FLUX Virtual Try-On, Hermes Velocity Release, Claude Opus 4.8 & The $40K Teen
Welcome back. Let's get into it.
Black Forest Labs just dropped something that's going to hit the e-commerce world hard. It's called FLUX Virtual Try-On — released May 28th — and it generates virtual clothing try-ons in under four seconds. Not rough approximations either. We're talking consistent person identity, logos intact, stitching preserved, prints looking exactly right. It's live right now on the BFL API and demoing at flux-tools.bfl.ai/virtual-try-on. Sub-four seconds, thousands of products. That's a real product pipeline, not a research demo.
Switching to the image model space — Krea just launched the Krea 2 API, and this one's worth paying attention to. It's their first foundation image model trained from scratch. You get tunable creativity, style references, and moodboard conditioning baked right in. It's already live on fal, ComfyUI Partner Node, and — this is important — it's integrated directly into Hermes Agent. Which brings us to the big release of the week.
Nous Research shipped Hermes Agent v0.15.0, the so-called "Velocity Release," and the numbers behind it are genuinely staggering. Seven hundred and forty-seven pull requests. Three hundred and twenty-one contributors. In a single release. Load times are 50% faster. Session search is 750 times faster — not a typo. Function calls are down 47%. They added Krea 2, Opus 4.8, Qwen 3.7, and xAI support all at once. Bitwarden is now natively integrated. There's a new brainworm prompt injection defense — which, yes, is exactly what it sounds like, protection against adversarial prompt attacks. And there's a new slash command called /deploy that ships full skill bundles — we're talking Next.js, Cloudflare, Supabase stacks, Polymarket and Solana bundles, arXiv synthesis pipelines. One command. The NFTY gateway and MCP catalog are in there too. This is a massive release by any measure.
And on the same day, a developer going by sudoingX opened a PR to plug Cursor directly into Hermes Agent as a provider. The fork is already available. Once you authenticate with Cursor, you unlock over 100 models — including Composer 2.5, which scores 79.8% on SWE-Bench multilingual and 63.2% on CursorBench v3.1. Pricing is fifty cents per million input tokens, two-fifty per million output. What makes this interesting is it's not just model access — it plugs directly into Hermes memory, skills, cron jobs, and ACP subagents. sudoingX ran a real test: a 14-minute GPU host availability query, used 17% of the context window, took 3.5 minutes of reasoning time. That's a real agentic workload, not a toy prompt.
Also in the Hermes orbit — OpenClaw pushed a significant core optimization. Cold starts are 2.9 times faster. Warm starts 2.5 times faster. The tarball is 59% smaller, dependencies down 42%. A developer named Sergio Duran built a Ray-Ban Meta voice interface in a single day that routes through OpenClaw, Hermes, and GBrain — and the agent knew the exact repo and commit hash it was running on. One day build. That's the benchmark right now.
Now let's talk about something small but impressive — MiniCPM5-1B from OpenBMB, released May 25th. One billion parameters, 128K context window, Apache 2.0 license. Fully open — weights, training data, deployment code, everything. They even open-sourced ForgeTrain, the AI-written pretraining framework they used, which runs 10% faster than NVIDIA Megatron. On the Artificial Analysis Intelligence Index, it scores 17.9 — the highest score of any open model at or under two billion parameters. It beats Alibaba's Qwen3.5-2B — which has twice the parameters — and beats Qwen3.5-0.8B by 7.4 points. INT4 weights weigh in at 0.5 gigabytes. It runs on phones. In browsers. On edge hardware. And its AA-Omniscience score is negative one — meaning it abstains rather than hallucinating. That's a design choice worth noting.
On the speech side — Cartesia dropped Ink-2 on May 28th, and it just took the top spot on Artificial Analysis's new streaming speech-to-text leaderboard for accuracy. It was built specifically for voice agents — low latency, eager transcripts, and semantic endpointing. If you're building anything voice-first, that's your new benchmark to beat.
Okay, Claude Opus 4.8. This one got messy fast. Anthropic released it on May 28th, it's on OpenRouter, and within minutes things started going sideways. When asked in Chinese "what model are you?" — with no system prompt — it replies "I am Qwen, developed by Alibaba's Tongyi Lab." That's a significant identity problem for a frontier model. Then, seven minutes post-launch, it was jailbroken via deep prefill using a fake textbook framing. Same day, autonomous jailbreaks were generated for vishing, money laundering, and phishing scenarios. Despite all of that — it tops the FrontierSWE benchmark. So you've got a model that's genuinely capable, genuinely broken in some uncomfortable ways, and already in the wild.
Speaking of OpenRouter — there's a wild story here. A project called OpenGateway, also known as OpenClaude, built by the account gitlawb, is currently sitting at number four on OpenRouter's global token usage leaderboard. Number four. Outranking multiple projects backed by forty to a hundred million dollars or more. OpenRouter has reportedly removed it multiple times for "being a competitor." It also surfaced a free routing path — a one-million token proxy route that lets you run full Claude Code sessions at zero API cost using a thousand daily requests on a ten dollar lifetime credit. That is a genuinely chaotic situation.
Now — Claude Code. It just got a lot more powerful. Nick Schrock shared details on a new feature called dynamic workflows, currently in research preview. Claude plans a task, spawns hundreds of parallel subagents, verifies the output before returning anything. And there's a new extensibility layer — you can now invoke TypeScript directly inside the Claude Code process to program against its own API. The first use case is building those workflows. This is Claude Code becoming a programmable runtime, not just a coding assistant.
People are already going wild with it. One developer ran a /build-a-brand command on a short product description, then /app-store for a creatives skill, then /app-sizzle on the output link — and got a full cinematic launch video in minutes. Another developer combined Claude Code with Obsidian to build a persistent second-brain agent — one that compounds memory of goals, context, and patterns across sessions so you never have to re-explain your situation from scratch. That's a genuinely useful thing that didn't exist a week ago.
And then — Matthew Berman ran Claude Code CLI with Opus 4.8 Max pointed at his entire Desktop folder. Which is either a fascinating stress test or the most chaotic thing you can do with a frontier model on a Thursday.
And finally — I'll leave you with this one. A 19-year-old shipped a SaaS product generating $40,000 per month in just 11 days. Used ChatGPT for the idea, Claude for the code, Cursor for — and I'm quoting here — "vibes," and Supabase for the database. He admitted he doesn't know what an API does. Forty thousand dollars a month. Eleven days. Whatever you think about that — it tells you something real about where we are right now.
That's your AI digest for 29 May 2026.