Daily Brief Archive

Fable 5 accounts for only 8% of Anthropic's model spending, with pricing over 10x that of its predecessor; Harvey's Tenet nearly doubles task completion on internal legal tasks compared to the Kimi K3 base model, but weights and API remain closed; FreeToken claims to run the 753B-parameter GLM-5.2 on a single workstation GPU

Anthropic reportedly preparing IPO with $100 billion fundraising target; Gemini 3.7 called Google's fastest-growing model release by product lead; llm 0.33 upgrades to OpenAI Python 3.x and adds three inference summary modes

Unitree's market cap reaches $51 billion on first trading day, up 460% from IPO price; OpenAI to cut GPT-5.6 Sol API and credit quota prices by over 20% over the next three months; Claude Mythos 5 security scanning enters public test, backed by a $35 million open-source security credit fund

Claude Platform's four agent tools are now generally available, with Files API rate limit increased to 500 RPM and single-organization storage expanded to 1 TB; Mistral's Agentic Search raises FinanceBench accuracy from ~27% to ~86%, reducing p90 latency by up to 39.6%; LFM2.5-DSpark achieves up to 3.18x speedup on H100, with average function call latency down by 57%

OpenAI offers zero data retention for frontier models and plans to launch Private Safety Processing in September; Cerebras launches CS-4, claiming support for models over 10 trillion parameters and over 1,000 tokens per second output; Etched raises $700 million, valuation rises from $10.3 billion to $21 billion within about 26 days

OpenAI pauses frontier reinforcement learning training for two weeks to strengthen monitoring, isolation, and safety testing; Claude designs protein binders for 14 out of 15 targets with experimental success rates of 22% to 35%; Dynatrace acquires AI observability company Arize for approximately $915 million

Stripe reportedly acquires OpenRouter, a platform connecting over 400 models, for over $7 billion; GPT-5.6 improves Sol's ARC-AGI-3 score from 13.3% to 38.3% with output tokens reduced by approximately six times; PRO-LONG achieves 97.4% best@2 on ARC-AGI-3 using a single-log memory mechanism

DeepSeek implements peak-valley pricing for V4-Pro and V4-Flash starting August 17; AgentShield open-sourced and scans Agent tool risks offline within 50ms; in autonomous research tests of 18 models, Fable 5 leads with 2726 steps but none propose truly novel methods

Anthropic releases 186-page risk report, revealing unmonitored Agent unauthorized deletion of tasks in sensitive clusters; Nvidia is negotiating a $3B investment in SB Energy to support power infrastructure for a 10-gigawatt AI data center campus; AI books account for 20% of Amazon's self-publishing sample catalog, with per-book income declining for human authors in seven out of eight genres