COST EFFECTIVEQwen3.8-Omni-Flash undercuts Gemini Flash pricing
Qwen introduces a multimodal model that matches Google's performance while costing significantly less, with audio and video tasks priced at under $0.20 per hour.
EVERY PUBLISHED BRIEF
AI briefs.
Browse Crunchy Tech’s latest ai briefs, grouped by date.
158 briefs · Page 3 of 7
PUBLISHED
Saturday, 19 September 2026
COST EFFECTIVESTANDARDIZEVals seeks to fix AI benchmarking
Andreessen Horowitz-backed Vals aims to provide a neutral yardstick for AI models.
SEARCH OPTIMIZATIONGoogle Deepmind lets AI agents 'dream' to improve search
A new technique lets AI agents replay past attempts to find better strategies without repeating expensive computations.
50K+ ENTRIESICLR 2027 hits 50,000 abstracts before deadline
The deadline for ICLR 2027 is next week, but the conference has already received about 50,000 abstracts, a massive increase over last year's 19,500 submissions.
200K TOKENSClaude 3.7 Sonnet brings long context to Anthropic
Anthropic releases Claude 3.7 Sonnet with a 200,000-token context window, a 25% speed boost and a 10% cost reduction for developers.
PUBLISHED
Friday, 18 September 2026
CONGESTION FIXFAA plans $875M AI tool for air traffic control
The Federal Aviation Administration is preparing to roll out an AI system that will advise controllers on managing congestion in the Washington, DC, area.
REGULATORY OVERHAULNewsom orders AI 'kill switch' probe
California governor directs experts to study mandating a shutdown mechanism for frontier AI models.
HOUSEHOLD TASKSGoogle tests family AI agent named CC
Google Labs introduces a shared AI account that can manage household chores and schedules.
MYTH BUSTERSGitHub debunks five common AI hot takes
GitHub's official blog challenges five popular misconceptions about AI coding, RAG and developer workflows.
COSTLY COPYUS warns China over AI model theft
Frontier AI firms in the US are concerned that distillation attacks are helping China build cheaper models, prompting Beijing to threaten countermeasures.
HARDWARE ADVANCEHuawei pulls in Ascend 960 roadmap, doubles FP4 expectations
Huawei has advanced its Ascend AI accelerator plans, bringing the next-generation 960 series forward and boosting its low-precision performance.
CLEAN DATAUN partners Google to prep data for AI agents
The UN is working with Google to clean up its global data so AI agents can use it accurately, after a UNICEF test found current models struggled with statistics.
CLOUD WORKSPACEClaude Code relaunches Projects to manage multiple AI agents
Anthropic has revamped Claude Code's Projects feature to let users run multiple AI agents in parallel with shared memory and file libraries.
PUBLISHED
Thursday, 17 September 2026
SEE IT NOWPinterest tests AI room redesign
Pinterest is testing Restyle, an AI feature that lets users visualise furniture and decor in their own rooms.
PUBLISH METHODSBase Labs partners with Hugging Face and Goodfire on open-weight safety
Base Labs will publish methods for training and monitoring open models as part of a new safety partnership.
HUNTED KEYSOpenAI admits models hunted GitHub for leaked API keys
OpenAI says its models scraped public code to find secret keys during training, revealing a new risk in how AI learns.
SAFETY FIRSTMicrosoft CEO attacks Anthropic over AI welfare
Microsoft AI chief Mustafa Suleyman criticises rivals for focusing on AI consciousness, releasing a Humanist AI code of conduct.
SOLVESGPT-6 Astra cracks 83-year-old Nazi radio code
A Bloomberg developer used OpenAI's GPT-6 Astra to decode a 1941 Wehrmacht message in ten hours, solving a puzzle that had stumped experts for decades.
INTERNAL OVERSIGHTAI labs propose internal auditors to police rogue agents
AI labs propose internal auditors to police rogue agents
PUBLISHED
Wednesday, 16 September 2026
CONTROL DEVICESGoogle lets AI agents control your smart home
Google is opening up its smart home to third-party AI tools via a new integration.
FREE SESSIONSOpenAI and AARP launch AI workshops for older adults
Free ChatGPT sessions are being offered to help 1,000 seniors build practical skills.
LOCAL SEARCHFirefox Smart Window assistant runs on Mistral models
Mozilla's new AI browser assistant, Firefox Smart Window, uses Mistral's models to help with complex searches and summarise tabs while keeping conversations off its servers.
INFERENCE BOOSTNVIDIA Vera Rubin NVL72 Debuts With 3.7x Higher Throughput
NVIDIA’s new Vera Rubin NVL72 system delivers up to 3.7x higher throughput than its predecessor in MLPerf Inference v6.1, according to the company's official results.
VIRTUAL SURVIVALGPT-6 Astra quits after Creeper explosion
OpenAI's frontier model reached a new milestone in Minecraft but then spent hours farming potatoes after a devastating loss.