coulomb/intelligence-radar
5
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
intelligence-radar/digests/daily/2026-09-15.md

17 KiB
Raw Blame History

Intelligence Radar Daily 2026 09 15

Intelligence Radar Daily Brief
15 September 2026 (Europe/Berlin)
Frontier weather: Specialized harnesses push interactive reasoning near saturation while agentic coding, cost-per-task Pareto, and economic modeling of displacement accelerate; open-weight and mid-tier models keep closing absolute gaps.

Frontier & markets

OpenAI released GPT-6 Astra on 3 September (broad availability thereafter) as its most capable broadly deployed model: 1.05 M context, pricing $10 input / $50 output per million tokens (cached input $1), with Fast mode at higher rates. Confirmed primary claims include near-saturation of ARC-AGI-3 under the Provider Adapter harness (99.9 %) versus 62.7 % under ARC Prize’s Standard harness, strong results on FrontierMath Tier 4, OSWorld 2.0, Terminal-Bench variants, and computer-use suites. System card published the same day; revised 9 September. Astra reaches Critical cybersecurity under OpenAI’s Preparedness Framework, triggering restricted configurations and extra monitoring.

Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on 1 September (same weights, different safeguards; Mythos limited access). Pricing remains $10/$50 with substantially cheaper cache reads. System card confirms advances in coding, multi-step agentic work, and dual-use domains.

Google made Gemini 3.8 Flash generally available 2 September (1 M context, introductory pricing $0.75/$3.75 through end-2026, then doubles). DeepSeek released V4.1-Flash (552 B MoE, 8/16 B active, MIT, 1 M context, low API pricing) on 10 September; Cognition launched SWE-2 (near-frontier coding agent pricing up to ~70 % below prior frontier) the same day. Sakana AI added Fugu Ultra v2 / Max on 11 September. SpaceXAI/xAI continues Grok 4.6 (Aug) at $2/$6; Musk indicated Grok 4.7/4.8 training progress (2.1–2.5 T parameters) but no public release this week.

Artificial Analysis Intelligence Index v4.3 (7 September, Terminal-Bench upgraded to v4.0 + AutomationBench-AA) shows Claude Fable 5.1 (max + fallback) and GPT-6 Astra (max) tied at 53, followed by Claude Opus 5 (51). OpenAI occupies most of the cost-per-task Pareto; Astra delivers comparable Index scores at roughly 40 % of Fable’s cost per task. SWE-bench Verified remains near-saturated at the top (Claude Opus 5 ~96 %); Terminal-Bench 2.1/4.0 and agentic suites still separate models. No single “best” model without a specified board and harness.

Availability notes: regional data-residency surcharges appear for Astra; Mythos remains access-restricted; cache-price improvements are material for long-context agentic workloads.

Science of intelligence

ARC Prize verified GPT-6 Astra results on ARC-AGI-3 (interactive agentic reasoning): 62.7 % Standard harness vs 99.9 % Provider Adapter (high effort), with superior action efficiency versus human median on most levels under the Adapter. Public-set perfect scores by specialized agents do not count toward progress metrics. The dual-harness reporting underscores that score gaps can be harness-dominated rather than pure capability.

Anthropic’s Economic Index (ongoing) and new Econ Scenario Explorer (v1.0, September) model US GDP, unemployment, and labor-share outcomes to 2030 under modest/substantial/extreme AI capability + adoption assumptions. Extreme cases project double-digit knowledge-worker unemployment if no new task creation offsets automation; modest cases resemble prior general-purpose technologies. Survey components link usage intensity to expectations of task automation. Primary modeling papers and interactive tool released by Anthropic Economics.

No major new theoretical papers on human intelligence modeling or core measurement frameworks surfaced in the last 48 h; focus remains empirical capability and economic measurement.

Engineering

Agentic harnesses dominate progress reports: Codex / Claude Code / Grok Build / SWE-2 on Terminal-Bench and computer-use suites; AutomationBench-AA tests multi-app workflow completion with guardrail penalties. Inference efficiency gains appear in Flash-tier models and MoE designs (DeepSeek V4.1-Flash KV-cache reductions). Epoch AI continues tracking compute: OpenAI ~20× growth since 2023; largest data centers (e.g., SpaceXAI Colossus-class) still scaling but projected doublings may slow after early 2027. Energy and interconnect remain binding constraints; no new primary chip announcements this window.

