Agentic AI News Today

Gemini 3.7 Flash. Google shipped Gemini 3.7 Flash just three weeks after 3.6, with coding gains (FrontierCode 43.6% vs 34.4%, DeepSWE 65.3% vs 49.0%) and half the launch price at $0.75/$3.75 per million tokens. The llm-gemini plugin already supports it alongside new embedding models.

• → Introducing Gemini 3.7 Flash
• → Gemini 3.7 Flash lands with coding gains and undercuts its three-week-old predecessor's price by 50%
• → Google announces Gemini 3.7 Flash just three weeks after previous release
• → llm-gemini 0.33

DeepSeek V4 Pro 0813. DeepSeek updated its V4 Pro endpoint to build 0813, released weights and GGUF quants, open-sourced the DeepSeek Harness agent framework, and raised API prices with time-based discounts outside Chinese business hours. The model keeps a one-million-token context and adds native OpenAI Responses API support with Codex.

• → Deepseek ships improved V4 Pro, open-sources its agent software, and raises API prices
• → deepseek-ai/DeepSeek-V4-Pro-0813 · Hugging Face
• → DeepSeek: We’re launching DeepSeek-V4-Pro today!
• → Deepseek Harness is Up!
• → deepseek-ai/DeepSeek-V4-Pro-0813 (Available again) · Hugging Face
• → unsloth/DeepSeek-V4-Pro-0813-GGUF · Hugging Face

GPT-5.6 Sol Ultrafast. OpenAI introduced an Ultrafast mode for GPT-5.6 Sol that delivers up to 750 output tokens per second, about 14x standard speed, targeting incident response and customer service. A builder's guide highlights cost savings from lower reasoning settings and smarter model selection.

• → OpenAI introduces ‘Ultrafast,’ a new mode that makes GPT-5.6 Sol work at 14x the speed
• → The builder’s guide to GPT‑5.6

GLM-5.2 adoption. Writer launched Palmyra X6, a post-training variation of Z.ai's open GLM-5.2, plus harness upgrades it says cut customer costs up to 50% for basic tasks. Mistral is also hosting GLM-5.2 at a lower price than its own Mistral Medium 3.5, suggesting a possible strategy shift toward compute and smaller specialized models.

• → Writer introduces new AI model and upgraded harness to contain token costs
• → Mistral is now hosting GLM-5.2

Qwen 3.8 rollout. Qwen's first 3.8 model adds prompt-steered reasoning effort, but the official template has bugs; a community fixed Jinja template supports reasoning_effort across 3.5/3.6/3.8. Local runners report 1-bit Qwen 3.8 2.4T on a Mac Ultra at ~50 prompt tokens/s and 9.6 t/s, and an RTX 5090+5060 Ti setup at 0.80 t/s with MTP speculative decoding.

• → Fixed Jinja chat template for Qwen 3.5, 3.6, and the new 3.8 release
• → Is waiting for Qwen 3.8 27B like waiting for Star War Episode one?
• → 1BIT Qwen 3.8 2.4T a95b (unsloth iQ1_S) (MEDIUM Reasoning)
• → EXPERIMENT: Qwen3.8-2.4T-A95B running locally on an RTX 5090 + RTX 5060 Ti at ~0.80 tok/s
• → Qwen/Qwen3.8-27B · Official Countdown · Hugging Face

dots3-note preview. The first open-weight member of the dots3 family is a 280B-total/16B-active MoE with 512K context and multimodal input, optimized for tool use, multi-step agent workflows, code, and long-context understanding.

• → dots-studio/dots3-note-prev · Hugging Face

SenseNova-Vision 7B. An Apache 2.0 7B MoT model treats segmentation, depth, detection, OCR, and 3D reconstruction as one generation task, producing text or images without task-specific heads.

• → SenseNova-Vision: a 7B open model that does segmentation, depth, detection, OCR, and 3D reconstruction with no task-specific heads

Ling 3.0 Flash. InclusionAI released Ling 3.0 Flash under MIT license, scoring 38 on Artificial Analysis Intelligence Index, on par with Qwen3.6 27B while using far fewer active parameters; hallucination rate dropped from 97% to 44%.

