# AI News Flash — Daily Brief **Date:** Friday, May 29, 2026 **Type:** daily **Source:** https://ainewsflash.co/brief/17 **Editors:** Justin Bunnell (https://www.linkedin.com/in/justinbunnell/), Laz Manrique (https://www.linkedin.com/in/laz-m-5a218b81/) **Publisher:** AI News Flash (https://ainewsflash.co) **License:** Republish with attribution. --- ## Key takeaways - Anthropic ships Claude Opus 4.8 with honesty and agentic gains - Mistral rebrands Le Chat to Vibe, ships Leanstral math-proving model - Qwen3.7-Max Claims 35-Hour Autonomous Agent Sessions With 1,158 Tool Calls - GPT-5.5 Tops Terminal-Bench 2.0 at 82.7%, Setting New CLI Agent Ceiling - Zyphra's ZAYA1-8B MoE trained end-to-end on AMD Instinct hardware - LLaVA-OneVision-2 processes video as native codec bitstream, not frames --- ## Platforms ### Anthropic ships Claude Opus 4.8 with honesty and agentic gains Anthropic released Claude Opus 4.8 on May 28, just 41 days after Opus 4.7—the shortest gap between consecutive Opus releases. Agentic coding scores jump from 64.3% to 69.2% on SWE-Bench Pro, and knowledge-work scores rise from 1753 to 1890 on GDPval; the company frames the headline improvement as calibration: Opus 4.8 is described as four times less likely to let flaws in its own code pass unremarked. Fast mode is 2.5x quicker and 3x cheaper than before, and the release bundles a research preview of Dynamic workflows, which lets Claude orchestrate hundreds of parallel subagents inside a single session. - Source: https://www.anthropic.com/news/claude-opus-4-8 ### Mistral rebrands Le Chat to Vibe, ships Leanstral math-proving model Mistral renamed its Le Chat product to Vibe and restructured it as a unified work-and-code agent with two modes: Work Mode for long-running multi-stage tasks and Code Mode for remote coding and pull-request workflows, plus a new VS Code extension. Alongside the rebrand, Mistral released Leanstral, an Apache 2.0 math-proving model that beats Claude Sonnet on formal verification benchmarks at pass@2 while costing roughly 15x less per run. - Source: https://releasebot.io/updates/mistral ## Capabilities ### Qwen3.7-Max Claims 35-Hour Autonomous Agent Sessions With 1,158 Tool Calls Alibaba's Qwen3.7-Max, launched May 20 at the Alibaba Cloud Summit, is built specifically for long-horizon agent execution rather than chat quality—Alibaba claims it ran autonomously for 35 hours, fired 1,158 tool calls without human intervention, and delivered a 10x speedup on a GPU kernel it had never seen during training. On third-party benchmarks, it scores 56.6 on the Artificial Analysis Intelligence Index v4.0 (a 4.8-point gain over its predecessor, ranked 5th globally and highest among Chinese models), 69.7 on Terminal-Bench 2.0, and 60.6 on SWE-Bench Pro. The 35-hour run figure comes from Alibaba's own internal benchmark with no independent reproduction yet, so treat it as directional rather than guaranteed. - Source: https://qwen.ai/blog ### GPT-5.5 Tops Terminal-Bench 2.0 at 82.7%, Setting New CLI Agent Ceiling Terminal-Bench 2.0, which simulates a real software engineer working in a sandboxed terminal with a 5-hour timeout across 89 tasks (compile, train, configure, debug), now shows GPT-5.5 at 82.7% as the May 2026 leader—ahead of Claude Opus 4.7 at 69.4% and Qwen3.7-Max at 69.7%. The benchmark is notable because it resists the reward-hacking vulnerabilities that have inflated scores on SWE-Bench Verified, with tasks requiring multi-step tool use under real time constraints rather than patch generation alone. - Source: https://www.digitalapplied.com/blog/ai-model-releases-may-2026-complete-tracker ## Technology & Research ### Zyphra's ZAYA1-8B MoE trained end-to-end on AMD Instinct hardware Zyphra released ZAYA1-8B, an Apache 2.0 MoE model with 8B total parameters but only ~760M active per token, trained entirely on AMD Instinct GPUs—proving a viable frontier-quality training path outside the