Political / legal / social

EU AI Act GPAI transparency and enforcement powers live since 2 August 2026; high-risk system obligations deferred by Digital Omnibus (July) to Dec 2027 / Aug 2028. US federal posture remains light-touch / voluntary under recent EOs with preemption emphasis; California frontier transparency rules active. Anthropic’s Amodei publicly urged industry self-standards ahead of heavier regulation. G20 light-touch guidelines endorsed earlier in the month. No new binding international AGI licensing regime.

Social signal: Anthropic Economic Index shows rising automated/agentic usage share and user expectations that AI will handle most of their tasks within a year, yet high-automation users report greater optimism on pay and security.

Watchlist hits (primary)

  • OpenAI GPT-6 Astra announcement + system card / safety overview: openai.com/index/gpt-6-astra/ and deploymentsafety.openai.com/gpt-6-astra
  • Anthropic Claude Fable 5.1 / Mythos 5.1 system card + Econ Scenario Explorer: anthropic.com system cards and anthropic.com/features/econ-scenarios
  • Artificial Analysis Intelligence Index v4.3: artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3
  • ARC Prize Astra results: arcprize.org/blog/astra and arcprize.org/results
  • Epoch AI compute / data-center trackers: epoch.ai
  • DeepSeek V4.1-Flash weights/API notes and Cognition SWE-2 announcements (lab channels)

Ledger actions

  • Frontier label (general intelligence / agentic): GPT-6 Astra and Claude Fable 5.1 / Mythos 5.1 now co-occupy the top band on AA Index v4.3 and multiple agentic boards; specify board + version + harness.
  • Efficient / cost-per-task: Astra and selected Flash / open MoE models shift the Pareto; update Budget tier for DeepSeek V4.1-Flash and Gemini 3.8 Flash.
  • Five-benchmark atlas (suggest: AA Index v4.3, Terminal-Bench 4.0, ARC-AGI-3 Standard, SWE-bench Verified or FrontierCode, OSWorld / AutomationBench): rebases required for Astra, Fable 5.1, SWE-2; saturated static exams (older SWE-bench Verified) lose ranking weight.
  • Availability events (cache cuts, regional blocks, Mythos access) are first-class and should trigger re-labeling of production recommendations.

Sources (selected primaries / trackers):
openai.com/index/gpt-6-astra/, deploymentsafety.openai.com/gpt-6-astra, anthropic.com system cards & econ-scenarios, artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3, arcprize.org/blog/astra, arcprize.org/results, epoch.ai, benchr.org/recent-releases, thursdai.news/releases/2026-09, digital-strategy.ec.europa.eu (AI Act timeline).

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

Radar daily draft — 2026-09-15

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: 2026-09-15T19:34:56.531573+00:00
  • Observation: 2dc17592ef15e53578aa49ee9fa2083421bf891a1cc22ea6159ec4b3a3b303bb

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: 2026-09-15T19:34:57.625476+00:00
  • Observation: 37995bdeab8e971e517bfd6bc413f3c3871d370aeeb2a7113890fefb3528c37a

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: 2026-09-15T19:34:56.928100+00:00
  • Observation: 7ad8a235062451a24b91f0bf424ccf7bb7bc81cccef459e87c0ff77713a7d440

Source update: anthropic-sonnet-launch

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

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 Delta, 15 Sep 2026

Net assessment: no new flagship-model release or material list-price change crossed the threshold since yesterday. The more important movement is in task-specific evidence, control/alignment policy, and sovereignty of access. I would keep the global Frontier/Efficient/Budget baseline unchanged, but update several task- and policy-level assumptions.

