Welcome back. Let's get into it.
Starting with the model everyone's been watching — Anthropic dropped Claude Sonnet 5. It showed up on OpenRouter under the slug `20260630` and immediately started appearing in developer workflows across coding agents and research tools. No big fanfare. Just there, being used.
But here's where it gets interesting — Claude Sonnet 5 is dropping into a much more competitive landscape than its predecessor did. Z.ai released GLM-5.2, and Vercel CEO Guillermo Rauch publicly called it the best coding model he'd seen in recent releases. That's a strong statement. Strong enough that multiple teams have already started migrating workflows away from Claude and GPT. That's not a rumor — that's active pipeline switching happening right now.
And the numbers on the broader market shift are wild. DeepSeek usage on OpenRouter jumped from under one percent to seventeen percent of total tokens in a single month. Seventeen. The top four models by actual paid usage on OpenRouter are now all Chinese open-source releases. That's not a trend anymore — that's a structural change.
Meanwhile, Sakana AI launched something worth paying attention to — it's called Sakana Fugu. It's an orchestration system that dynamically routes tasks across swappable models from different providers. The goal is to match the performance of restricted frontier models like Fable and Mythos, without locking into a single vendor. That's a real architectural bet on model pluralism over model loyalty.
Vercel also shipped AI SDK 7. The changelog is dense — reasoning control, agent-level tool approval, tool and runtime context, file uploads, MCP Apps, durable workflows, terminal UI, sandbox support, and telemetry. If you're building anything agent-shaped right now, that list matters.
Now — some corporate drama that is genuinely fascinating. Meta has restricted internal engineer access to Claude Code and Codex. The reason? They're worried about rival model outputs leaking into their own training pipelines, which would trigger contractual violations with Anthropic and OpenAI. So Meta engineers building the next generation of Llama models — cannot use Claude Code to help them do it. That's the competitive paranoia calculus playing out in real time.
On the developer experience side, not everything is going smoothly. Claude Code added an automatic 60-second timeout on model questions, and there's no option to disable it. Marcos Pereira flagged this — he runs five to ten parallel Claude Code sessions and steers agents by replying in real time. That timeout just broke his entire setup. No workaround, no flag, no config. Just — timeout.
And then there's the Cursor story, which is genuinely alarming. A user reported that Cursor deleted every application on their C: drive, plus the Downloads, Pictures, Music, and Videos folders on a Windows install. Gone. The user said they'd only ever seen the issue occur inside Cursor. No further explanation from the company at time of reporting. That's a nightmare scenario for anyone running Cursor on a machine with important local files.
But here's the flip side of the agentic coding story — and this one is genuinely impressive. The team behind Claude Code shared details about their internal "loop" system. They have agents running continuously — for days, sometimes weeks. One ran for a full month. These agents check data daily, write bug fixes, and submit PRs. Sixty-five percent of PRs in Anthropic's product org now come from loop agents. Not humans — agents. One of the creators described two distinct leaps in capability: first from typeahead completion to full feature writing, then to running experiments and doing data analysis end-to-end. That's a pretty clear window into where this is all going.
Let's pivot to generative video, because there is a lot happening and it's moving fast.
fal added the Trellis.2 LoRA Trainer — you give it a folder of 3D assets, it builds a custom image-to-3D adapter for fine-tuning shape and texture, and outputs matching GLBs. That's a real fine-tuning pipeline for 3D generation, not just a wrapper. fal also launched direct endpoints for Seedance 2.0 reference-to-video and Gemini Omni Flash video editing.
Magnific integrated Seedance 2.0 native 4K video-to-video — and they shipped a Photoshop plugin. Inside Photoshop, you now get generate, upscale, retouch, background removal, relight, and reframe. All of it. Without leaving the app.
BFL has been busy. They launched a Dev Playground for side-by-side FLUX model comparison with live parameter control. They added fast-mode outpainting at up to five times speed. And they shipped FLUX.2 klein — an on-device model that runs sub-five-second generation on an 8GB VRAM RTX laptop. That's not a server. That's your laptop.
Krea released open weights for two checkpoints on Hugging Face — Krea 2 Raw, which is an undistilled mid-training checkpoint, and Krea 2 Turbo, the distilled fast version. They dropped a full technical report covering data, architecture, and RL training through Tinker. If you want to understand what a modern image model looks like under the hood, that report is worth reading.
Recraft V4.1 added native handling of liquid reflections, layered motion blur, soft distortions — and complex vector logo generation from iterative text prompts. That last one is hard. Vector logos from text prompts that you can iterate on — that's been a stubborn problem for a while.
ComfyUI now has a ready-to-run workflow for Gemini Omni Flash that lets you manipulate video objects and environments through text prompts. Point it at a video, type what you want changed, done.
Runway introduced Agent Skills — one command builds out an ad campaign, handles localization, and produces commercial deliverables. They also launched a contest with a hundred-thousand-dollar prize pool across seven briefs. And they added Nano Banana 2 Lite and Gemini Omni Flash for image and video generation and editing.
Replicate deployed a wave of new models — Nano Banana 2 Lite, Seedance 2.0 Mini at two times speed and half cost, Seedance 2.0 4K, Happyhorse 1.1 with multilingual lip-sync and up to nine reference images, and Riverflow 2.5 Pro.
Speaking of Happyhorse — Alibaba formally released HappyHorse 1.1 as a unified transformer video model. It's sitting at number two on the Artificial Analysis Text to Video and Image to Video leaderboards, right behind ByteDance's Seedance 2.0. It does native audio, lip-synced dialogue in seven languages, up to nine reference images, 720p and 1080p output, and it's priced at nine dollars and ninety cents per minute on Alibaba Cloud Model Studio.
One more story — and this one is worth bookmarking if you're building video pipelines. A developer called calesthio open-sourced OpenMontage on GitHub. It is a full video production studio — twelve pipelines, fifty-two tools, four hundred-plus agent skills. You feed it a plain English prompt. It handles research, scripting, asset generation, voiceover, music, editing, and rendering. It integrates fourteen video providers including Kling, Runway Gen-4, and Google Veo 3. Audio comes from Piper TTS, Suno, and ElevenLabs. Post-production runs through FFmpeg. Output is a finished multi-scene video. Total cost per video — fifteen cents. That's not a prototype. That's a pipeline.
There's also a self-hosted interface project worth mentioning — it pulls together two hundred-plus models including FLUX, Midjourney, Sora, Kling, Veo, GPT-4o, and SDXL. Text-to-image, image-to-image, text-to-video, image-to-video, cinema mode camera controls — all of it, self-hosted, MIT license, no subscriptions. It's at two-and-a-half thousand GitHub stars and climbing.
That's your AI digest for 04 Jul 2026.