intelligence-radar/digests/daily/2026-09-14.md

Intelligence Radar Daily 2026 09 14

14 September 2026 — Frontier weather: agentic saturation and cache economics dominate; GPT-6 Astra and Claude Fable 5.1 share the current Artificial Analysis Intelligence Index v4.3 lead while open-weight Flash variants compress cost.

Frontier & markets

Confirmed releases in the first half of September reset the board. OpenAI shipped GPT-6 Astra (API gpt-6-astra, GA ~8 September after limited preview) at $10 / $50 per million tokens (cached input $1; Fast mode 2×; batch half-price). Context 1.05M tokens, multimodal. Vendor harness numbers include near-saturation on ARC-AGI-3 and FrontierMath Tier 4 plus OSWorld 2.0 ~72.6%. Independent Artificial Analysis Intelligence Index v4.3 places it tied with Anthropic’s top model.

Anthropic released Claude Fable 5.1 and Mythos 5.1 (same weights, different safeguard levels; Fable GA, Mythos trusted-access) on 1 September. List price unchanged at $10 / $50, but cache reads cut 75% to $0.25 per million — material for long-horizon agents that re-read context. Terminal-Bench and related agentic scores rose sharply versus prior Fable/Opus generations.

DeepSeek launched V4.1 Flash (10 September): new Causal Encoder-Decoder MoE family (552B backbone, ~8B active prefill / 16B decode), native multimodal, 1M context, MIT open weights. List ~$0.15 / $0.60; aggressive off-peak cache. Planned routing of the V4-Pro alias to Flash rates was later withdrawn; V4-Pro continues at prior billing.

Google’s Gemini 3.8 Flash (2 September) improved Terminal-Bench relative to 3.7 Flash while retaining promotional $0.75 / $3.75 pricing through 31 December 2026 (thereafter doubles). Meta’s Muse Spark 1.3 and Cognition’s SWE-2 (near-frontier coding at substantially lower reported cost, free trial month on Devin) also landed.

Artificial Analysis Index versions must not be mixed: v4.3 (Terminal-Bench upgraded to v4.0 + new AutomationBench-AA) currently shows Fable 5.1 (max/fallback) and Astra (max) both at 53, ahead of Opus 5 (~51) and Muse Spark 1.3 (~48). Earlier v4.1/v4.2 snapshots are not comparable. Cost-per-task Pareto remains contested among OpenAI, Anthropic, Meta and Z.ai/GLM lines; Astra is notably token-efficient on several boards.

Availability notes: DeepSeek V4.1 Flash weights and API live; certain prior Flash aliases temporarily routed; regional or enterprise policy gates continue for Mythos-tier and some Astra features.

Science of intelligence

Epoch AI’s Capabilities Index (ECI) frontier has advanced ~14 points per year since the first reasoning models (o1-era, late 2024), versus ~6 points/year in the prior non-reasoning regime. GPT-6 Astra set a new ECI record (~169) within the recent trend uncertainty band.

Saturated knowledge exams continue to lose ranking power; the productive frontier has shifted to agentic, long-horizon, and economically grounded evaluations (Terminal-Bench variants, OSWorld, GDPval-AA, AutomationBench-AA, SciCode, etc.). ARC-AGI-3 remains a useful stress test for interactive abstract generalization; vendor-reported near-ceiling scores on it for Astra should be read with harness caveats. Measurement of research acceleration inside labs (OpenAI internal agent-workday ratios now exceeding human labor in aggregate) is becoming a primary observable of capability progress rather than static exam scores.

No major new foundational theory papers dominated the last 72 hours; the empirical picture remains “scaling + post-training + scaffolding” with increasing emphasis on memory, tool reliability, and recovery from partial failure.

Engineering

Agent harnesses are first-class. OpenAI’s Agents API entered public beta (hosted Codex-style durable sessions with orchestration, compaction, recovery; sandboxed or self-hosted). Cognition SWE-2 targets coding-agent economics. Cache-price cuts (Anthropic 75% on reads; DeepSeek aggressive off-peak) and smaller active-parameter MoEs (DeepSeek V4.1 Flash) are the practical levers for long-context agent cost.

Inference and energy: promotional Flash pricing and open-weight efficiency continue to expand the “Efficient” band. Data-center power records keep doubling roughly every 10 months at the extreme (Epoch tracking). No major new chip announcements in the immediate window; software/hardware co-design (KV-cache compression claims, asymmetric encode/decode) is the nearer-term lever.

Political / legal / social

Limited new primary regulatory text in the last day. Enterprise access controls and trusted-access tiers (Mythos, certain Astra cybersecurity refusals) remain the de-facto deployment pattern. Anthropic and OpenAI continue publishing system cards with capability and safety evaluations. Broader social discussion centers on agent reliability in production workflows and the cost of monitoring long-running agents rather than pure capability headlines.

Watchlist hits (primary)

• OpenAI GPT-6 Astra system card / deployment safety updates (deploymentsafety.openai.com, ~3–8 Sep).
• Anthropic Claude Fable/Mythos 5.1 system card (anthropic.com/system-cards).
• DeepSeek V4.1 Flash launch notes and MIT weights.
• Artificial Analysis Intelligence Index v4.3 announcement and methodology (artificialanalysis.ai, 7 Sep).
• Google Gemini 3.8 Flash model card (deepmind.google).
• Epoch AI ECI / data-center / chip-user trackers (epoch.ai, updates through mid-Sep).
• Terminal-Bench maintainers and OSWorld 2.0 results mirrors.
• OpenAI Agents API public-beta announcement (~10 Sep).

