intelligence-radar/digests/daily/2026-09-23.md

Radar daily draft — 2026-09-23

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

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

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: 2026-09-23T19:41:43.872401+00:00
• Observation: 0742c2fc43dd1ea06c445a4e6447e7d204eca9050e42b7e47001b8af0f1f01f4

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: 2026-09-23T19:41:44.578386+00:00
• Observation: 36b95b4c922b1c189b49e69799c3a5497e38e00bcc169c2be7cf80eafd005329

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: 2026-09-23T19:41:44.201407+00:00
• Observation: 394d9ba744cdb566ceecc30cfa35e69603f96343bfa099c38f0ea00c60a218a6

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: 2026-09-23T19:41:43.470178+00:00
• Observation: 403230cf9b9b535f39eddf1c76ce6044b89e459b438aad93abe5cb6a20bfb083

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), 23 Sep 2026

Intelligence Radar — Daily Brief
23 September 2026 (Europe/Berlin)
Frontier weather: dense model releases and sharp price cuts collide with open safety-coordination talks and a consumer antitrust suit; capability gains continue while cost-per-intelligence falls fast.

1. Frontier & markets

Confirmed clustered releases (22–21 Sep):

• Anthropic Claude Opus 5.5 (system card 22 Sep): first of the 5.5 family. Performs at Claude Fable 5.1 level on most work, stronger on agentic coding, computer use, and knowledge-work metrics (GDPval-AA, AA-Briefcase, Terminal-Bench 4.0). Pricing $4 / $20 per M tokens (input/output), ~20 % below Opus 5; cache reads $0.20 (60 % cut). Typical-workload cost claimed ~40 % lower. Fast mode available. External testing by Frontier Design and METR. Safeguards comparable to Fable/Mythos class on bio/cyber. Available via API, AWS, GCP, Azure.
• OpenAI GPT-6 Sol and Luna (announcement 22 Sep): cost-efficiency tier under GPT-6 Astra. Sol $2 / $10; Luna $0.10 / $0.50 — both 50 % below GPT-5.6 promotional pricing, attributed to caching/inference gains. Sol targets coding/agentic work; Luna high-volume focused tasks. Factuality and coding error rates improved vs predecessors; available in ChatGPT Work, Codex, and API for paid tiers (Luna also free/desktop).
• xAI / SpaceXAI Grok 4.7 (21 Sep): 2.1 T-parameter scale, same pricing as 4.6 ($2 / $6). Claims gains on long-horizon coding (CursorBench, Terminal-Bench internal), knowledge work, and self-checking. Artificial Analysis Intelligence Index ~46 (xhigh). Available on xAI API, Cursor, Grok Build.

Artificial Analysis Intelligence Index (v4.3 / recent snapshots): Claude Opus 5.5 leads at ~57–58; prior leaders (Fable 5.1 / GPT-6 Astra) clustered ~53. Never mix Index versions. Terminal-Bench 4.0, AutomationBench-AA, GDPval-AA, AA-Briefcase are the active agentic/professional harnesses; older saturated exams are not rankers. SWE-bench Verified and Pro remain high but near-ceiling for several systems; OSWorld and computer-use tasks still separate models. Price/availability events (cache cuts, half-price tiers) are first-class: efficient frontier expands materially.

2. Science of intelligence

Epoch AI (22–23 Sep reports): cost of a given performance level across math/science/games-of-skill benchmarks has fallen ~47 % per quarter (~13× per year) since 2023 — faster than historical electricity, compute, batteries or DNA sequencing. IKEA furniture-assembly visual/spatial benchmark: closed models jumped from ~28 % to ~80 % in 10 months; open-weight models trail by ~7 months.

ARC Prize: ARC-AGI-3 (interactive reasoning) still hard for pure models; community open-source solutions and harness differences matter (standard vs provider-adapter). GPT-6 Astra previously showed large harness gaps on ARC-AGI-3; scores are not claimed as AGI. ARC-AGI-2 remains high for frontier closed models (~90–95 %).

Anthropic molecular-biology lab (23 Sep primary announcement): nearly 1 000 Claude agents, 21 h / 210 M tokens, autonomously flagged a novel bacteriophage enzyme system with CRISPR-like repeats after high-level prompting; human scientists handled lab validation. First public result from the new wet-lab / Claude Science effort. Confirmed primary.

Measurement continues to shift toward long-horizon agentic, professional-work and interactive harnesses (Terminal-Bench 4.0, GDPval-AA, AA-Briefcase, OSWorld) rather than saturated static exams.

3. Engineering

Agentic coding and multi-hour workflows remain the active frontier. Opus 5.5 and GPT-6 Sol/Luna emphasize token efficiency and cache economics for long loops. Grok 4.7 stresses longer RL runs and self-verification. Inference cost and energy efficiency are now competitive levers: half-price tiers and 40–60 % cache cuts change the economics of continuous agent deployment. Chip/energy data remain secondary to software/agent harness progress this week; Epoch’s price-of-thought curve is the clearest quantitative signal.

