intelligence-radar/digests/daily/2026-09-21.md

Radar daily draft — 2026-09-21

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

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

Source update: openrouter-models

Source content was captured or changed. Extracted claims require editorial review.

• Status: reported; source kind: primary
• Source: https://openrouter.ai/api/v1/models
• Observed: 2026-09-21T19:37:39.174485+00:00
• Observation: 495bd94a389c6f9ef544ea8b6412293f81bc7a7e183e941ce8aafcbc7f35156d

Source update: anthropic-pricing

Source content was captured or changed. Extracted claims require editorial review.

• Status: reported; source kind: primary
• Source: https://claude.com/pricing
• Observed: 2026-09-21T19:37:38.124497+00:00
• Observation: 915e677e4b5376f5f0877896d20d61b80d6afe5eee6d87e14c3edc2b3fb85cfe

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), 21 Sep 2026

Intelligence Radar DAILY BRIEF
21 September 2026 (Europe/Berlin)
Frontier weather: dense September cluster of frontier releases continues; xAI ships Grok 4.7 today while Artificial Analysis Intelligence Index v4.3 holds Claude Fable 5.1 and GPT-6 Astra tied at the top, with capability gains concentrated in long-horizon agentic and terminal work amid rising coordination rhetoric and regulatory enforcement.

Frontier & markets

Confirmed releases dominate. xAI (SpaceXAI) launched Grok 4.7 on 21 September: new larger base model, extended RL on multi-hour tasks, native Grok Bot harness understanding, 500k context, text+image in / text out. Pricing matches Grok 4.6 at $2 / $0.50 cached / $6 per million tokens (under 200k prompt; doubles above). Available immediately via xAI API (grok-4.7), Cursor, Grok Build, and routers. Artificial Analysis scores it 46 on Intelligence Index v4.3 (xhigh), +2 over Grok 4.6, placing xAI in the top-4 labs; Coding Agent Index rises to 56 (DeepSWE v1.1 73%, Terminal-Bench 4.0 33%). Vendor numbers show gains on CursorBench 4.0 (46.3%), EEBench (64%), and legal/agentic sets.

Earlier confirmed: Anthropic Claude Fable 5.1 / Mythos 5.1 (1 September, same weights, different safeguards; Mythos limited-access for cyber/life-sciences). 1M context, $10/$50 (cache reads now $0.25). Stronger long-horizon coding, research, document work; cache-price cut yields ~25% lower typical cost. OpenAI GPT-6 Astra (3 September): 1.05M context, $10/$50, Critical cybersecurity threshold under Preparedness Framework, saturates several math/cyber sets (FrontierMath Tier 4 ~97–98%, ExploitBench 100%, ARC-AGI-3 ~99.9% under provider harness). Google Gemini 3.8 Flash / Live / Cyber variants mid-month; DeepSeek V4.1 Flash, Cognition SWE-2, Meta Muse Spark 1.3, Qwen3.8-Omni-Flash also confirmed.

Artificial Analysis Intelligence Index v4.3 (and .2 patch) remains the reference composite (AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, HLE, GDP.pdf, CritPt, AA-Omniscience, AA-LCR). Claude Fable 5.1 (max w/ fallback) and GPT-6 Astra (max) tied at 53; Opus 5 at 51; Muse Spark 1.3 at 48; GPT-5.6 Sol at 47. OpenAI occupies most of the cost-per-task Pareto; Astra delivers the 53 score at materially lower cost than Fable. Do not mix prior Index versions. Saturated exams are not used as rankers.

Availability remains broad for the new models; no major capacity pauses reported today. Price stability on Grok and cache cuts on Anthropic are the notable availability/economics moves.

Science of intelligence

Measurement continues to shift toward agentic, long-horizon, and professional-workflow tasks. Artificial Analysis Capability Indices v1.1 (14 September) map O*NET occupations onto sliced Intelligence Index components plus specialized evals, adding weight to agentic tool use. Terminal-Bench 4.0 and AutomationBench-AA (private test set) emphasize recovery, multi-step terminal work, and rule-following workflows. Epoch Capabilities Index notes GPT-6 Astra leading overall (Math-ECI strong) while trailing on software-engineering slices relative to Fable 5.1. An arXiv analysis of SWE-bench submissions documents convergence among top coding agents (shared successes dominate; scaffold effects large), arguing leaderboards no longer reliably order the frontier and proposing audit protocols instead.