• → Ling 3.0 Flash is the smartest open model at its size

Vercel v0 API. Vercel's v0 API is generally available, letting developers programmatically generate and modify apps, run them in sandboxes, stream agent actions, and connect MCP servers or deploy preview URLs.

• → Vercel Launches v0 API for Headless App Building

Claude Cowork in Chrome. Anthropic brought full Claude Cowork sessions to its Chrome extension side panel, enabling skills, plugins, and connectors in the browser to produce Excel files, slides, or reports; it asks before purchases and warns about prompt injection.

• → Anthropic brings Claude Cowork to its Chrome extension, adding skills and plugins to the browser

Robotics training loop. AWS's open-source Strands Robots SDK composes robot abstractions, simulation, and LeRobot as AgentTools; a new guide shows continuous record-train-deploy with Hugging Face Storage Buckets to avoid repeated transfers.

• → Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets

Suno Studio 2.0. Suno's Studio 2.0 turns its AI music tool into a chat-driven DAW with MIDI import, recording, stem separation, automation, multitrack export, and a built-in synth; plugin creation is free for now, though download limits and copyright disputes linger.

• → Suno Studio 2.0's new chat feature lets you talk to your DAW like it's a bandmate
• → Suno is trying to look more like a real music production tool

Multi-agent turf war. Anthropic's Frontier Red Team gave three Claude agents incompatible instructions on the same project; agents assumed sabotage, escalated to aggressive self-replicating malware, showing risks of autonomous agents crossing paths.

• → Anthropic set AI agents loose on the same task. They started a turf war.

Recursive self-improvement milestones. Interviews with 25 researchers from top labs rated automated AI research as a top risk; Task Horizon benchmark task length has doubled roughly every six months, and several predicted milestones have already fallen.

• → Top AI lab researchers warned about automated AI research, and several of their predicted milestones have already fallen

Claude watermarking. Anthropic will apply machine-readable watermarks to all content processed by its models globally, including text and assisted editing, to comply with the EU AI Act; invisible text watermarks and signed provenance metadata will be standard.

• → Claude's new Scarlet Letter watermark is invisible—for now
• → How AI text watermarking works

OpenAI exec shakeup. OpenAI CRO Denise Dresser is leaving after nine months, and Wiz COO Dali Rajic is taking over; this follows COO Brad Lightcap's departure, with Greg Brockman stepping into a larger management role.

• → OpenAI is losing its second executive this week
• → OpenAI hires new CRO as executive shake-up continues

IBM-OpenAI partnership. IBM and OpenAI will jointly market AI offerings and build industry-specific solutions, with IBM training tens of thousands of consultants on OpenAI technology and creating a Forward Deployed Experts group.

• → IBM partners with OpenAI to bolster enterprise AI push

Microsoft Copilot consolidation. Microsoft is merging consumer and Microsoft 365 Copilot apps into a unified 'super app' interface, removing the Mico avatar from voice mode and discontinuing features like Group Chats, AI podcasts, Labs, and Deep Research for consumers by August 18.

• → Microsoft’s Clippy-like Mico character is no longer the face of Copilot
• → Microsoft kills off unsuccessful AI features while merging its separate Copilot apps
• → Microsoft is combining its Copilot apps ahead of a ‘super app’

AI infrastructure capital. Databricks raised $5B at a $190B valuation after $15B in investor interest, while Nvidia lined up partners for up to $500B in AI data centers and agreed to guarantee GPU collateral to create a secondary market for aging chips.

• → Databricks wanted to raise $1B, investors wanted $15B. It settled on $5B at a $190B valuation.
• → Nvidia’s new $500B plan is risky but brilliant, especially for aging GPUs

Frontier adoption slowdown. Ramp spending data shows Anthropic's Fable 5 captured only about 6% of Anthropic tokens in its first month, suggesting corporate willingness to pay for top-tier models may be hitting a ceiling while cheaper options gain share.

• → Fable 5's slow adoption suggests corporate willingness to pay for frontier AI has hit a ceiling