Live Feed
AWS Kiro Adds $100/Month Pro Max Tier and Self-Correcting Subagents — Amazon Q Developer Sunsets April 2027
Kiro Pro Max lands at $100/month with 5,000 credits and all frontier models, filling the gap between $40 Pro+ and $200 Power. A self-correcting subagent pipeline also shipped this week. Amazon Q Developer new signups closed May 15; full platform EOL is April 2027.
OpenAI Pulled GPT-5.2 From ChatGPT 61 Days Before Its Documented End-of-Life
A June 10 model picker rollout silently removed GPT-5.2 from ChatGPT despite published deprecation docs listing August 10 as the shutdown date. Pro subscribers at $100/month found out through a missing model option, not an announcement.
3.2 Million Math Records: AI Lifts Completion Speed 23%, Drops Proctored Retention 25%
A study tracking 3.2 million ALEKS math problems across 10 years found AI assistance sped up completion on word problems but reduced proctored retention scores by 25%. Graph problems, which resist AI handoff, showed no effect. The friction that slows problem-solving is also what makes knowledge stick.
Anthropic Launches Claude Corps: $150M to Place 1,000 AI Fellows at 400 Nonprofits
Anthropic is committing $150 million to embed 1,000 early-career fellows inside 400+ nonprofit organizations for a year. No degree required. CodePath and Social Finance operate the program.
Semafor: China-Linked Group Had Already Accessed Mythos Before the Export Order Landed
New reporting ties the White House export control on Fable 5 and Mythos to a specific intelligence finding: a China-linked group had accessed Mythos and could distill its capabilities into a separate model. The jailbreak dispute was a secondary concern.
Google's 10th-Gen Icefish TPU Breaks With Broadcom and TSMC — Samsung and MediaTek Step In
Google is designing its 10th-generation Tensor Processing Unit with MediaTek instead of Broadcom, and discussing manufacturing of I/O dies with Samsung rather than relying exclusively on TSMC. Mass production is expected in 2028. The shift redraws Google's AI chip supply chain for the first time in a decade.
Salesforce Acquires Fin for $3.6B to Own Enterprise AI Agent Customer Service
Salesforce has agreed to acquire Fin, formerly Intercom, for $3.6 billion. The deal puts an AI-native customer service platform inside the CRM giant at a moment when enterprise AI agents are the most competitive space in B2B software.
Intel Foundry Locks 3M+ Google TPU Orders by 2028 as Nvidia Runs 18A Trials for Feynman GPU
Google has booked Intel to manufacture more than 3 million TPUs in 2028, roughly half its projected AI chip output. Nvidia is running multi-project wafer tests on Intel 18A for its next-gen Feynman multi-die GPU. Intel stock jumped 12%. The TSMC-dominated AI chip supply chain has a credible second source.
Apple Ships Five Third-Gen Foundation Models at WWDC26: Gemini Refines the Weights, Nvidia Runs the Cloud Tier
Apple revealed five AFM-3 models at WWDC26: two on-device (up to 20B sparse params), three cloud-based. Four run on Apple Silicon with weights refined via Gemini RL outputs. The top tier, AFM 3 Cloud Pro, runs on Nvidia GPUs in Google Cloud.
Kimi K2.7-Code Ships Open-Source: 30% Fewer Thinking Tokens, 81% MCP Mark Verified, $4/M Output
Moonshot AI's K2.7-Code cuts reasoning token usage 30% over K2.6 and beats Claude Opus 4.8 on MCP Mark Verified real-world tool benchmarks. At $4/M output it is 12x cheaper than Fable 5 — though all headline figures are vendor-run and independent SWE-bench results are pending. A high-speed version at 180 tok/sec launches today.
Cohere Ships North Mini Code: 30B MoE Agentic Coding on a Single H100 Under Apache 2.0
Cohere's first developer-facing model activates just 3B of its 30B parameters per token, fits on one H100 at FP8, and beats models four times its size on agentic coding benchmarks. One catch: it generates three times as many output tokens as comparable systems.
Rio de Janeiro's 'Homegrown' LLM Rio3.5 Is a Model Merge, Not a Training — Qwen3.7 Benchmark Claims Disputed
Rio de Janeiro's city government announced Rio3.5 as a locally trained language model, claiming benchmark performance above Qwen3.7-Max. Analysis reveals the model is a merge of existing open-source weights, not an independently trained system.
HLL Benchmark: AI Agents Fail 10 CAPTCHA Types, Exposing a Gap Static Tests Can't See
A new arXiv paper proposes HLL — Humanity's Last Line — a benchmark where agents must solve interactive human-verification tasks: checkbox grids, image selection, drag-to-verify. Even frontier agents that score well on static benchmarks break down when pages are cluttered, tasks compound, or the system checks whether actions were actually valid.
Anthropic Pulls Fable 5 and Mythos 5 Globally After Refusing to Patch a Jailbreak
Commerce Secretary Howard Lutnick issued an export control directive Friday that treats Anthropic's two frontier models as restricted dual-use technology. Anthropic declined the White House's request to patch the jailbreak, triggering a full global shutdown that blocks even the company's own international employees.
Artificial Analysis Launches AA-AgentPerf: Production Hardware Benchmark for the Agent Era
AA-AgentPerf measures how many concurrent AI agents a provider can serve while meeting real production service levels — not peak throughput. First results show DeepSeek V4 Pro maxing out at Tier 3 (180 tokens/s, TTFT under 3s). GPT-OSS-120B targets a Tier 4 with 2,000 tokens/s and sub-1s first token.