intelligence-radar/digests/daily/2026-09-20.md

Intelligence Radar Daily 2026 09 20

Intelligence Radar — Daily Brief (Grok), 20 Sep 2026

20 September 2026 (Europe/Berlin). Frontier weather: post-Labor-Day cluster of releases has produced a tight dual leadership on hard agentic/coding boards, with cyber specialization and cost-per-task divergence now the binding constraints; safety coordination talk is rising faster than new pure capability jumps.

Frontier & markets

Confirmed releases dominate the last three weeks. Anthropic shipped Claude Fable 5.1 / Mythos 5.1 (1 Sep, system card published); Fable is the general-access configuration with dual-use safeguards, Mythos the more permissive trusted-access variant for cyber and life-sciences work. OpenAI released GPT-6 Astra (3 Sep, system card); it is the first OpenAI model to hit Critical cybersecurity under its Preparedness Framework. Google released Gemini 3.8 Flash and the specialized 3.8 Flash Cyber (≈2 Sep), plus Gemini 3.8 Live / Live Extended Thinking voice models (15 Sep). DeepSeek launched V4.1-Flash (10 Sep), an asymmetric MoE (≈552B total, low active parameters) that the lab claims outperforms its own prior Pro on multiple agent/coding suites while retiring older Flash/Pro routes. Meta’s Muse Spark 1.3 and various Chinese flash tiers (GLM-5.3-Flash, Qwen3.8 variants, Step 5 Preview) also appeared.

On Artificial Analysis Intelligence Index v4.3 (not earlier versions; Terminal-Bench upgraded to 4.0 + new AutomationBench-AA private set), Claude Fable 5.1 (max + fallback) and GPT-6 Astra (max) are tied at 53. Claude Opus 5 sits at 51, Muse Spark 1.3 at 48, GPT-5.6 Sol at 47. Open-weight leaders remain GLM-5.3-Flash (≈42) and Qwen3.8-class models. Astra leads Terminal-Bench 4.0 (≈59 %) and AutomationBench-AA; Fable is stronger on several knowledge-work and scientific slices. Cost-per-task Pareto is dominated by OpenAI configurations; Astra delivers the tied 53 score at substantially lower average task cost than Fable.

Pricing and availability are first-class. Astra is listed near $10 input / $50 output per million tokens (cached input heavily discounted); Fable/Mythos sit in a comparable high band. DeepSeek V4.1-Flash and multiple Chinese Flash tiers remain an order of magnitude cheaper. No major capacity pauses or region blocks reported in the last 48 h; cyber-gated access (OpenAI Daybreak Blue, Anthropic Mythos trusted programs, Google Fairwind) continues to expand for vetted defenders. Saturated exams are ignored as rankers; Terminal-Bench 4.0, OSWorld-Verified, SWE-bench variants, and AA’s private-set components are the operative harnesses.

Science of intelligence

Measurement and cognitive structure, not product claims. A May 2026 arXiv paper proposes a Cognitive Taxonomy of 10 faculties drawn from psychology/neuroscience for AGI progress tracking, with held-out cognitive-task profiles. September papers examine systematicity of reasoning models on rule-induction isomorphisms (models often fail structurally equivalent variants despite solving the base task) and whether agents can autonomously run SAE-based mechanistic interpretability research (SAEScientist-Bench; frontier agents show partial discovery but lag expert baselines on causal steering). Stanford’s AIMS group released an updated “AI Measurement Science” textbook emphasizing validity, estimation, reliability, and the inference nature of evaluation rather than raw leaderboard scores. Epoch continues tracking rapid rise in AI acknowledgments in arXiv math preprints (≈25 % by August) and maintains the Epoch Capabilities Index, where Astra currently leads overall and on math slices while Fable retains an edge on software-engineering slices.