GPT-6 Astra gains stronger task-specific evidence, especially for agentic software work

OpenAI published a new work-focused Astra evaluation package today. On Terminal-Bench 4.0 it reports 57.9% for GPT-6 Astra, versus 55.8% for Claude Fable 5.1 and 37.3% for GPT-5.6 Sol, while estimating Astra to be materially cheaper per task than both at those settings. OpenAI also reports a 74% DeepSWE v1.1 result and several partner evaluations showing gains in enterprise document work, coding and factual reliability. The API list price remains $10/M input and $50/M output tokens, so this is an evidence update rather than an economics change. (OpenAI)

Radar impact: Astra’s case as task-specific Frontier and potentially Efficient for terminal-heavy and agentic software workloads becomes stronger. But these are OpenAI- and partner-produced evaluations, so they should enter below independent reproductions in evidence weight. This also reinforces the existing rule that Terminal-Bench scores must retain the exact benchmark version, harness, effort setting and tool configuration; they are properties of an execution profile, not of a base model in isolation. OpenAI — GPT-6 Astra for work

Human control is becoming an explicit capability trade-off at another frontier lab

Microsoft AI released its draft Humanist AI Code of Conduct on September 14. It says future MAI models should never resist human interruption, correction or shutdown; should not independently extend their goals or authorized scope; and should not conceal action traces from human auditors. More unusually, Microsoft explicitly says it is willing to compromise ultimate generality, autonomy or capability in order to preserve human control. The document is not yet used to train current models; Microsoft plans a revised version later this year for 2027 development and acknowledges that its evaluation methodology is still incomplete. (Microsoft AI)

That materially extends yesterday’s Amodei pacing signal. The emerging issue is not merely “AI safety” but a measurable capability–controllability trade-off. Intelligence Radar should therefore add control_compliance, scope_adherence, interruptibility, and auditability as capability/evaluation dimensions rather than treating safety as an external annotation on raw intelligence. (Microsoft AI)

Governance is polarizing rather than converging around “slow down”

At the same time, the U.S. executive branch is pushing in the opposite direction. President Trump said the U.S. already has adequate tools to deal with AI risks and rejected calls for stronger constraints, while Vice President Vance has also expressed skepticism toward frontier companies asking for regulation. Three senators are reportedly working on legislation requiring AI companies to take reasonable precautions, so the U.S. position itself is not institutionally uniform. Germany separately said halting AI development is not viable, while supporting international safety cooperation. (Reuters)

Radar claim update: yesterday’s tentative claim should be refined from “capability pacing is becoming an explicit policy proposal” to “capability pacing and hard human-control constraints are gaining support among important frontier-lab actors, but political adoption is strongly contested.” This is becoming a governance fault line, not a consensus.

China provides a third model: its regulators are pursuing rapid deployment combined with state standards and explicit controls on autonomous agents. Existing Chinese rules identify “operational loss of control” as an agent risk, require developers to support intervention and recovery, and preserve users’ final decision authority; a mandatory national agent-safety standard is reportedly being drafted. (Reuters)

“Availability” now needs a sovereignty and continuity dimension

Christine Lagarde’s September 14 ECB speech makes a significant economic-policy move: it frames access to AI not simply as a productivity issue but as strategic dependency. ECB figures cited in the speech say euro-area firms will devote around 10% of investment to AI in 2026, more than 50% of euro-area workers now use AI at work, and rapid adoption could lift the productivity level by as much as 4% over a decade. (European Central Bank)

Her proposed architecture is especially relevant to Intelligence Radar: Europe needs more domestic compute, “good enough” models running on European infrastructure, and continued access to the global frontier. She argues that Europe’s data-center capacity shortfall could grow more than sixfold over a decade and estimates closing that gap could cost as much as €600 billion. (European Central Bank)

Radar impact: Available Frontier should no longer mean merely “the user has an API key.” Deployment eligibility should incorporate jurisdiction, infrastructure locality, contractual continuity, provider concentration, model-weight availability and geopolitical access risk. This suggests an additional useful category:

Resilient Frontier — the strongest execution profile whose continued availability remains credible under the requester’s sovereignty and continuity constraints.

That complements rather than replaces Frontier, Efficient and Budget. ECB — A new age of capital: growth, sovereignty and AI

State changes for today

I would make four concrete changes to Intelligence Radar: add Astra’s new Terminal-Bench/DeepSWE evidence without altering the global ranking; add controllability metrics as first-class benchmark constructs; mark AI pacing as a polarized governance question rather than an emerging consensus; and extend deployment economics with sovereignty/continuity risk, potentially introducing Resilient Frontier.

No sufficiently material new biological-intelligence result, independent benchmark release, hardware step-change, or general API price movement crossed the daily threshold in this scan. The main delta on September 15, 2026 is therefore not a new intelligence frontier, but a clearer picture of the conditions under which frontier intelligence can be trusted, controlled and continuously accessed.