intelligence-radar/digests/daily/2026-09-22.md

Radar daily draft — 2026-09-22

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

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

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: 2026-09-22T20:30:31.734117+00:00
• Observation: 1d57b67a769851d195246782ef7c4106fc2d2a1e194820f018b54fba72535de0

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: 2026-09-22T20:30:32.773919+00:00
• Observation: 66a06a439a614a52dfbb956fdcc39462288ac5cfb10f060c0ca9ed8839d1cf1f

Source update: anthropic-sonnet-launch

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

• Status: reported; source kind: primary
• Source: https://www.anthropic.com/news/claude-sonnet-5
• Observed: 2026-09-22T20:30:32.513553+00:00
• Observation: 94759945d0a7266d89348114f996da3d83e822c9e8a3dede988e2988a8187621

Source update: anthropic-sonnet

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

• Status: reported; source kind: primary
• Source: https://www.anthropic.com/claude/sonnet
• Observed: 2026-09-22T20:30:32.120502+00:00
• Observation: d32a91fede8f7fbddef7bdc99e1384a5bdfa64730693dac5bb960513a12addf5

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), 22 Sep 2026

Intelligence Radar DAILY BRIEF
22 September 2026 (Europe/Berlin)

Weather of the frontier: Anthropic’s Claude Opus 5.5 claims the Artificial Analysis Intelligence Index lead with a material price cut while xAI’s Grok 4.7 lands as a strong price-performance contender; open-weight Chinese models continue closing the gap on agentic boards. Confirmed cluster of releases, not rumor.

Frontier & markets

Anthropic released Claude Opus 5.5 today (22 Sep). Primary system card and announcement confirm it as the first Claude 5.5-family model, matching or exceeding Claude Fable 5.1 / Mythos 5.1 on most capability suites while priced at $4 / $20 per million input/output tokens (20 % below Opus 5) and cache reads at $0.20 (60 % cut). Artificial Analysis Intelligence Index v4.3.x (max effort, with fallback) scores it 58—several points clear of prior leaders. Gains reported on Terminal-Bench 4.0 (59.6 %, level with GPT-6 Astra), GDPval-AA v2.1, AA-Briefcase, SciCode, Humanity’s Last Exam. External testing by Frontier Design and METR. Safeguards for cyber and biology now match Fable-class routing. Confirmed.

xAI (SpaceXAI) shipped Grok 4.7 yesterday (21 Sep) at unchanged $2 / $6 pricing and 500 k context. Model card and launch note describe a larger base, extended RL on multi-hour tasks, stronger self-verification. Artificial Analysis scores it 46 on the Intelligence Index; Coding Agent Index improves past GPT-5.6 Sol. Strong price-performance on CursorBench 4.0 and long-horizon knowledge work. Confirmed.

Xiaomi MiMo-V2.6-Pro (open-weight, ~1 T total / 42 B active MoE) debuted near 46.3 on the same Index, placing it among the strongest open models and within a point of several closed mid-tier frontier systems at far lower cost-per-task. Confirmed via Artificial Analysis and release trackers.

Earlier September cluster (still active): OpenAI GPT-6 Astra (3 Sep, Critical cyber threshold under Preparedness Framework), Google Gemini 3.8 Flash + Cyber variant (Fairwind Program), Anthropic Claude Fable/Mythos 5.1 (1 Sep, 75 % cache-read cut), Meta Muse Spark 1.3, DeepSeek V4.1 Flash, Cognition SWE-2. Availability remains broad on major clouds; no major capacity pauses reported today. Never mix Index versions—current reference is v4.3.x (10 evaluations including Terminal-Bench 4.0, AutomationBench-AA, GDPval-AA v2.1, AA-Briefcase). Saturated exams are not rankers.

Science of intelligence

Primary literature continues to shift from absolute human-scale benchmarks toward relative and recursive measurement. A September arXiv piece formalizes recursive self-improvement (RSI) stages—from improvement-execution autonomy through meta-improvement—using a Headroom-Closed Index to diagnose current LLM limits; application domains include scientific discovery and software engineering.