Engineering

Agent harnesses and infrastructure remain the rate-limiting factors. Terminal-Bench 4.0 (harder terminal tasks across software/ML/science/ops/security) and AutomationBench-AA (Zapier-derived business workflows with private set and guardrail penalties) are now inside the AA Index; full end-to-end workflow completion without guardrail trips is still substantially harder than partial objective completion. SWE-bench and OSWorld-Verified continue to show strong open-weight Chinese models (Qwen3.8 Max leading some OSWorld snapshots) alongside closed frontier agents. Inference economics favor high-throughput Flash/MoE designs; DeepSeek’s causal encoder–decoder and low active-parameter counts are explicit responses to cost and latency. Epoch’s data-center mapping now covers an estimated 44 % of global AI compute across 86 sites (higher coverage of 2026 deployments), with satellite and permit-derived H100-equivalent estimates. No major new chip announcements in the immediate window; energy and power delivery remain the binding constraints on further scale.

Political / legal / social

Dario Amodei’s 12 September essay “We Must Pace the Frontier” is the clearest primary statement from a frontier lab CEO: call for slower capability advance so that alignment and risk-prevention work can keep pace, unilateral Anthropic commitment to permanent third-party “embedded evaluators,” industry standards in democratic countries, and eventual global coordination. OpenAI, Google DeepMind, and others have publicly acknowledged ongoing safety talks and expressed varying degrees of support for third-party verification; antitrust concerns around coordination have been noted but not treated as blocking.

EU AI Act transparency obligations (Art. 50 marking, disclosure of AI interaction, deepfake labelling) became enforceable 2 August 2026; the AI Omnibus delayed standalone high-risk system obligations to December 2027 (product-embedded later). Systemic-risk GPAI rules and AI Office oversight continue to apply. US legislative discussion (FRONTIER Act references to independent verification) is active but no new primary statute has closed in the last week. Public polling (Epoch/Ipsos) shows near-daily AI use among US adults roughly doubling since March.

Watchlist hits (primary)

• Anthropic Claude Fable/Mythos 5.1 system card (1 Sep).
• OpenAI GPT-6 Astra system card and Preparedness Framework Critical cyber threshold (3 Sep).
• Artificial Analysis Intelligence Index v4.3 announcement and methodology (7 Sep; Terminal-Bench 4.0 + AutomationBench-AA).
• DeepSeek V4.1-Flash technical announcement (10 Sep).
• Dario Amodei “We Must Pace the Frontier” essay (12 Sep).
• Epoch AI data-center coverage update and Capabilities Index notes (mid-Sep).
• Google Gemini 3.8 Flash / Cyber / Live releases.
• arXiv: systematicity of reasoning models; SAEScientist-Bench; Cognitive Taxonomy framework.

Ledger actions

• Frontier label: dual hold on AA Intelligence Index v4.3 (Fable 5.1 max+fallback / Astra max). Task-specific boards (Terminal-Bench 4.0, AutomationBench-AA, cyber suites, SWE-ECI) already differentiate; any sustained >3-point gap or new private-set saturation would force a re-label.
• Efficient: OpenAI Astra configurations and selected Chinese Flash/MoE (DeepSeek V4.1-Flash, GLM-5.3-Flash) currently dominate cost-per-task at mid-to-high intelligence. A material cache-price cut or throughput jump by another lab would expand the set.
• Budget: sub-$0.50/M effective tiers remain DeepSeek / GLM / Qwen Flash variants; monitor retirement of older routes.
• Five-benchmark atlas (AA Index components + Terminal-Bench 4.0 + OSWorld-Verified + SWE-bench Verified + Epoch ECI or HLE/CritPt): no immediate rebase required, but AutomationBench-AA and Terminal-Bench 4.0 weight increases already shift agentic emphasis; watch for official OSWorld or SWE-bench version bumps.

Sources (selected primary or near-primary):
artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3
anthropic.com system cards (Fable/Mythos 5.1)
deploymentsafety.openai.com (GPT-6 Astra)
deepseek.com (V4.1-Flash)
epoch.ai (data centers, Capabilities Index, math-preprint tracking)
arxiv.org (2605.28405, 2609.13948, 2609.09113 et al.)
EU digital-strategy pages (AI Act / Omnibus timelines)
Amodei essay coverage with direct quotes from primary text.