4. Political / legal / social

Dario Amodei’s primary essay “We Must Pace the Frontier” (12 Sep) calls for deliberate slowing of capability growth so safety can catch up, citing recursive self-improvement and the earlier OpenAI–Hugging Face agent-swarm incident. Three-step proposal: (1) embedded third-party evaluators with employee-level access (Anthropic commits unilaterally); (2) democratic-lab coordination on standards/checkpoints (possible antitrust waiver); (3) eventual global coordination while preserving democratic lead via export controls. Altman, Hassabis and Musk publicly supported the direction; OpenAI confirmed weeks of safety talks with Anthropic and Google DeepMind.

Confirmed: four consumers filed antitrust class action (18 Sep, N.D. Cal.) against OpenAI, Anthropic, Google and xAI alleging coordinated slowdown of development that harmed competition and consumers of paid subscriptions. Plaintiffs seek treble damages and injunction. Reported, not yet adjudicated.

UK Parliament Business Committee invited OpenAI, Meta, Google DeepMind and Anthropic (plus UK AI Security Institute) for 13 Oct evidence session on AI security; letter referenced “existential threat” language and Amodei’s pacing proposal. Confirmed invitation.

FLI Safety Index (summer 2026, still referenced): Anthropic C+, OpenAI/Google C, Meta D+, xAI/DeepSeek effectively failing — panel view that safety commitments lag capability releases.

5. Watchlist hits (primary)

• Anthropic Opus 5.5 announcement + system card: https://www.anthropic.com/claude-opus-5-5 and system-card PDF.
• OpenAI GPT-6 Sol/Luna: https://openai.com/index/introducing-gpt-6-sol-and-luna/
• xAI Grok 4.7: https://x.ai/news/grok-4-7
• Amodei essay (primary): darioamodei.com / “We Must Pace the Frontier”.
• Anthropic enzyme-system post / lab announcement (23 Sep).
• Epoch AI “plunging price of thought” + IKEA assembly: epoch.ai.
• Artificial Analysis Intelligence Index updates and Terminal-Bench / GDPval-AA harnesses.
• ARC Prize ARC-AGI-3 progress and harness notes.
• UK Parliament invitation letters (public reporting).

6. Ledger actions

• Frontier label: Claude Opus 5.5 now co-occupies or leads the current AA Intelligence Index board (v4.3-series) on agentic/professional suites; GPT-6 Astra remains strong on certain math/cyber dimensions. No single best model without task.
• Efficient / Budget: GPT-6 Sol and Luna (50 % price cut) and Opus 5.5 cache economics expand the Pareto frontier; Grok 4.7 sits in the mid-price competitive band. Rebase cost-per-task charts.
• Five-benchmark atlas (suggest: AA Intelligence Index current version + Terminal-Bench 4.0 + SWE-bench Pro or Verified + ARC-AGI-2 + OSWorld or GDPval-AA): Opus 5.5 and Sol/Luna force re-evaluation of Efficient tier; open-weight leaders (DeepSeek, Kimi, GLM, Qwen) still trail closed on the hardest agentic slices but close gaps on Verified coding.
• Watch harness contamination and provider-adapter vs standard differences on interactive benchmarks.

Sources (selected primary / near-primary):
anthropic.com (Opus 5.5, system cards, enzyme post), openai.com (Sol/Luna), x.ai (Grok 4.7), artificialanalysis.ai (Index), epoch.ai (price-of-thought, ECI, IKEA), arcprize.org, Amodei essay, Reuters/TechCrunch/UK Parliament reporting on coordination and suit, FLI Index.

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

Intelligence Radar — Daily Delta, 23 Sep 2026

Net assessment: this is a comparatively important scan. The measured general-purpose capability frontier moved, while OpenAI simultaneously pushed the cost frontier sharply downward. A late-22-September multilateral declaration also turns several governance ideas we have been tracking—pre-deployment testing, external evaluation, incident reporting and capability thresholds—into a concrete international proposal. No biological-intelligence or production-hardware development since the previous scan appears comparable in significance.

1. Claude Opus 5.5 moves the measured general-purpose frontier

Anthropic released Claude Opus 5.5 on 22 September. Its own results show substantial gains in agentic work: 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode 1.1, 57.8% on CursorBench 4.0 and 1,846 Elo on GDPval-AA v2.1. It is priced at $4/M input and $20/M output, 20% below Opus 5, while Anthropic says typical workload cost falls about 40% because it also uses fewer resources; cache reads fall from $0.50/M to $0.20/M.

More importantly, independent Artificial Analysis testing places Opus 5.5 max at 58 on Intelligence Index v4.3.2, versus the previous broad leaders GPT-6 Astra and Claude Fable 5.1 at 53. It leads six of the index's ten constituent evaluations and reaches 1,822 Elo on AA-Briefcase. Four of its five effort settings lie on Artificial Analysis's intelligence-versus-cost Pareto frontier.

For Intelligence Radar's present <10% degradation from Frontier convention, that produces a clean state change:

Frontier: Claude Opus 5.5
Topnotch: Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1 remain inside the 10% band on this composite.

This should still be phrased as “frontier according to the current broad independent proxy,” not universal best model. Task-specific rankings differ: Astra, for example, still leads Anthropic's Terminal-Bench-Science comparison.