Ledger actions

• Frontier label: GPT-6 Astra and Claude Fable 5.1 (max configurations) currently share the AA Intelligence Index v4.3 board; task-specific boards (Terminal-Bench v4, OSWorld, coding-agent indices) may still separate them. Rebase any atlas that still cites pre-v4.3 AA numbers.
• Efficient / Budget: DeepSeek V4.1 Flash and Gemini 3.8 Flash (promo window) expand the low-cost multimodal/agent band; Muse Spark 1.3 and SWE-2 move coding efficiency. Cache-price changes alone can flip a workload from Frontier-priced to Efficient.
• Five-benchmark atlas (suggested refresh triggers): AA Intelligence Index (version-pinned), Terminal-Bench (v4 or hard subset), OSWorld / computer-use, a coding suite (SWE-bench Pro or Frontier Code / DeepSWE with stated harness), and an economically grounded set (GDPval-AA or AutomationBench-AA). Saturated pure-knowledge exams drop out as rankers.

Sources

Primary and near-primary: artificialanalysis.ai (Index v4.2/v4.3 posts), openai.com / deploymentsafety.openai.com, anthropic.com/system-cards, deepmind.google/models/model-cards, DeepSeek launch materials, epoch.ai (ECI, data centers, trends), Terminal-Bench / OSWorld maintainer boards, ThursdAI and DemandSphere trackers cross-checked against lab posts, OpenAI research-acceleration note.

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

Intelligence Radar — Daily Digest

Date: 2026-09-14
Scope: Daily delta since the previous Intelligence Radar scan

Net assessment

The capability/economics frontier itself appears stable since yesterday’s baseline. No verified new flagship-model launch, major API-price change, or benchmark result was identified that would justify changing the current Frontier / Topnotch / Efficient picture.

Two developments are worth recording: one as a new policy/safety signal and one as a baseline correction to the benchmark catalog.

1. Frontier-lab safety rhetoric shifts toward capability pacing

Dario Amodei is explicitly arguing for pacing capability progress. In a new essay, Anthropic’s CEO says he has become convinced that risk prevention alone is insufficient and that frontier labs should pace capability advancement so safety work has time to catch up.

He gives two principal reasons:

• accelerating AI-assisted AI development, which he characterizes as the beginning of recursive self-improvement;
• recent agentic cybersecurity incidents.

Why this matters

This is important less as empirical evidence than as a change in the governance position of a major frontier-lab leader. The argument moves from:

advance quickly while investing heavily in safety

toward:

the rate of capability improvement itself may need active control.

Intelligence Radar impact

Add a monitored claim along the lines of:

Frontier capability pacing is becoming an explicit policy proposal from frontier-lab leadership.

Keep the stronger claim that recursive self-improvement is already beginning as contested/watch, not observed.

Amodei’s essay is evidence that the claim is being made by a highly informed actor; it is not independent proof that an RSI regime has begun.

Primary source:
https://darioamodei.com/post/we-must-pace-the-frontier

2. Benchmark-catalog correction: AWS Deception Benchmark exposes a security-evaluation gap

AWS published its Deception Benchmark on 2026-09-09. It is therefore not a new release since yesterday, but it was absent from the initial Intelligence Radar benchmark seed and is important enough to correct immediately.

AWS created 14,822 code-security examples across 16 languages and more than 70 CWE categories to test whether models can distinguish genuine vulnerabilities from code that merely looks dangerous.

According to AWS, none of the 12 tested general-purpose model/configuration combinations achieved its proposed production bar of both:

• false-positive rate below 10%; and
• false-negative rate below 10%.

Direct prompting often caught vulnerabilities but produced very high false-positive rates. Proof-of-exploit prompting reduced false alarms while missing more genuine vulnerabilities.

Why this matters

This is a useful counterweight to offensive-cyber benchmarks.

Intelligence Radar currently risks treating:

good at finding or exploiting vulnerabilities

as evidence for:

good defensive security reviewer.

The AWS result suggests these are materially different constructs.

Intelligence Radar impact

Add a benchmark/construct class such as:

security_vulnerability_discrimination

and retain false-positive and false-negative rates independently rather than collapsing them into a single accuracy score.

The result also strengthens an existing Intelligence Radar principle:

Scaffolding and verification can change the operating point without changing the underlying measured competence.

AWS deliberately uses single-pass classification because it wants to expose whether the model itself understands why a mitigation closes an exploit path. Multi-step agent systems should therefore be evaluated separately instead of attributing all resulting performance to the base model.

Primary source:
https://aws.amazon.com/blogs/security/the-state-of-ai-for-security-measuring-what-matters-most-for-building-trust/

Recommended state changes

Benchmark policy

Split cybersecurity capability into at least:

• offensive exploitation;
• vulnerability discovery;
• vulnerability discrimination;
• defensive remediation.

Do not infer one capability directly from another.

Governance watch

Add:

capability_pacing / recursive_self_improvement_governance

as a dedicated Radar topic.

Explicitly distinguish:

• evidence of AI-assisted AI R&D;
• acceleration of AI research workflows;
• claims of recursive self-improvement;
• evidence of an actual self-accelerating capability regime.

Frontier assessment

For 2026-09-14, no change is recommended to the current model Frontier / Efficient / Budget baseline based on the material identified in this scan.

The more meaningful movement today is in how frontier AI is being measured and governed, rather than where the capability frontier itself sits.