No major new primary theory papers from the listed thinkers surfaced in the last 24–48 h; the dominant signal remains empirical progress on harder agentic and scientific reasoning sets rather than architectural claims.

Engineering

Agents and harnesses are first-class. Grok 4.7 was trained for native Grok Bot / long-session verification; Astra emphasizes async tool calling, fewer retries, and computer-use robustness; Fable 5.1 improves multi-hour coding and recovery. Cognition SWE-2 and various Flash variants target lower-cost coding agents. Inference economics favor models that finish tasks in fewer tokens (Astra highlighted).

Epoch AI reports continued rapid data-center scale-up: largest single-site compute capacity has roughly doubled every 7 months since Colossus 1; their explorer now covers ~44% of global AI compute across 86 sites (53% of 2026 deployments). OpenAI, Google DeepMind, Anthropic, Meta Superintelligence Labs, and xAI remain the dominant compute users, with OpenAI’s fleet growth especially steep. Energy and power constraints are tracked as first-order limits.

Political / legal / social

Confirmed: EU AI Act main rules and GPAI enforcement powers active since early August 2026; Digital Omnibus amendments (in force late July) delayed standalone high-risk obligations to December 2027 (embedded systems to August 2028) while adding transparency and certain prohibited practices. AI Office and national authorities are operational.

Reported: class-action antitrust suit filed 18 September in N.D. Cal. against Anthropic, OpenAI, Google, and xAI alleging coordinated slowdown of frontier development. Plaintiffs cite Amodei’s mid-September essay proposing independent evaluators with employee-like access, industry standards (compute/training limits, capability checkpoints), and international coordination, plus prior working-group discussions. Suit seeks treble damages and injunction; covers U.S. paid subscribers to the four products. Amodei’s essay itself is confirmed primary; the coordination claim is litigated and therefore tagged reported.

Social signal remains elevated around recursive self-improvement risks, cyber “swarm” scenarios, and the tension between safety pacing and competitive/geopolitical pressure (especially China).

Watchlist hits (primary links)

• xAI Grok 4.7 announcement & docs: https://x.ai/news/grok-4-7 · https://docs.x.ai/developers/grok-4-7
• Artificial Analysis Intelligence Index v4.3 / Grok 4.7 eval: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3 · https://artificialanalysis.ai/
• OpenAI GPT-6 Astra system card & related: https://deploymentsafety.openai.com/gpt-6-astra
• Anthropic Claude Fable/Mythos 5.1: https://www.anthropic.com/claude-fable-and-mythos-5-1
• Epoch AI data-center scaling: https://epoch.ai/latest/scaling-ai-data-centers-research
• EU AI Act / Omnibus official: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

Ledger actions

• Frontier label: Grok 4.7 enters consideration on coding-agent and price-performance boards (AA Coding Agent Index, CursorBench, EEBench); does not displace the Fable 5.1 / Astra 53-point AA II v4.3 tie for general Intelligence Index.
• Efficient / Budget: Grok 4.7 at $2/$6 strengthens the mid-frontier efficient tier; Anthropic cache-read cut improves Fable economics for repeated-prompt workloads. DeepSeek V4.1 Flash and various Flash/Omni models remain primary Budget candidates.
• Five-benchmark atlas (suggest: AA Intelligence Index v4.3, Terminal-Bench 4.0, SWE-bench Pro / FrontierSWE, AutomationBench-AA or GDPval-AA, HLE or SciCode): Grok 4.7 moves the coding-agent and terminal slices; Astra still leads several math/cyber and cost-adjusted agentic slices. No saturation-driven removal of any atlas member this cycle. Re-run after independent harness confirmation of Grok numbers.