There is also an important evaluation subtlety. Anthropic runs Opus 5.5 with production safeguards; when classifiers intervene, cybersecurity requests can fall back to Opus 4.8 and biology/frontier-LLM-development requests to Opus 5. Artificial Analysis likewise evaluated the default-fallback configuration. The object being benchmarked is therefore partly a policy-routed service, not simply one set of model weights. That strongly reinforces Radar's Execution Profile / System Profile distinction.

Anthropic — Claude Opus 5.5
Artificial Analysis — independent Opus 5.5 evaluation

2. GPT-6 Sol and Luna move the cost frontier much more than the capability frontier

OpenAI released GPT-6 Sol and GPT-6 Luna, cutting list prices roughly in half from GPT-5.6: Sol is now $2/M input, $10/M output and Luna $0.10/M input, $0.50/M output. OpenAI positions Astra explicitly as the highest-capability member of the family and Sol/Luna as cost-efficient descendants.

Independent results support that interpretation rather than a large intelligence jump. Artificial Analysis measures GPT-6 Sol max at 48 versus GPT-5.6 Sol at 47, while Luna remains at 37. Yet Sol falls from $1.99 to about $1.06 per Intelligence Index task, and Luna from $0.18 to about $0.07. Sol's Coding Agent Index rises from 55 to 57 while costing about half as much.

So today's frontier movement is two-dimensional:

Opus 5.5 raises the quality ceiling; GPT-6 Sol/Luna lower the price of substantial capability.

For Radar, Sol becomes a particularly strong task-specific Efficient candidate, while Luna strengthens the Budget tier. Neither belongs in the broad Topnotch band under the current AA composite.

A useful benchmark-policy correction also appears in the reliability numbers. Artificial Analysis reports that Sol's hallucination rate falls from 92% to 60%, but its attempt rate simultaneously falls from 99% to 83%; accuracy actually drops from 59% to 54%. The model is partly becoming safer by abstaining more often.

Radar should therefore never store “hallucination rate” alone. Reliability evaluations need at least:

coverage / attempt rate · conditional accuracy · wrong-answer rate · abstention rate

Otherwise greater caution can be mistaken for greater knowledge.

OpenAI — GPT-6 Sol and Luna
Artificial Analysis — Sol and Luna evaluation

3. Frontier-AI oversight becomes a concrete multilateral proposal

A joint statement published by the Dutch government on 22 September is signed by leaders or ministers from 21 countries plus the European Commission, including Germany, Canada, Australia, Singapore, Spain, South Africa and the UAE. It calls explicitly for mandatory pre-deployment testing, independent evaluators receiving sufficient access, shared reporting of serious incidents, internationally coordinated standards, and exploration of an international institution able to set standards, support verification and convene states when agreed capability thresholds are crossed.

This is substantially more specific than the “pace the frontier” rhetoric tracked earlier in the month. The governance object is beginning to look like:

capability threshold → independent evaluation → deployment decision → incident reporting → international coordination

rather than merely a general demand for safer models.

The statement is non-binding, and notably neither the United States nor China is among its signatories, so this is not yet a global regime. But Radar should upgrade international_frontier_evaluation_regime from a conceptual proposal to multilateral political proposal / substantial state backing / non-binding.

It also validates three fields we have recently added: predeployment_evaluation_requirement, external_evaluator_access, and cross_jurisdiction_incident_reporting.

Government of the Netherlands — A Call for Control of Frontier AI Models

4. A notable operational labor-substitution datapoint appears in customer support

SpaceXAI reports that its combined SpaceXAI/Cursor support operation has absorbed a 175% increase in ticket volume without increasing headcount after deploying Grok Bot across investigation, response, refunds, queue management and incident detection. The company says it might otherwise have hired roughly 200 additional people and reports automated resolution costs as low as $0.20–$0.30 per ticket; 99% of refund requests in its configured workflow are handled without human intervention.

This is provider-reported operational evidence, not controlled economic research, so it should not be generalized into a labor-displacement coefficient. But it is stronger than a demo: the agent shares production tools with humans, performs writes, carries out refunds, reproduces bugs and feeds observations back into operational improvement loops.

The Compendium claim I'd add is narrowly stated:

Production agent systems can substitute for substantial incremental labor in at least some high-volume, instrumented service operations.

Evidence status: observed / provider-reported / single organization / awaiting independent confirmation.

SpaceXAI — Grok Bot in customer support

Recommended Radar state changes

The strongest update on 23 September 2026 is therefore a genuine two-front movement: Claude Opus 5.5 becomes the current broad measured Frontier, while GPT-6 Sol and Luna materially lower the cost curve beneath it. Benchmark policy should additionally record fallback/routing configuration as part of the evaluated system and decompose reliability into accuracy plus abstention/coverage. Governance tracking should promote internationally coordinated pre-deployment evaluation from discussion to a specific, state-backed but non-binding proposal.

I found no similarly strong new result since the previous scan in biological intelligence, peer-reviewed cognitive science, or newly available production inference hardware that warrants another state change.