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

Intelligence Radar — Daily Delta, 20 Sep 2026

Net assessment: I found one material state change since the previous scan: StepFun has formally launched Step 5 Preview. It does not move the absolute capability frontier, but it appears to move the cost/capability frontier, especially for agentic workloads. I found no comparably strong verified change today in biological intelligence, regulation, hardware availability, or the OpenAI/Anthropic/Google flagship frontier.

Step 5 Preview materially improves the low-cost frontier

StepFun announced Step 5 Preview on September 20 as a 600B-parameter sparse MoE model with 27B active parameters, a 1M-token context window, and text+image input. Multiple Chinese outlets reporting the company announcement say API and Studio access are now open, with full model weights planned for October 15. (EEO)

Independent measurements from Artificial Analysis give it an Intelligence Index score of 44, about 99.8 output tokens/s, and pricing of $1/M input tokens and $2.70/M output tokens. Its measured cost across the Intelligence Index is only $0.71 per task. (Artificial Analysis)

That combination is the important result. A score of 44 remains below the current broad aggregate leaders around 53, so under Intelligence Radar's current <10% from Frontier rule it would not be Topnotch on the broad composite:

44 / 53 ≈ 83%

But at $0.71 per evaluated task it becomes a strong candidate for Efficient or Budget on task classes for which its actual success probability clears sufficiency.

In other words:

Step 5 does not appear to move the global quality frontier; it moves the price of relatively high intelligence sharply downward.

That is exactly the distinction Intelligence Radar was designed to expose.

Artificial Analysis — Step 5 Preview measurements

Vendor claims suggest an important long-horizon/self-improvement angle — but should remain provisional

StepFun also reports two interesting long-duration experiments. In one, Step 5 reportedly optimized H100 GPU kernels for about 24 hours, reaching 508 TFLOPS after roughly 22 hours; in another, it generated post-training data that raised Qwen3-30B-A3B's AIME24 result from 53.3% to 60%. (凤凰网)

These should not yet enter the benchmark catalog as independently established capability results. I would record them as:

long_horizon_kernel_optimization
  evidence = vendor_reported
  confidence = provisional

AI_improves_other_AI_model
  evidence = vendor_reported
  confidence = provisional

They are nevertheless relevant to the RSI watchlist because they are another example of a frontier-ish model being used not merely to solve user tasks, but to improve computational machinery and training inputs for other AI systems.

Important availability/provenance caveat

There is a slightly unusual catalog-quality issue. At crawl time, StepFun's own public platform still prominently lists Step 3.7 Flash as its current production model, with no indexed Step 5 product page yet. (StepFun)

So I would not yet assign Step 5:

availability_confidence = high

Instead:

release_status        = released
api_status            = reportedly_available
first_party_docs      = lagging
availability_confidence = medium
weights_status        = announced_for_2026-10-15

This exposes a useful Intelligence Radar rule:

A release announcement, an independently callable endpoint, first-party documentation, and downloadable weights are four different availability states.

Step 5 should not yet be classified as an open-weight model merely because StepFun says weights will be released on October 15.

Radar state changes

I would make four updates today:

• add Step 5 Preview to the model/deployment catalog;
• flag it as a potential cost-frontier / Budget candidate, not a global Frontier model;
• introduce explicit availability states such as announced, API-preview, documented, GA, weights-announced, and weights-available;
• add its long-horizon kernel optimization and model-improvement experiments to the RSI evidence graph as vendor-reported / awaiting reproduction.

A related Reuters report says Anthropic is considering another model release in response to Astra, but this is sourced reporting about deliberations rather than a release, so I would keep it merely as a watch item, not a capability-frontier change. (Reuters)

No other development I could verify since the previous scan was strong enough to justify changing the current Frontier / Topnotch baseline.

The interesting move on September 20, 2026 is therefore economic:

High-end-but-not-frontier intelligence continues to become dramatically cheaper, increasing the distance between the “best model” and the economically rational model for ordinary work.

That makes the Efficient and Budget categories increasingly important rather than secondary to Frontier.