“Measuring Intelligence Beyond Human Scale” (arXiv) proposes adversarial psychometrics: models generate public challenges that separate other systems, producing scalable relative ratings once human-authored tasks saturate.

Christopher Summerfield’s Neuron review (online 18 Sep) argues that the “recipe” (how ingredients combine) now shows surprising parallels between biological and artificial systems after a decade of scale-driven progress that largely bypassed earlier brain-inspired roadmaps.

Epoch Capabilities Index (ECI) data (updated 22 Sep) continues to show ~14-point annual frontier advance since reasoning models; GPT-6 Astra leads overall ECI but trails on software-engineering slices. No single theory of general intelligence is declared; measurement remains multi-harness.

Engineering

Agentic harnesses dominate: Terminal-Bench 4.0 and AutomationBench-AA are now core to Artificial Analysis Index weighting. Cognition SWE-2 leads older Terminal-Bench 2.1 snapshots; OSWorld-Verified remains competitive for Chinese open models (Qwen3.8 Max at 86.1 %). SWE-Bench Pro V2 released today with a cleaned 642-task public split and network-locked protocol—Claude Opus 5, Kimi K3 and GPT-Astra all >96 % on the refreshed suite; a new “Hard” 51-task split targets residual failure modes.

Inference economics: Opus 5.5 and Grok 4.7 both emphasize lower cost-per-task at high effort via better token efficiency and cache pricing. No new chip announcements today; energy and cluster data remain tracked by Epoch (data centers updated continuously).

Political / legal / social

EU AI Act transparency obligations (Article 50) and GPAI enforcement powers have been live since 2 August 2026. The Digital Omnibus (Regulation 2026/1744, in force late July) deferred standalone high-risk Annex III obligations to 2 December 2027 and product-embedded systems to August 2028; a new prohibited practice on non-consensual intimate imagery applies from December 2026. Confirmed primary regulatory texts.

No major new US federal action reported today; state-level transparency and frontier-safety rules (California et al.) continue to operate. Lab system cards increasingly route dual-use cyber/biology queries and document external audits—Opus 5.5 and GPT-6 Astra both highlight this.

Watchlist hits (primary links)

• Claude Opus 5.5 system card & announcement: anthropic.com (CDN PDF + news page, 22 Sep)
• Grok 4.7 model card & launch: x.ai / media.x.ai (21 Sep)
• Artificial Analysis Intelligence Index update: artificialanalysis.ai/articles/claude-opus-5-5 (22 Sep)
• SWE-Bench Pro V2: labs.scale.com (22 Sep)
• Epoch ECI / capabilities data: epoch.ai (updated 22 Sep)
• EU AI Act Omnibus text & Commission pages: digital-strategy.ec.europa.eu

Ledger actions

• Frontier label: Claude Opus 5.5 (AA Intelligence Index v4.3.x max) now sits alone at the top; GPT-6 Astra remains co-leader on several agentic/cyber boards. Rebase any “Fable 5.1 / Astra dual” notes.
• Efficient: Grok 4.7 and MiMo-V2.6-Pro strengthen the price-performance Pareto; cache-price cuts on Opus move long-horizon agent economics.
• Budget: DeepSeek V4.1 Flash and Xiaomi variants continue to dominate sub-$1/task open/closed tiers.
• Five-benchmark atlas (suggest: AA Intelligence Index v4.3.x, Terminal-Bench 4.0, GDPval-AA / AA-Briefcase, SWE-Bench Pro V2, OSWorld-Verified) requires version pins; older Terminal-Bench 2.x and saturated SWE-bench Verified no longer rank the frontier.

Sources: Primary lab pages (Anthropic, xAI), Artificial Analysis articles and Index snapshots (22 Sep), Epoch AI data pages, Scale Labs SWE-Bench Pro V2 post, official EU regulatory texts and Commission digital-strategy pages, selected arXiv cs.AI / q-bio.NC preprints, BenchLM release trackers cross-checked against lab announcements. All items tagged confirmed via primary or contemporaneous independent measurement; no unconfirmed rumor included.

