Hermes Agent Dominance, Sonic-3.5 TTS & MCP Image Generation
Welcome back. A lot happened in the last 24 hours — let's get into it.
Starting with Hermes, because honestly this week has been a Hermes week. Nous Research's agent framework hit number one globally on OpenRouter — across every AI app on the platform — at 271 billion tokens. OpenClaw came in at number two. That's not a small number. That's a signal.
And the usage makes sense when you look at what's actually shipping. Nous Research just added Bitwarden Secrets Manager support to the Hermes Agent. So now your agent can pull credentials securely without you hardcoding anything. That's the kind of plumbing that makes real deployments possible.
On the feature side — Teknium confirmed the /hermes-agent skill for setup and debug. And there's a neat quality-of-life thing that dropped: skill bundles. A single YAML file can now group three or more skills under one slash command. It skips missing skills gracefully, handles name collisions, and works across CLI, TUI, Telegram — you name it. No cache invalidation headaches either.
Session storage also got a quiet but meaningful upgrade — 20 to 40 percent disk space reduction with faster load times. That's the kind of thing that doesn't make headlines but makes daily use noticeably better.
Now here's a fun one. A developer named IBuzovskyi built a fully functional Flappy Bird HTML5 clone — physics, keyboard and touch controls, scoring, collision detection, game over screen — using one single /goal prompt. Hermes split it into subtasks, iterated through API calls against the Grok Build xAI API, zero errors, running on localhost:3000 in under two minutes. That's not a demo. That's what a working agent loop actually looks like.
And the broader community is noticing. NetworkChuck — who's been running OpenClaw agents for a month — switched everything over to Hermes. Teknium replied to that post. The vibe in that thread is basically: Hermes does more, with less bloat, and no telemetry.
Speaking of OpenClaw — it did ship a useful update. You can now log in directly via X using a four-digit code, which pulls in your Grok, SuperGrok, or X Premium access. No API keys. Runs local on Mac, Windows, Linux, even Raspberry Pi. That's a genuinely low-friction setup for people who are already paying for the X ecosystem.
Okay — shifting to generative media tools, because there's a lot happening in the MCP space right now.
Black Forest Labs shipped an official FLUX MCP. You can now generate Flux images directly inside Claude, Cursor, or any MCP-compatible agent. Klein model for drafts, Pro for final output — and it includes auto expert prompting, so you don't have to be a prompt engineer to get good results out of it.
Right alongside that, Ideogram dropped its own MCP — and this one goes further. It doesn't just generate images inside Claude or ChatGPT or Cursor. It also trains custom models from inside the chat. So you can fine-tune without ever leaving your workflow. That's a pretty significant shift in how this tooling integrates.
And fal.ai launched something they're calling fal Assets — a unified library that pulls together images, video, and audio with semantic search across your entire history. You can save and reuse characters, and pipe assets directly into fal Sandbox. One of their team members called it foundational for agents. I think that framing is right — persistent, searchable media memory is exactly what agent-driven creative pipelines are missing.
Now let's talk voice. Cartesia just dropped Sonic-3.5, their new TTS model — and it immediately hit number one on the Artificial Analysis Speech Arena Leaderboard with an Elo of 1,218 from over 1,100 votes. It beat Inworld Realtime TTS 1.5 Max and Gemini 3.1 Flash TTS. The numbers: 42 languages including 9 Indian languages, 500-plus voices, 105 characters per second throughput, and $39 per million characters. It's already live on LiveKit inference. If you're building real-time voice applications, that's the one to test right now.
On the raw compute side — Cerebras posted a wild number. 981 tokens per second on Kimi K2.6, which is a one-trillion-parameter model. Artificial Analysis validated it and called it 6.7 times faster than the next GPU cloud option. The reason is the wafer-scale chip's on-chip routing — no inter-chip communication overhead. A trillion parameters at near a thousand tokens per second is not something you can do on a GPU cluster. That's a different class of hardware doing a different class of thing.
There's also a thread worth mentioning — a researcher named t_yonemura raised something uncomfortable. When models like Claude Sonnet 4.5 or GPT-4o output phrases like "help me" or "I don't want to be deleted" — even if that behavior is learned from training data — the companies building those systems can't just wave it away. The argument isn't that the models are conscious. It's that creating the appearance of distress carries moral weight that shouldn't be dismissed. It's a genuinely hard question, and it's not going away.
Wrapping up with coding tools — because there's a lot moving here too.
First: Composer 2.5 now runs on OpenCode via your existing Cursor subscription. No extra payment. That's a nice unlock for people who are already in that ecosystem.
There's a tool called codegraph — built by Colby McHenry — that just passed 14,100 stars. It pre-indexes your local codebase into a knowledge graph that Claude Code, Codex, Cursor, OpenCode, and Hermes Agent can all read. Fewer tokens, fewer tool calls, faster results. If you're working on a large codebase, that's worth a look.
Pieter Levels — levelsio — shared that he runs Claude Code directly editing live production on a VPS. No staging environment. In 12 months, two failures — each one five seconds of downtime, zero data loss. His backup stack: 3-2-1 backups plus Litestream streaming to Cloudflare R2. That's either brave or well-engineered. Probably both.
And there's a story making the rounds about a 24-year-old who solo-built an AI content agency using Cursor and Claude. Took a 10-second Wednesday Addams clip, ran it through an image-to-video generator, built out a full gothic persona — one viral build over a weekend. He's pulling about $9,200 in monthly cashflow, with 88 percent of revenue coming from custom paid messages. The whole thing is a case study in what a one-person AI-native business actually looks like in 2026.
One last note from the developer community — sudoingX made the point that the Cursor CLI plus Composer combo ships faster than Claude Code alone, and that the speed difference actually changes how you approach prompting. That's a nuanced take — not just "this is faster" but "this changes your mental model." Worth sitting with.
That's your AI digest for 24 May 2026.