Live Feed
Agents Don't Fail at Tool Calls. They Fail at Tool Chains — New Survey Quantifies the Orchestration Gap
A comprehensive survey of multi-tool LLM agents finds that single-call accuracy is no longer the binding constraint. The gap is in coordination: planning, state management, recovery, and reliable execution across chains of 10 or more tool calls.
Fine-Tuning RAG Embeddings for Precision Silently Cuts Retrieval Accuracy by 40%, Redis Research Finds
A new paper from Redis shows that training dense embedding models to distinguish near-identical sentences reduces broad retrieval generalization by up to 40%. Enterprise teams optimizing RAG for precision may be quietly degrading the pipelines they depend on.
Claude Gets 9 Creative Tool Connectors: Blender, Adobe CC, Autodesk Fusion, Ableton, and More
Anthropic launched nine native connectors that put Claude inside the software creative professionals already use. Blender gets a Python API bridge, Autodesk Fusion gets conversational 3D modelling, and Adobe Creative Cloud exposes 50-plus tools across Photoshop, Premiere, and Express.
Google Signs Classified Pentagon AI Deal for 'Any Lawful Government Purpose' With No Veto Clause
Google has signed a classified agreement giving the US Department of Defense access to its AI models for any lawful government purpose, including classified intelligence work. The deal contains no provision allowing Google to veto specific uses of its AI.
OpenAI Brings GPT-5.5, Codex, and Managed Agents to AWS Bedrock in Limited Preview
OpenAI and AWS have expanded their partnership to put the full OpenAI stack inside Amazon Bedrock. Three products launch simultaneously: OpenAI models including GPT-5.5, Codex on AWS, and Amazon Bedrock Managed Agents powered by OpenAI.
Alibaba's HappyHorse-1.0 Tops Artificial Analysis Video Arena at Elo 1389 T2V, 1411 I2V
HappyHorse-1.0 leads both text-to-video and image-to-video on Artificial Analysis, beating SeedDance 2.0 by 115 Elo points. The 15B-parameter model from Alibaba's Taotian Future Life Lab launched on fal API on April 27.
IBM Bob Launches Globally: Enterprise AI Coding Agent Claims 45% Productivity Gains Across 80,000 Employees
IBM's full-SDLC coding agent Bob goes generally available after a 10-month internal rollout. Multi-model routing across Claude, Mistral, and IBM Granite covers planning through deployment. 30-day trial at bob.ibm.com.
Microsoft VibeVoice Hits 40,000 GitHub Stars as ASR-7B Joins HuggingFace Transformers
Microsoft's open-source voice AI family cleared 40,000 GitHub stars and landed ASR-7B in HuggingFace Transformers in March 2026. The 7B speech recognition model handles 60-minute audio in one pass with speaker attribution and timestamp alignment across 50+ languages.
GPT-5.5-High Enters All Six Arena Categories — OpenAI and Anthropic Now Head-to-Head Everywhere
Arena added GPT-5.5-high to Code, Text, Expert, Search, Document, and Vision on April 27, the first time any model has spanned all six categories simultaneously. Claude Opus 4.7 also joined the Search leaderboard the same day, completing a direct matchup across every Arena benchmark.
OpenRouter's Million-Request Study: Opus 4.7 Costs 27% More in the Agentic Sweet Spot
Analysis of over one million real API requests shows Opus 4.7's new tokenizer adds 12–27% to actual costs for prompts above 2K tokens, with the worst hit at the 2K–10K range that dominates agentic coding workflows. Short prompts under 2K are 1.6% cheaper.
Intel DCAI Revenue Hits $5.1B in Q1 2026, Up 22% — Agentic AI Puts CPUs Back in Demand
Intel reported Q1 FY2026 data centre and AI revenue of $5.1 billion, beating Wall Street consensus by over a billion dollars. CEO Lip-Bu Tan attributed the surge to a structural shift: agentic and inference workloads need CPUs alongside GPUs, reversing four years of GPU monoculture.
Anthropic Locks Opus Behind Pay-as-You-Go for Claude Code Pro — $20/Month No Longer Sufficient
Anthropic has updated Claude Code documentation to require Pro plan users to purchase extra usage credits before accessing any Opus model. Max and Team Premium subscribers retain Opus as included. The change affects Claude Code only, not the Claude.ai chatbot.
Sullivan and Cromwell Apologises to Bankruptcy Judge Over 40 AI-Hallucinated Citations
Wall Street firm Sullivan and Cromwell filed a corrected emergency motion in the Prince Global Holdings bankruptcy after 40 incorrect legal citations slipped through its AI review process. Partner Andrew Dietderich apologised directly to the judge and to opposing counsel.
GPT-5.5-High Joins Arena at 1504 ELO — OpenAI Ties Anthropic at the Top for the First Time
OpenAI's GPT-5.5-high entered Chatbot Arena on April 27, hitting 1504 on text and 1558 on code ELO — tying Claude Opus 4.7 for the top text spot and pulling ahead on coding. The model was added across six leaderboards simultaneously.
GitHub Copilot Kills Flat-Rate Pricing June 1 — All Plans Move to Token-Based AI Credits
GitHub announced all Copilot plans transition to usage-based billing on June 1, 2026. Flat subscription fees give way to GitHub AI Credits priced by token consumption. A preview billing dashboard launches in early May.