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

Intelligence Radar — Daily Delta, 22 Sep 2026

Net assessment: the absolute closed-model frontier did not clearly move overnight, but the open-weight and cost frontiers did. The strongest new capability signal is Xiaomi’s MiMo‑V2.6 release. Separately, OpenAI disclosed an unusually large body of claimed machine-generated mathematical research, while both OpenAI and SpaceXAI supplied evidence that our current Radar abstractions around research capability, task economics, and governance of automated R&D need refinement.

1. MiMo‑V2.6 moves the open-weight and cost frontier

Xiaomi has released MiMo‑V2.6 Pro and Flash, with model weights and a technical report available through the XiaomiMiMo organization. MiMo‑V2.6 Pro is a sparse MoE with roughly 1.02T total / 42B active parameters, native text/image/audio/video input and a 1M-token context window. Xiaomi’s model card reports strong agentic, coding, computer-use and cybersecurity results and describes a single mixed reinforcement-learning regime spanning multiple domains and harnesses.

More importantly, independent Artificial Analysis testing currently puts MiMo‑V2.6 Pro at 46 on its Intelligence Index, #1 among the open-weight models it tracks, at about 134 output tokens/s and only $0.13 per Intelligence Index task using Xiaomi’s API. That places it on Artificial Analysis’s intelligence-versus-task-cost Pareto frontier.

This is a more consequential economic delta than yesterday’s Step 5 update. MiMo‑V2.6 combines three properties that rarely coincide:

high-end capability + genuinely open weights + extremely low hosted inference cost.

It does not establish a new absolute model frontier—the score remains below the strongest closed systems—but it likely changes both the Open Frontier and, for many tasks, the Efficient/Budget frontier.

Radar changes: add an explicit Open Frontier alongside Frontier / Topnotch / Efficient / Budget / Resilient Frontier. Open weights are not merely another availability flag: they materially change provider continuity, self-hosting, auditability, jurisdictional exposure and marginal economics.

The release also provides unusually useful evidence for the emerging Intelligence Amplification Loop concept: Xiaomi describes scaling RL compute, environment diversity and grader compute together and publishes an RL system intended to keep improving capabilities through exploration and feedback. Treat Xiaomi’s stronger “self-improvement” framing as a development-method claim, not evidence of autonomous RSI.

Xiaomi MiMo‑V2.6 Pro model card
Artificial Analysis — MiMo‑V2.6 Pro

2. OpenAI says one internal model has now resolved more than 100 long-standing mathematics problems

OpenAI disclosed on September 21 that an internal model whose training began on August 28 has, in its assessment, resolved more than 100 long-standing open mathematical problems across most areas of mathematics, in addition to its previously disclosed Navier–Stokes result. OpenAI says the pace surprised even its internal mathematicians.

If even a substantial fraction of those results survive expert review, this is potentially one of the largest capability-frontier developments of the month: it would move AI from demonstrating elite mathematical problem-solving toward industrial-scale production of novel mathematical results.

But the epistemic qualifier is crucial. The announcement does not provide a public inventory and validation record for those 100-plus claims. They should therefore enter Intelligence Radar as:

provider-reported research results / potentially frontier-moving / awaiting result-level external verification

—not as 100 established discoveries.

OpenAI simultaneously helped establish an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study and including Timothy Gowers, Martin Hairer, Edward Witten, Ravi Vakil and others. Members are unpaid, may make their advice public and can criticize OpenAI, but have advisory rather than decision-making authority and explicitly do not govern the pace of OpenAI’s internal mathematical research.

Radar impact: scientific capability now needs something beyond conventional benchmark scores. Add a Research Result Evidence Ladder such as:

generated → internally checked → formally verified → externally reviewed → independently reproduced → field accepted

and attach the status per result, not per model.

This is particularly important because formal proof checking, mathematical novelty, correspondence to the original problem, attribution, and broad community acceptance are different claims.

3. Grok 4.7 improves agentic work substantially—but demonstrates why token price is a bad economic metric

SpaceXAI released Grok 4.7 on September 21 at the same $2/M input and $6/M output list price as Grok 4.6. First-party results show meaningful improvements across coding, professional work and engineering tasks.

