Peanut T2I Model, Autonomous Agent Businesses & Claude Code Dominance
Welcome back. Let's get into it.
First up — a mysterious new image model just dropped, and nobody knows who built it. It's called Peanut. Anonymous developers released it quietly, and it debuted at number eight overall on the Artificial Analysis Text-to-Image Arena. That puts it ahead of Z-Image Turbo, Qwen-Image, and FLUX.2 dev — which is a serious claim for a model that showed up with no company name attached. Weights aren't out yet, but the release is expected soon. Keep an eye on this one.
Shifting to voice — Inworld AI just launched something genuinely interesting. It's called Realtime TTS-2, and the team behind it is the same group that built the number-one ranked TTS on Artificial Analysis. What makes TTS-2 different? It's the first voice model that actually processes full multi-turn audio context before it responds — meaning it reads your emotion, your tone, your pacing across the whole conversation, not just your last sentence. You can give it voice direction in plain language. No emotion tags. No XML. Just — "sound more tired here" — and it adjusts. It holds a single voice identity across more than a hundred languages and can switch on the fly. You can even generate a completely new voice from a prose description. And the time-to-first-audio is under 200 milliseconds. That's fast enough that you wouldn't notice the gap in a real conversation.
Now — this next story is going to make you uncomfortable, and it should. A 21-year-old student created a fictional woman on OnlyFans and made $43,000 in the first 30 days. Here's the stack: Claude for behavior-constrained chats — so the persona never breaks character. Flux with a LoRA for consistent face images, locked to a specific appearance including a scar for uniqueness. ElevenLabs for personalized voice notes. And four Markdown files handling persona rules, visual guidelines, voice parameters, and — this is the part that really lands — customer memory. The system remembers subscriber names, habits, and spending patterns. The result? 1,247 paid subscribers at an average of $34 each. One fan alone spent $1,847. The whole operation runs from a phone. Make of that what you will — but from a pure systems architecture standpoint, it's one of the more sophisticated automation setups we've seen at this scale.
Alright, moving into the world of autonomous agents — and there are a few stories here that stack up in a pretty striking way.
Shubham Saboo, a Senior AI PM at Google, has been running a one-person company for months using OpenClaw and Hermes agents on 24/7 schedules. Engineering, sales, marketing, ops, design — all handled by agents. When agents conflict with each other, the system uses self-review and self-correct loops, plus kanban boards to track state. One person. Full company. No contractors.
Then there's an 18-year-old American solopreneur who bought a $400 Mac Mini, connected it to OpenClaw through Telegram, and built a 3D product design business. He sends prompts like — "add table under Mac Mini and screen" — the system auto-opens Blender, renders the design with AI, and he sends it to clients. First month revenue: $16,000. Traditional designers charge $50 to $150 an hour for the same output.
And then — this one's for the hardware nerds — a developer going by sudoingX ran a Hermes agent on a ROG Ally with an RTX 5090 mobile chip. The model was carnice-v2-27b at Q4_K_M quantization, running through llama-server with a 262,000-token context window. The agent was asked to benchmark its own hardware and configuration — and it did it autonomously, making 19 tool calls across 42 messages in 12 minutes. Eleven terminal probes pulling from nvidia-smi, nvtop, curl. Six TODO updates. File writes and reads to verify its own output. Peak VRAM: 21 gigabytes. GPU utilization: 99%. Speed: 16.71 tokens per second. The agent essentially audited itself. That's a loop that would've required a systems engineer two hours to do manually.
Now let's talk Claude Code — because there is a lot happening here.
A developer built a seven-agent sales automation system orchestrated by Claude Sonnet 4.6 through Claude Code Router and MCP, running on a MacBook and an iPhone. The agents are named and specialized. Scout crawls Google Maps and pulls 220 businesses a day, filtering down to 30 qualified leads based on years in business, review count, and website age. Diagnoser writes a 50-word diagnosis and a cold message under 70 words. Builder creates Lovable landing pages for the top five leads. Pitcher sends 30 messages a day — email for roofers, SMS for trades, Instagram for salons, LinkedIn for realtors — and gets a 14% reply rate. Checker evaluates every message for personalization and AI markers. And a Mobile agent handles real-time iPhone replies and Calendly bookings. The system serves 47 businesses a month at $400 each — $18,800 in monthly revenue — with $480 in API costs. It burns through 3 million tokens a day. A human only gets woken up for deals over $3,000 or if the reply rate drops below 12%.
Erik Schluntz — the lead on Claude Code at Anthropic — said publicly that he has written zero lines of hand-code for months. He shipped 49 full features in two days using Claude Code, and gave a 30-minute conference talk explaining how. Zero. Lines. Hand-coded.
A developer called LandseerEnga posted that Claude Code autonomously wrote and ran a test plan on a live iPhone app — no user-written tests, no manual device interaction — and caught a real bug in the process.
And a builder named mylesxstute is running four ventures with Claude Code as his entire engineering team. He flagged that Anthropic doubled the Pro and Max five-hour usage limits and removed peak throttling — which he says eliminates the daily bottleneck that was the last real friction point in his workflow.
That's your AI digest for 06 May 2026.