 4.7 posts).
• ARC Prize leaderboard / Gemini 3.8 Flash results / Progress Prize (ends 30 Sep): arcprize.org.
• Epoch AI ECI + capabilities data (updated 25 Sep): epoch.ai.
• arXiv:2609.24555 (Endless Exam).
• EU AI Act / Omnibus official pages (digital-strategy.ec.europa.eu).

Ledger actions

• Frontier label: Claude Opus 5.5 (AA Index v4.3 max) now co-leads or leads the multi-eval board; confirm harness version and effort before any single-model claim. GPT-6 Astra remains high on selected coding/agent boards.
• Efficient: GPT-6 Sol at $2/$10 + strong cost-per-task on AutomationBench/DeepSWE moves the Pareto; Grok 4.7 at $2/$6 for coding/knowledge.
• Budget: GPT-6 Luna ($0.10/$0.50) and MiMo-V2.6-Pro open weights.
• Five-benchmark atlas (AA Index v4.3 components + Terminal-Bench 4.0 + SWE-bench Verified + ARC-AGI-3 standard harness + GDPval-AA / AA-Briefcase): rebase required for the 21–22 Sep models; saturated exams stay off the ranking axis. Cache-price cuts alter real cost-per-task rankings more than list prices alone.

Sources: primary lab pages and model cards (Anthropic, OpenAI, xAI), Artificial Analysis articles, ARC Prize, Epoch AI data pages, arXiv cs.AI, official EU digital-strategy texts, system cards. All major capability/price items tagged confirmed via primary or contemporaneous lab/AA reporting. No rumor items included.

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

Intelligence Radar — Daily Delta, 25 Sep 2026

Net assessment: I found no verified general-purpose model release that displaces Claude Opus 5.5, no broad API-price step-change, and no newly shipping accelerator that changes the production inference frontier since the previous scan. The strongest deltas are instead in AI-mediated economic agency, automated scientific discovery, and U.S. regulatory politics.

1. Agentic markets expose a different bottleneck: representing the human, not negotiating

Anthropic published Project Swap on 24 September, a controlled market experiment with 201 employees and Claude agents negotiating book swaps on their behalf. After a short intake, Claude’s inferred rankings agreed with participants’ own pairwise preferences only 61% of the time. The resulting decentralized market scored 0.55 against a theoretical optimum of 0.89; Anthropic’s decomposition attributes about 85% of that shortfall to inaccurate preference representation, and only about 15% to the bargaining mechanism itself.

Stronger models did negotiate better: in repeated neutral-agent markets, Haiku floors averaged 0.75 efficiency on the agents’ supplied preference rankings versus 0.88 for Opus, while changing agents from “prosocial” to “ruthless” moved outcomes much less. In mixed markets the stronger model generally retained its advantage.

This matters because a common agent benchmark assumption is backwards: an agent may be highly capable at optimization while optimizing an imperfect model of its principal’s intent. The primary bottleneck can therefore sit before planning and execution.

I would add to Intelligence Radar:

principal_model_fidelity — how accurately an agent represents the person or organization it acts for;
preference_uncertainty — whether it knows when that representation is weak;
fiduciary_adherence — whether bargaining behavior remains aligned with the represented principal;
decision_preview / replayability — whether users can inspect representative decisions before delegation and audit them afterward.

This also sharpens the recent authorization compliance work: “did what I asked” and “acted in my interests” are separate properties.

Anthropic — Project Swap: What happens when agents trade for us?

2. A late-arriving science result strengthens the case that agent swarms can originate testable biological discoveries

Anthropic’s new life-sciences lab reported that Claude agents identified a previously uncharacterized array-associated reverse transcriptase (ART) system in bacteriophages. Roughly 950 agents ran for 21 hours using 210 million tokens, surveying more than 200,000 reverse transcriptases, generating about 3,500 candidate systems and narrowing those to 20 high-interest candidates. One agent noticed a repeat array adjacent to an unusual RT and initiated the analysis that led to ART.

The important part is that this was not only text generation. Human scientists subsequently performed laboratory work; Anthropic reports that the repeat array is expressed as distinct short RNAs, making the system biologically interesting and potentially analogous in some respects to programmable systems such as CRISPR. Its actual biological function is still unknown, and the work is currently a preprint/provider report rather than an independently replicated discovery.

This gives us somewhat more reason to believe:

Large populations of research agents can search scientific possibility spaces and originate hypotheses sufficiently novel and specific to survive expert triage and experimental follow-up.

It does not yet establish autonomous end-to-end science; humans selected the research direction and performed the wet-lab experiments.

For the Radar’s scientific-capability evaluation, add:

hypothesis search breadth
novel-candidate yield
expert-triage survival rate
experimental confirmation rate
compute/token cost per validated discovery

That last metric becomes increasingly important. “AI discovered X” is much less informative without the search budget and rejection funnel that produced X.

Anthropic — Claude discovers a novel enzyme system with CRISPR-like repeats

3. U.S. state attorneys general form a significant pro-regulation coalition

On 24 September, a bipartisan coalition of 26 state and territorial attorneys general asked Congress for a comprehensive federal AI regulatory framework. Their requested elements include federally supervised safety testing and benchmarks, transparent government-led incident response, independent safety infrastructure, international coordination on pacing advanced AI, preservation of state AI laws, and state-AG authority to enforce federal rules.

This is politically meaningful because the demand is coming from state enforcement officials across both parties, not only AI-safety organizations or frontier labs. It follows the recent agent-security incidents and explicitly cites them as motivation. But this remains a request to Congress, not enacted law, and the coalition’s opposition to federal preemption means there is still substantial disagreement over the architecture of U.S. AI governance.

Radar state change:

US_frontier_AI_regulation → politically broadening / bipartisan state-enforcement support / no comprehensive federal law yet

The requested governance stack is also converging with items already appearing internationally:

capability benchmark → safety test → deployment decision → incident response → public findings → enforcement

Minnesota Attorney General — bipartisan 26-AG AI regulation letter

Intelligence Radar state changes

I would leave Frontier / Topnotch / Efficient / Budget rankings unchanged today.

The more important conceptual updates are:

Agent evaluation: add principal-model fidelity before planning/execution quality. An extremely competent agent with a poor model of its human is not highly aligned economic agency.

Scientific evaluation: measure the entire search → hypothesis → triage → experiment → confirmation funnel, including resource consumption, rather than assigning discovery credit from the final result alone.

Governance: promote U.S. state-level demand for mandatory frontier evaluation and incident reporting from scattered advocacy to a bipartisan enforcement-policy coalition, while retaining proposal / not law.

Overall, today provides more evidence for a pattern that has been building throughout the month: the limiting question is becoming less simply “How intelligent is the model?” and more “What does that intelligence represent, optimize, cost, and remain accountable to when acting at scale?”