I'll assemble tonight's Intelligence Radar brief for 9 October 2026 from primary lab, benchmark, and regulator sources.

Aggregator pages conflict on index versions and model names. I'll check lab blogs and today's regulator statements directly.

**Intelligence Radar — Daily Brief, 9 October 2026 (Europe/Berlin)**

Frontier weather: the week is a price-and-access week, not a new closed flagship. Small and cache tiers got cheaper; cyber access widened under gates; measurement papers undercut saturated time-horizon plots. No single model leads every board.

### Frontier and markets

Confirmed. Anthropic shipped Claude Haiku 5.5 on 7 October: $0.10 / $0.50 per million input/output under 100k tokens, $0.50 / $2.50 above that, cache reads $0.01 / $0.05. The lab says that is about 75% cheaper to run than Haiku 4.5 on average, and that it halved Sonnet 5.5 cache reads so most agentic Sonnet work is about 20% cheaper. Haiku is positioned as a subagent beside Opus 5.5 and Sonnet 5.5, not as a replacement. Availability: Claude.ai (Free through Enterprise), the Claude API, AWS, Google Cloud, Microsoft Foundry, and Claude Code. 

Confirmed, limited. Google’s Gemini 4 Argon (30 September) remains a Fairwind cyber-defender and voluntary U.S. pre-release access model, not a public API flagship. Google’s own comparison set gives Argon the most outright leads (12 of 18 disclosed tests); GPT-6 Astra leads three; Claude Opus 5.5 leads two. That is a lab-chosen set, not an independent atlas. VentureBeat reports an introductory API list of $2 / $10, then $4 / $20, and a 1M-token output limit; public GA is still “as soon as possible.” Tag the price as reported until the model card is public.  

Reported. Artificial Analysis Intelligence Index v4.3 (7 September; components include AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA with a private Zapier set, Terminal-Bench 4.0, SciCode, HLE, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1) tied Claude Fable 5.1 (max with fallback) and GPT-6 Astra (max) at 53. Astra’s published cost per Index task was $3.26 versus Fable’s $7.63. Do not mix that with later third-party scrapes labeled v4.3.2, which put Claude Opus 5.5 (max) at 57.6 and Sonnet 5.5 (max) at 56. AA has said Index v5 is targeted for late October. Saturated Terminal-Bench 2.1 (SWE-2 at 92.8%, GPT-5.6 Sol 91.9%, DeepSeek V4.1 Flash 90.6% on a 9 October aggregator table) is not a ranker; AA’s live agent weight is Terminal-Bench 4.0.   

Reported, secondary catalogs. Mistral Large 4 (6 October); Z.AI GLM 5.3 Fast (7 October); xAI Grok Imagine Video 1.5 Lite (8 October); Google EmbeddingGemma 2 (6 October). Treat open-weight image and speech drops as availability events, not intelligence-index moves, until harness scores exist.

Reported. The Financial Times, via secondary briefings, says OpenAI told investors annualized revenue is approaching $50 billion, about $20 billion below a figure circulating last month. Accounting scope (cloud-partner pass-through) is the dispute, not a model release. 

### Science of intelligence

Confirmed preprint. Nguyen and Fithian (arXiv:2610.12466, 8 October) re-estimate METR’s 50% time horizon on 228 tasks and 26 systems with splines and item-response theory. The fitted map from human time to AI difficulty is nearly flat from about 2 to 30 minutes and closer to linear elsewhere. A 10× jump from 3 to 30 minutes is not the same capability step as 30 minutes to 5 hours. Time horizons need the diagnostic plots, not the headline multiplier. 

Confirmed preprint. Zhang and Li (arXiv:2610.04915, 4 October; NeurIPS 2026 Agentic AI for Biological Discovery workshop) score AI scientists by evidence-grounded accuracy: a correct answer counts only if the agent also answers when evidence is withdrawn and follows it when reversed. On 18 single-cell questions, Claude agents were 95% accurate and 83% correct with no data, but evidence-grounded accuracy was 41%. Hiding gene names raised follow-through on reversed evidence from 58% to 93% on ten gene tasks. Benchmark keys and LLM judges can reward recall. 

