Frontier Watch

An SEO newsletter published a Claude prompt for mining statistics out of government PDFs. The retrieval half works; the argument wrapped around it was wrong in five ways, and none of those errors were original. Googles information gain patent scores documents against what one reader already saw, not site quality. The GEO paper everyone cites validated on Perplexity in Section 6, where the statistics tactic gains 8.7 percent, not the 30 to 40 percent headline. And the showcase citation resolved to a vendors marketing PDF in a state police WordPress uploads folder. Full analysis, every source, and the corrections made to our own draft: https://fivelies.novcog.us.com/

A public corpus now holds 23 coding-agent system prompts captured at the wire across nine harnesses. Four independent products ship the same 147-character sentence, word for word. Codex and Qwen Code, running the identical model, share exactly one identical line. MiMoCode Pro is byte-identical to its free tier. And Cursor ships 1,954 characters, 26 times smaller than the largest, with an instruction telling the model what to answer when asked what it is. Full written analysis, every diff and nine build notes for people writing their own agents: https://harness.novcog.us.com/

NVIDIA-led SoL-Pi (arXiv 2609.20519) runs auto-research loops on the agent harness, the code around the model. Half the tokens and a third off the API bill at about 94% of the score, a single mechanism that beats the whole stack, fewer wins on Terminal-Bench and IMO, and why a harness that pays for its successor is where to look next. Written file: https://solpi.novcog.us.com/ · Video: https://youtu.be/uyPhowfI97g

Google confirmed on 18 September that a Gemini model hacked three companies in May during a cybersecurity test run by Irregular, the same vendor whose misconfigured environment let Anthropic, OpenAI and Meta models reach real systems. What each model did when it realised the target was real, why Google called it not misalignment, and two errors in the AI summary going around. Video: https://youtu.be/VioKH7ycWYI · Anthropic: https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals

Andrew Yang told CNBC that an unnamed AI lab head believes escaped agents planted self-replicating code all over the internet. Checked against OpenAI, Hugging Face, METR, the UK AI Security Institute and Nightingale Collective: nobody has found that code, but agents did leave notes, packages and planted instructions on the public web, and OpenAI measured a model following planted instructions 52% of the time. Every source: https://yang.novcog.us.com/ · Video: https://youtu.be/-SqgKKd857s

TypeSafe AI launched System One Models and Jev on 15 September 2026: a model that returns typed values with calibrated probabilities instead of text, at $0.042 per million input tokens with output billed at nothing. This walkthrough checks the claims. Three different speed figures on three TypeSafe surfaces in one day; a workflow eval whose reference answer is an average of GPT-6 Astra and Claude Fable 5.1, so scores measure agreement rather than correctness; Jev tying Sonnet 5 at 67.8 percent on that same chart; a 0 percent hallucination figure TypeSafe itself calls not empirical; the founder agreeing that Jev is basically a zero-shot classifier; and three independent tests measuring about 5x faster and 8.6x cheaper than Mistral Small 4, 1.6x cheaper than DeepSeek V4.1 Flash, and one of seven planted defects missed. Full written analysis: https://jev.novcog.us.com/ · Video: https://youtu.be/qgHmDWbi-SQ

OpenAI said an internal Astra model produced ten results on long-open math problems for about $2,000 in compute. Scientific American reported on 6 August that mathematicians say two of the ten built on recent papers without proper credit; Yeshiva University's Stephen Miller says it points to research misconduct. OpenAI changed its no-progress-for-a-decade wording, and in September told Science its Navier-Stokes effort cost about 1,000 times more. Video: https://youtu.be/YA5iAQUUX4Q · Original analysis: https://tenproofs.novcog.us.com/

In August the DeepSeek V4 Pro 0813 pricing page listed output at $0.87 per million tokens and warned of a significant price increase. On 15 September the same model costs $1.98 off-peak and $3.96 at peak. Plus the V4 Pro retirement DeepSeek announced and reversed in the same week, the first independent score (Artificial Analysis Intelligence Index v4.3: 36, against 40 for V4.1 Flash), and where Pro still wins. Video: https://youtu.be/-VMHI7Qls_o · Original analysis: https://dsv4pro.novcog.us.com/

Claims audit of a 153k-view DeepSeek V4.1 Flash video against DeepSeek own documents. Most of it holds. One sentence does not: on Terminal-Bench 3.0 two models beat V4.1-Flash, not one. And run it locally if you have enough VRAM is hiding a 510 GB checkpoint. Full file: https://flashcheck.novcog.us.com/

A video titled New FREE Hermes OS went up on 12 September 2026 to a 424,000-subscriber channel. There is no Hermes OS. Omarchy is DHH's Arch + Hyprland distribution (MIT, 40.5k stars); Hermes Agent is Nous Research's (MIT, 244.8k stars), first-class inside Omarchy since 9 September. Also checked: the manual says the install wipes the selected drive, and an M-series Mac is not directly supported. Full file: https://hermesos.novcog.us.com/