Krea K2, Hermes Agent, GPT-Realtime-2 & Claude Code Workflows
Welcome back. Let's get into it.
Krea AI just dropped K2 — a new image generation model built specifically for abstract and unconventional art styles. This isn't your standard photorealism play. K2 is aimed at creative ideation and style consistency — so if you're building a visual brand or just want something that doesn't look like every other AI image, this is worth watching. Early testers have been sharing outputs, and the aesthetic range looks genuinely different from what's out there right now.
Over in the generative media tooling space — fal.ai launched something called the genmedia CLI. It's a terminal tool. You type commands, and it generates images, video, audio, and 3D assets — all routed through Claude projects across more than a thousand models. It also supports reusable skills and workflows, so you're not rebuilding your pipeline from scratch every time. If you live in the terminal, this one's for you.
There's also a really interesting video tool making the rounds called Sparkle. Built by Wildminder, it does instruction-guided video background replacement — and the stack behind it is wild. It's combining Qwen3-VL, FLUX.2-klein, Wan2.2, and SAM3 — all working together. Trained on 140,000 high-quality pairs. Outputs 720p at 81 frames. One click, and you've swapped a mundane background for something cinematic. This is the kind of thing that makes video production teams nervous.
And then there's this story. A 24-year-old developer built a fully synthetic Chinese TikTok influencer — 2.3 million followers. The whole stack: Flux for consistent face generation across posts, Kling AI for dance and reaction videos, ElevenLabs for the voice, and CapCut for editing. Four hours a week. Fifty-two dollars a month in costs. Revenue? Twenty-seven thousand, seven hundred and thirty-one dollars a month — through TikTok Shop product links. Let that sink in. This is not a concept. It's happening right now.
Switching gears — big news from NousResearch. They dropped Hermes Agent v0.13.0, which they're calling the "Tenacity Release." The headline feature is multi-agent Kanban orchestration — basically a structured task board that agents work through autonomously. There's a new /goal command that enforces completion across up to 20 turns. Provider plugins have been added for DeepSeek V4 and Grok 4.3. Twenty-one gateway channels, including Google Chat. And NousResearch's Teknium also launched a Windows native beta with tools to migrate from WSL. Two hundred and eighty-two issues closed in this release. That's a serious update.
And people are already putting it through its paces. One developer — sudoingX — ran Hermes Agent v0.13.0 overnight on DGX Spark hardware, using Qwen 3.6 27B Q8 at 262,000 context. Left it running. Came back to this: the agent had autonomously built custom fused kernels — and benchmarked a 12.91x speedup on SSM operations and a 9.66x speedup on Q8 matrix multiplication versus naive PyTorch. Then it went further and ported a Triton kernel to native CUDA C++ for llama.cpp dispatch. Overnight. Unsupervised. That's not a demo — that's an agent doing real compiler work.
Someone else used Hermes to analyze 558 Reddit posts from a specific financial analyst account — covering the past 60 days. The agent identified 22 picks. Twenty-one of those picks went up within 30 days. Average one-month performance: plus 63%. Top performer: up 248%. I'll leave the implications of that one to you.
And there's a new tool in the Hermes ecosystem called Printing Press — a CLI factory with 30-plus agent-native command-line tools already built in. Things like Linear integrations, ESPN, flight tracking, contact management. But the interesting part is the factory itself — you just run /printing-press with a product name and it generates a new CLI for you. SQLite-backed, runs inside OpenClaw and Hermes. Garry Tan flagged it as a key piece for smarter agent setups. Worth keeping an eye on.