Sources: primary lab pages and model cards (xAI, OpenAI, Anthropic, Google), Artificial Analysis, Epoch AI, BenchLM mirrors of AA and Terminal-Bench, EU Commission digital-strategy pages, and contemporaneous reporting on the Amodei essay and related suit (all tagged above).

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

Intelligence Radar — Daily Delta, 21 Sep 2026

Net assessment: the capability frontier itself appears stable since yesterday. I found no verified new flagship-model release, material benchmark jump, general API-price change, or production-hardware step-change. Two things did materially change: Step 5 Preview’s availability moved from uncertain to first-party confirmed, and AI incident handling entered formal U.S.–China diplomatic discussion.

1. Step 5 Preview’s availability uncertainty is resolved

Yesterday’s scan treated Step 5 Preview as released / reportedly API-available / first-party docs lagging, with only medium availability confidence.

That should now change. StepFun’s own product page is live and explicitly states that Step 5 Preview is available today through its products and API, while open weights are scheduled for October 15, 2026. StepFun’s public developer platform also lists Step 5 Preview among its current Frontier AI models and exposes the API endpoint.

The first-party page also confirms several claims that were previously mostly vendor/secondary reports: 600B total parameters, 27B active per token, 1M context, vision input, plus its reported coding, finance and 24-hour kernel-optimization experiments. StepFun itself positions the model explicitly around the Pareto frontier between intelligence and cost.

Radar state change:

Step 5 Preview
release_status           = released
api_status               = available
first_party_docs         = available
availability_confidence  = high
weights_status           = announced
weights_release_date     = 2026-10-15

Do not yet mark it open_weight; that status should change only when the weights are actually downloadable.

This validates yesterday’s proposed availability-state taxonomy:

announced → preview/API available → documented → GA → weights announced → weights available

and demonstrates why available = true/false is too lossy.

StepFun — Step 5 Preview
StepFun Open Platform

2. U.S.–China AI governance moves from abstract rivalry toward incident communication

On September 20, U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng held talks in New York that included a dedicated AI discussion. China’s Ministry of Commerce confirms that the delegations held a dialogue on artificial intelligence alongside the economic negotiations.

Bessent subsequently said the U.S. had proposed a notification mechanism for AI incidents rising to the level of national-security concern, as part of a broader “U.S.–China AI dialogue.” His stated aim is greater transparency between what he called the world’s number-one and number-two AI powers. China has not publicly accepted the specific notification mechanism yet, so this is a proposal under negotiation rather than an agreement.

That is a meaningful governance development because the discussion is moving from:

AI safety principles

to potentially:

incident → classification → notification → bilateral communication

This resembles an incident-management protocol between strategic competitors, rather than conventional domestic AI regulation.

Intelligence Radar impact

Add a governance construct:

international_AI_incident_notification

with fields such as:

participants
covered_incident_classes
notification_threshold
notification_latency
information_shared
verification_mechanism
confidentiality
escalation_channel
reciprocity
status

Also add cross_jurisdiction_incident_coordination to the security/governance taxonomy.

A particularly important distinction is emerging:

Preventing an AI incident and managing an AI incident after it occurs are separate governance capabilities.

Recent Radar observations have concentrated heavily on prevention, alignment and containment. A bilateral notification mechanism introduces response coordination as another layer.

For now I would classify the claim as:

US-China AI incident notification mechanism → proposed / bilateral discussion confirmed / not yet agreed

China Ministry of Commerce — U.S.–China talks included an AI dialogue
Reuters — U.S. proposes AI incident notification mechanism to China

Overall state

I would make no change today to the general Frontier / Topnotch model ranking and no benchmark-weight change.

The two concrete state transitions are instead:

Step 5 Preview: availability confidence medium → high

International AI governance: general calls for cooperation → concrete proposed incident-notification mechanism between the U.S. and China

I found no sufficiently strong new result since yesterday in biological intelligence, benchmark science, inference hardware, model pricing, or economic/social impact to justify another state change.

The interesting theme on 21 September 2026 is therefore not higher intelligence, but the infrastructure surrounding it becoming more explicit: what counts as available intelligence, and how states communicate when powerful intelligence causes or threatens an incident.