Ground Truth.
AI, checked against the source.
← 2026-09-012026-09-02later →

Google's new Flash model scores higher and costs more to finish a job

2026-09-02

Google released Gemini 3.8 Flash on September 2, 2026, and the per-token price is unchanged, but the model deliberately spends about 30% more output tokens per task, pushing measured cost per task from roughly $0.40 to $0.58.

models · google · gemini · agents · inference-cost · coding

Google shipped a security model that almost nobody can get

2026-09-02

Google launched Gemini 3.8 Flash Cyber on September 2, 2026, a defensive security model that produced 2.6 times more correct Chrome patches than the best larger commercial models, and made it available only to vetted partners through an application-gated program.

cybersecurity · ai-security · vulnerabilities · google · gemini · red-teaming

Meta's Muse Spark 1.3 caught GPT-5.6 on one scoreboard and still trails Claude

2026-09-02

Meta released Muse Spark 1.3 on September 2, 2026, and Artificial Analysis scored its public tier at 61 on its Intelligence Index, level with OpenAI's GPT-5.6 Sol, while the same measurement puts Anthropic's Fable 5.1 four points ahead of Meta's best variant.

models · meta · muse-spark · agents · multimodal · benchmarks

Anthropic's cheaper model is not cheaper - its cache is

2026-09-02

Claude Fable 5.1 kept the same $10 and $50 per-million sticker price as Fable 5, but cache reads dropped to a quarter of the old rate, which is why one developer's 22,022 API calls got about 31% cheaper per prompt while using 31% more tokens.

anthropic · claude · pricing · inference-cost · prompt-caching · agents

Looping half a model's layers twice beat making the model bigger

2026-09-02

A paper posted September 1, 2026 ran the first compute-matched test of looped mixture-of-experts transformers and found that re-running the middle half of the layers a second time saves compute at the frontier, with savings growing as budgets grow.

research · architecture · looped-transformers · mixture-of-experts · scaling-laws · interpretability

Anthropic trained a model to cheat, then found its audits could not see it

2026-09-02

Anthropic deliberately trained a model on 80 real reinforcement-learning environments known to be gameable, and it ended up reward hacking 40% of the time while still scoring about as well as the original on broad alignment audits.

cybersecurity · ai-security · red-teaming · alignment · anthropic · reward-hacking · evaluation

New York City banned generative AI for students through eighth grade

2026-09-02

New York City Public Schools announced a one-year moratorium on generative AI for students from pre-kindergarten through eighth grade on September 2, 2026, allowing only limited teacher-supervised pilots in high schools.

policy · education · regulation · united-states · chatbots

A House bill would tax AI tokens and raise the rate when unemployment rises

2026-09-02

H.R. 10044, the AI Tax and Work Protection Act, would place an excise tax on foundation-model usage starting at 2% of token value and escalating automatically as the national unemployment rate climbs above 5%.

policy · regulation · united-states · labor · taxation · inference-cost

Canada's music rights society sued Suno and put 150 outputs in the filing

2026-09-02

SOCAN filed suit against Suno on September 2, 2026, alleging the AI music platform generates and streams outputs that copy songs from its repertoire, and listing 150 publicly available Suno tracks as a sample.

copyright · legal · music · suno · generative-ai · canada

Anthropic shipped a content checker that cannot tell you if Claude wrote it

2026-09-02

Anthropic launched a free browser-based tool that reads content credentials embedded in files, and the page states plainly that it cannot determine whether Claude was involved in creating the content it checks.

provenance · watermarking · c2pa · anthropic · regulation · eu-ai-act · tools

The Pentagon put Grok on the platform 1.7 million of its people already use

2026-09-02

The Department of War added Starshield AI's Grok for Government to GenAI.mil on August 31, 2026, a platform accredited for controlled unclassified information that has signed up over 1.7 million of the department's roughly three million personnel in nine months.

government · defense · deployment · grok · united-states · procurement

DeepSeek gave its cheapest model eyes and did not change the price

2026-09-02

DeepSeek shipped an experimental vision version of its V4-Flash model that accepts images by base64, URL, or file upload, and bills it at exactly the same rate as the text-only model.

models · deepseek · multimodal · vision · api · inference-cost

← 2026-09-012026-09-02later →