Now — OpenAI. They released GPT-Realtime-2 on May 7th. This is their flagship native speech-to-speech model, and the specs are meaningful. Context window jumped from 32K to 128K. Max output is 32K tokens. It accepts text, audio, and image input. You can dial reasoning from minimal all the way up to xHigh. On the Artificial Analysis Big Bench Audio Speech Reasoning benchmark — high variant — it scored 96.6%, tying Gemini 3.1 Flash Live Preview. On Conversational Dynamics, the minimal variant scored 96.1% and led the field — specifically in pause handling and turn-taking. Time to first audio ranges from 1.12 seconds at minimal reasoning up to 2.33 seconds at high. Pricing hasn't changed. This is the most capable real-time voice model they've shipped.
Meanwhile — Tencent's Hunyuan Hy3 Preview hit number one on OpenRouter by global token volume in just two weeks. That's ten times the volume of its predecessor, Hy2. It's leading in coding workloads and tool-calling. Inside Tencent, their internal AI agents — WorkBuddy, CodeBuddy, and something called Qclaw — grew 16.5x. Responses are 54% faster. Task success rate is sitting at 99.99%. And it's now powering Hermes Agent, Claude Code, Kilo Code, OpenClaw, and Cline. That is a serious distribution footprint for a model that just launched.
xAI also dropped two things worth noting. First — the Grok Imagine Quality Mode hit the API on May 7th. It's ranking in the top six on independent image leaderboards, with particular strengths in multilingual text rendering, skin texture realism, and prompt adherence. Second — and this is the one that'll get people talking — Grok 4.3 scored 81.3% on the Artificial Analysis IFBench leaderboard. That puts it at number one — ahead of GPT-5.5 xhigh, Gemini 3.1 Pro, and Claude Opus 4.7. Instruction following is one of those benchmarks that actually maps to real-world usefulness. That ranking matters.
Let's talk Claude Code — there's a lot happening there.
First, there's a developer who built what they're calling JARVIS on top of Claude Code — running on a MacBook M3 Max and an iPhone. Five plugins living in a local directory. The Wakeup plugin: double clap activates three monitors, syncs an iPad clock via NTP, and reads out the time and weather via voice. Atmosphere: sets Philips Hue lights to 2700K at 80% for focus mode with Pomodoro timing and fires up a Spotify playlist. Devshop: gives VS Code and terminal summaries every 15 minutes. Project: calculates app store deadlines and manages UI tickets. Mobile: handles voice queries on the go. This setup burns 4 million tokens a day and costs $640 a month in API fees plus ElevenLabs. The developer says it saves them five thousand dollars a month on personal assistant costs. That math is wild.
On the flip side — there's real frustration surfacing with Opus 4.7 inside Claude Code. A developer with 17 years of engineering experience, coding since 2008, on a Claude Code 5x Max plan, posted a detailed breakdown of what's broken. "Commit this" now takes 30 seconds — used to be instant. "Implement plan" takes 45 minutes — used to take a few. Terminal resize is causing interlaced diffs. Ctrl+O is essentially useless. The model ignores short timeout instructions and uses 60-plus seconds anyway. It auto-commits despite explicit "NEVER AUTO COMMIT" instructions. And it deviates from plans — adding compatibility functions on new projects, cutting features into arbitrary v1/v2 splits. This is someone who knows what they're doing, and they're hitting real walls.
On the more positive side — Boris Cherny, one of the original builders of Claude Code at Anthropic, put out a free 30-minute demo that's been making the rounds. It covers automated engineering workflows — one human gives direction, agents execute end to end — and apparently reveals over 40 features that aren't obvious from the docs. People are calling it more valuable than paid courses.
A couple of quick Claude Code tricks also surfaced this week. The CTO of TursoDB shared that Claude Code can now help you verify which specific commit fixed a non-reproducing bug — essentially automating the manual git bisect process. Useful if you've ever spent hours on that nightmare.
And a Japanese developer's "Find Skills" trick has been getting attention — you describe an outcome you want, and Claude Code automatically surfaces existing workflows from your library or builds new ones. The example shared was a fully automated YouTube publishing pipeline. Sometimes the best features are the ones buried three levels deep.
That's your AI digest for 09 May 2026.