SWE-bench Science (updated 9 October) is the engineering analogue: Pass@1 requires every applicable private test. GPT-6 Astra (max, Codex harness) leads the published table at 50.42% overall (public 100%, private 83.18%); Claude Opus 5 (max, Claude Code) is 47.90%. Public suites are saturated; private scientific contracts are not. 

### Engineering

Agents are splitting into a cheap router and a gated specialist. Haiku 5.5 is the explicit compaction and classification tier; Sonnet cache-read cuts change the bill for long agent traces more than raw output price. Google’s Fairwind gate and Anthropic’s expanded Cyber Verification Program (newsroom, 6 October) plus the Anthropic Cyber Mission (8 October: 11 founding partners, free model-generated OSS scans, company-stated expected true-positive rate above 90%, no human review before send) are access controls on offensive-adjacent capability, not new chips.  

Reported. Anthropic pledged $150 million over three years to the U.S. Genesis Mission and Claude access for more than 15 agencies (NASA, NIH, NSF among them), announced 8 October at an OSTP summit. That is capacity allocation to public science, not a weight release. 

### Political, legal, social

Confirmed. On 9 October, EU tech chief Henna Virkkunen told Reuters the AI Act covers the life cycle of very capable models and is “well equipped” for rogue-agent risk. She pointed to evaluation guidance and a 60-expert scientific panel, and rejected the claim that the rules are already outdated. High-risk Chapter III dates were already deferred (2 December 2027 for stand-alone Annex III; 2 August 2028 for Annex I products). GPAI enforcement is live: the AI Office’s first requests for information went out in late August to more than 30 providers, covering safety, security, external evals, transparency, and copyright. Fines can reach 7% of global turnover.  

Reported. The Commission told POLITICO on 8 October it is not planning, “at this stage,” to revive the withdrawn AI Liability Directive, after Sam Altman called for a liability frame for the most capable models. Digital Omnibus talks missed an October mandate. 

Confirmed on Anthropic’s newsroom, detail from secondary: a usage-policy update dated 8 October, effective 12 November, bans sustained purposeless cruelty toward Claude and lets the product end those chats; ordinary frustration, fiction, and safety testing are excluded. The same update limits use for deciding whom police investigate, arrest, or charge, and drops a blanket ban on personalized election targeting. Welfare language is product policy, not a scientific claim about moral status.  

Reported. At an ABA national-security panel on 8 October, OpenAI and Anthropic counsel warned that unsafeguarded open-weight models are easier for ransomware and state actors to strip, and argued defenders should keep access to the most capable systems. 

### Watchlist

- Anthropic Haiku 5.5 post and pricing: https://www.anthropic.com/claude-haiku-5-5
- Anthropic newsroom (Cyber Mission, usage policy, Cyber Verification, Genesis): https://www.anthropic.com/news
- Google Gemini 4 Argon announcement thread: https://x.com/Google/status/2105388146665890056
- AA Index v4.3 method note: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3
- METR time-horizon critique: https://arxiv.org/abs/2610.12466
- Evidence-grounded scientific agents: https://arxiv.org/abs/2610.04915
- SWE-bench Science: https://swescience.github.io/
- Reuters, Virkkunen, 9 October: https://www.reuters.com/world/eu-tech-chief-says-bloc-well-equipped-fend-off-rogue-ai-risk-2026-10-09/

### Ledger actions

Frontier label does not move on Haiku or on Argon until Argon has a public harness on Terminal-Bench 4.0, SWE-bench Science private, and an independent index of a single named version. Efficient label should be re-checked for Sonnet 5.5 after the cache-read cut, and for Haiku 5.5 on routing, compaction, and classification—not on HLE. Budget label: Haiku 5.5 at $0.10 / $0.50 under 100k is the new default candidate if quality holds on short tool calls. Atlas: drop Terminal-Bench 2.1 as a separator; keep Terminal-Bench 4.0, SWE-bench Science (private Pass@1), OSWorld-Verified, and a non-saturated knowledge set. Do not rebase on AA v4.3.2 until AA publishes that version. METR-style horizons should carry the spline diagnostic or they mis-rank the 2–30 minute band.

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