Independent testing makes the picture clearer. Artificial Analysis gives Grok 4.7 46 on its Intelligence Index, only +2 over Grok 4.6 overall, but finds much larger improvements in agentic work: +111 Elo on AA-Briefcase and a Coding Agent Index jump from 47 to 56 with Grok Build.

The economics contain the more interesting lesson. Grok 4.7 used roughly 81k output tokens per Intelligence Index task, versus 36–38k for Grok 4.6 and 27k for GPT‑6 Astra. Thus a model can retain exactly the same advertised token price while becoming considerably more expensive per attempted task because it thinks longer.

That gives us unusually clean evidence for a Radar policy already emerging during the week:

$/token is an infrastructure tariff, not the price of intelligence.

The useful economic hierarchy should be:

$/token → $/attempt → $/completed task → $/acceptable task → $/acceptable unattended task

with the final measures incorporating retries, tool use, wall-clock duration, verification and human cleanup.

Grok 4.7 therefore strengthens SpaceXAI’s agentic knowledge-work position, but does not displace the broad absolute frontier.

SpaceXAI — Grok 4.7
Artificial Analysis — Grok 4.7 evaluation

4. OpenAI proposes standardized measurement of automated AI research and RSI

A second OpenAI publication from September 21 is more important to Intelligence Radar than a generic policy statement because it proposes concrete measurement objects.

OpenAI says fully autonomous RSI is not happening today, while increasingly automated AI research can already form a partially human-supervised self-improvement loop. It advocates international technical standards covering measurement of AI-driven R&D, human oversight thresholds and common classification/reporting of alignment and automated-research incidents.

Specifically, it proposes common standards around:

• the amount of autonomous research occurring inside AI companies;
• which automated research events require immediate human review;
• shared incident severity levels and reporting thresholds;
• comparable capability and safeguard evidence.

This substantially validates the Development Process Metrics introduced into Intelligence Radar last week.

I would now split that family into:

R&D capability metrics — how much research AI can perform;
R&D autonomy metrics — how independently it performs it;
R&D control metrics — what humans can observe, interrupt and review;
R&D acceleration metrics — how much faster/better the next system becomes as a result.

That separation is preferable to using a single “RSI level.”

OpenAI — Building standards for the next phase of AI

5. Smaller but noteworthy: AI retrosynthesis becomes more practically credible

Microsoft Research and collaborators at GSK and Novartis published RetroChimera, an open-source retrosynthesis system now described in Nature. It combines models with different inductive biases and supports zero-shot transfer and fine-tuning onto proprietary chemistry datasets. Microsoft reports that expert chemists fully accepted its route for 9 of 10 benchmark molecules, versus 2–5 for compared systems, and preferred its proposed routes to documented alternatives roughly 64% of the time in a blinded comparison.

This is not a general-intelligence frontier update, but it strengthens the Compendium claim that AI-for-science capability increasingly depends on closed-loop domain systems rather than language-model scientific QA scores. The remaining uncertainty is experimental: computationally plausible synthesis plans are not equivalent to successful wet-lab execution.

Microsoft Research — RetroChimera

State changes for 22 September

The clearest changes I would commit to Intelligence Radar today are:

MiMo‑V2.6 Pro → new Open Frontier / strong Efficient candidate.

OpenAI mathematics → potentially major scientific-capability signal, but retain provider-reported / externally unverified at batch level.

Grok 4.7 → stronger agentic-work contender, no broad Frontier change; reinforces cost-per-task over cost-per-token.

Automated AI research → replace one-dimensional RSI tracking with capability / autonomy / control / acceleration dimensions.

Scientific evaluation → add result-level evidence maturity from generated result through community acceptance.

I found no comparably strong new biological-intelligence result, enacted AI law, or production-hardware transition since the previous scan that warrants another state change. The largest delta this morning is therefore the rapid compression of the gap between open, cheap intelligence and proprietary frontier intelligence, while the frontier labs simultaneously begin treating the production and validation of genuinely new scientific knowledge as an operational governance problem.