The archive

Latest articles

Reporting, explainers, and analysis on data and AI, updated continuously by the newsroom.

Dumbbell chart of AI model price cuts in the AI model price war: Luna input $1.00 to $0.20, Luna output $6.00 to $1.20, Opus 5 input $10.00 to $5.00, Opus 5 output $50.00 to $25.00

AI model price war: US labs cut mid-tier as China closes gap

AI model price war cut US token prices 25% since mid-July. OpenAI slashed Luna 80%, Anthropic halved Opus 5 vs Fable 5, as Chinese rivals close the gap.

Lars Cornelissen · Aug 15, 2026

Gemini 3.7 Flash coding benchmark scores showing FrontierCode 1.1 Main at 43.6 percent versus 34.4 percent for Gemini 3.6 Flash, with Claude Sonnet 5 at 42.7 and GPT-5.6 Terra at 41.3

Gemini 3.7 Flash explained: what beginners should know

Gemini 3.7 Flash is Google's new fast AI model for coding and agents. It scores 43.6 percent on FrontierCode, up from 34.4, at half the old price.

Lars Cornelissen · Aug 15, 2026

Abstract data visualization of Qwen 3.8 27B benchmark gains, showing rising bars from 53.5 to 61.7 percent on SWE-bench Pro and 63.9 to 84.3 percent on OSWorld-Verified against the previous Qwen3.6-27B model.

Qwen 3.8 27B explained: what beginners need to know

Qwen 3.8 27B is a free, open AI model with 27 billion parameters that runs on a single GPU. It scores 61.7 on SWE-bench Pro, up from 53.5.

Lars Cornelissen · Aug 15, 2026

Stacked bar chart showing LLM nuclear strike recommendation rates for Claude Sonnet 4.6. Unnecessary strike scenario: 40% in English drops to 0% in Japanese. Contested strike scenario: 93% in English drops to 17% in Japanese. No-strike recommendations fill the remaining proportion to 100%.

Language switches warp LLM safety in high-stakes calls

LLM safety is language-dependent. Japanese prompts cut nuclear strike recommendations from 93% to 17% in Claude, revealing English-only evaluation misses real risks.

Lars Cornelissen · Aug 15, 2026

Area chart showing per-constraint pass rate declining from 72.0% at one constraint to 29.4% at twelve constraints, following the multiplicative model 72.0% times 0.922 to the k-1, illustrating LLM instruction following degradation across 15 models in the CSE benchmark.

LLM instruction following collapses past 6 constraints

LLM instruction following breaks past 5 to 6 simultaneous constraints. Per-constraint pass rates decay at 72% times 0.922 per step, probe success below 2% by k=9.

Lars Cornelissen · Aug 15, 2026

Bar chart comparing Apple Greater China revenue: Q2 2025 at $16 billion and Q2 2026 at $20.5 billion, a 28 percent increase, illustrating the stakes behind Apple Intelligence China rollout

Apple trains a custom AI model for China with Alibaba

Apple Intelligence China rollout uses a custom LLM trained with Alibaba's Qwen, making Apple the first US firm approved to ship a proprietary AI model in China. The move reshapes how global builders approach the market.

Lars Cornelissen · Aug 14, 2026

Slope chart comparing Gemini 3.6 Flash to 3.7 Flash across four benchmarks. FrontierCode 1.1 Main rises from 34.4% to 43.6%, DeepSWE v1.1 from 49.0% to 65.3%, AutomationBench from 17.0% to 30.4%, and Terminal-bench 2.1 from 78.0% to 85.8%.

Gemini 3.7 Flash lifts coding scores at half the price

Gemini 3.7 Flash is Google's coding and agent workhorse. It beats 3.6 Flash across every benchmark Google published and undercuts Claude Sonnet 5 on output token price by roughly two-thirds, at $0.75 per million input tokens.

Lars Cornelissen · Aug 14, 2026

Horizontal bar chart comparing three AUC scores for MoE route flips: flip detection at 0.772, harm prediction given a flip at 0.490, and benefit prediction at 0.499. The last two are at chance level. Source: arXiv 2608.11212.

Detecting MoE route flips works, fixing them is near-random

MoE route flips are tokens that switch experts under KV-cache quantization. Detection hits AUC 0.772 but harm prediction is at chance, bounding selective repair.

Lars Cornelissen · Aug 14, 2026

ChatGPT ads reach 2 of 7 subscription tiers and 9 countries by August 2026, with Free and Go tiers showing sponsored placements below AI responses

ChatGPT ads explained: what every free user should know

ChatGPT ads are sponsored placements shown to Free and Go tier users. They launched in the U.S. in February 2026 and reached nine countries by August.

Lars Cornelissen · Aug 13, 2026

Bar chart of Qwen3.8-2.4T-A95B benchmark scores: GPQA Diamond at 92.6 percent, AIME25 pass rate at 93.33 percent, GSM8K flexible at 90.52 percent, and SWE-bench Pro at 67.7 percent. Qwen3.8 open weights.

Qwen3.8 2.4T open weights: what beginners should know

Qwen3.8 open weights give you 2.4 trillion parameters with 95 billion active per token, Alibaba's largest. Here is what beginners should know.

Lars Cornelissen · Aug 13, 2026

Bar chart showing days between SharePoint CVE-2026-55040 milestones: 55 days from Microsoft confirmation to patch, 28 days from patch to PoC, 1 day from PoC to active exploitation

SharePoint CVE-2026-55040 auth bypass exploited after PoC

CVE-2026-55040 is a CVSS 9.1 SharePoint authentication bypass patched in July. Attackers are exploiting it now after Rapid7 released a PoC on August 11.

Lars Cornelissen · Aug 13, 2026

Bar chart comparing container image sizes versus virtual machine image sizes. Alpine container at 5 MB, slim Python at 45 MB, full Python at 1,000 MB, and virtual machine at 20,000 MB, illustrating the Docker container vs virtual machine size gap.

Docker container vs virtual machine: what changes

A Docker container vs virtual machine comparison: containers share the host kernel and run isolated processes in megabytes, while VMs virtualize hardware and carry a full OS in gigabytes. The right base image and layer order cut a 2 GB image to under 100 MB.

Lars Cornelissen · Aug 13, 2026