NVIDIA stack. On reasoning benchmarks it matches models activating 37–40B parameters. The model is available on Hugging Face and via a free Zyphra Cloud endpoint, making it the clearest demonstration yet that AMD can anchor a full training-to-deployment pipeline. - Source: https://huggingface.co/Zyphra/ZAYA1-8B ### LLaVA-OneVision-2 processes video as native codec bitstream, not frames Researchers at Glint Lab, AIM for Health Lab, and MVP Lab built LLaVA-OneVision-2, which ingests video as a continuous bit-cost stream drawn directly from codec structures rather than sampling discrete frames. The shift improves temporal grounding and fine-grained event localization, beating frame-centric baselines by 4.3 points on video understanding tasks and reaching 74.9 mAP on a new benchmark. Treating video as a compression signal rather than an image sequence could reduce the token cost of long-video inference substantially. - Source: https://arxiv.org/abs/2505.18111 ## Regulation & Policy ### CNN sues Perplexity for copyright theft in first TV network AI case CNN filed suit against Perplexity AI in the Southern District of New York on May 28, alleging the company unlawfully copied and distributed its journalism without authorization after licensing talks broke down. The case is the first AI copyright action by any television network and joins parallel suits by The New York Times and the Chicago Tribune against Perplexity, adding a new plaintiff class—broadcast news—to the growing litigation wave over AI scraping. - Source: https://www.cnn.com/2026/05/28/media/cnn-sues-perplexity-ai-copyright ### UK ICO closes recruitment AI consultation, pushes mandatory human oversight The UK Information Commissioner's Office concluded its public consultation today (May 29) on automated decision-making in recruitment, having published findings that employers must maintain more meaningful human involvement in AI-assisted hiring processes. The move is part of a broader shift in the UK's approach: the country's four main digital regulators—ICO, CMA, FCA, and Ofcom—jointly published a paper this month on agentic AI governance, signaling that existing transparency, fairness, and competition obligations apply fully to AI agents even absent new primary legislation. - Source: https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/05/ico-report-automated-decision-making-recruitment/ ## AI Stocks ### (CRM) Salesforce hits $1B Agentforce ARR in record Q1 FY27 beat Salesforce reported Q1 FY27 revenue of $11.13 billion, up 13% year-over-year, with GAAP EPS of $2.42 rising 52% and non-GAAP operating margin hitting a record 34.8%. Agentforce ARR crossed $1 billion — growing over 200% year-over-year — and total AI and data ARR reached $3.4 billion, with 3.8 billion Agentic Work Units delivered in the quarter. Despite the beat, full-year revenue guidance came in slightly below Street estimates and the stock was little changed in extended trading, as shares remain down 33% year-to-date on concern that AI disruption will pressure software growth. - Source: https://www.sec.gov/Archives/edgar/data/0001108524/000110852426000125/crm-q1fy27xexhibit991.htm ### (ARM) Morgan Stanley downgrades Arm, sees agentic AI revenue delayed Morgan Stanley downgraded Arm Holdings, citing that agentic AI revenue is likely to take longer to materialize than the market expects, even as the firm raised its price target to $150. The move sent ARM shares down 7%, a notable single-session decline for one of the AI infrastructure trade's key royalty plays. The call adds to debate over how quickly Arm's licensing model will monetize the wave of AI-edge and data-center chip designs built on its architecture. - Source: https://www.tikr.com/blog/broadcom-nasdaq-avgo-stock-climbs-6-after-securing-google-ai-chip-supply-deal --- Archive: https://ainewsflash.co/archive RSS: https://ainewsflash.co/rss.xml LLM index: https://ainewsflash.co/llms.txt