Krea 2, Hermes Agents, Claude Code Limits & Cursor Cloud
Welcome back. A lot happened today — generative image models, agent use cases getting genuinely wild, and some serious drama in the Claude Code billing world. Let's get into it.
Starting with generative media — Krea.ai dropped early access codes for their Krea 2 foundation model. This is their first model built entirely from scratch, and the pitch is interesting: it's tuned specifically for aesthetic diversity and stylistic control. You feed it single or multi-image moodboards, and it analyzes the underlying concepts — not just the surface-level style. Three codes went out: K2-PRFUF8, K2-NRWW9E, and K2-CAP48S — each good for 50 uses. @zazzygfx tested it same day and put it well: "it doesn't fight your taste." That's the key differentiator here. Flexible stylistic interpretation over that polished, homogenized AI look everything else seems to produce.
Also in the 3D space — fal.ai launched Pixal3D. Single image in, 3D model out — using pixel-aligned back-projection for direct pixel-to-3D mapping. The output includes detailed PBR textures, so it's actually pipeline-ready. Not a toy. Worth watching.
And there's a viral open-source project making the rounds from @m0h — a local video and image generator sitting at over ten thousand stars on GitHub. It's free, runs locally, and supports more than 200 models — Flux, Seedance Pro, Kling, Sora, Veo, Midjourney — all in one place. Cinema-style controls: lens, focal length, aperture. Plus Claude and Codex skills integration baked in. For anyone who doesn't want to pay for a dozen separate subscriptions, this is a big deal.
Now — the Hermes agent stories, and these are genuinely fascinating. Nous Research and Kimi Moonshot picked winners from 227 submissions to the Hermes Agent Creative challenge.
First place went to @daumerval's project called Brut-V. A Hermes agent generated 6,200 lines of RISC-V assembly code — and then bootstrapped a JavaScript assembler that achieved byte-perfect parity with a reference simulator — using self-correcting diff and patch loops. Let that sink in. The model debugged its own assembly output iteratively until it matched exactly.
Second place — @ya1sec's Ambien. A daydreaming sub-agent that samples an Obsidian corpus and synthesizes ideas into essays — using iterative voice calibration to match a writing style over time.
Third — @hughpyle's Teletype emulator. Hermes running on a Raspberry Pi, reviving a 1960s ASR-33 Teletype. There's a plugin that emulates the Ambisonic hum and mechanical keystrokes of the original hardware. That one's just wonderfully weird.
Staying in the Hermes world — @adiix_official documented something that's going to raise some eyebrows. A Polymarket trader — wallet address 0xce25e214d5cfe4f459cf67f08df581885aae7fdc — was identified using an OpenClaw agent. The setup: $1,000 deployed across five accounts — $5,000 total in paper trading. The agent ran 212 autonomous trades on prediction markets over 12 hours. 179 wins. 84% win rate. Grew to a $45,000 profit. And as of today — May 14 — real capital is being deployed. That's not a demo. That's production.
Meanwhile, @sudoingX showed off local optimization inside Hermes Agent — paste in an OpenAI-compatible endpoint, and it auto-detects all models from llama.cpp or vllm servers. Surfaces 10-plus models instantly, with per-model parsers, prompt templates, tool call formats, and thinking modes all handled automatically. The quote was direct: "local first from architecture" — and he called OpenClaw a "bloated tool." That's a shot. But the demo backed it up.
Moving over to Claude Code — Anthropic's @ClaudeDevs announced a 50% increase to weekly limits, stacking on top of prior doubled 5-hour limits. Live immediately through July 13, 6PM Pacific. Covers Pro, Max, Team, and seat-based Enterprise — across CLI, IDEs, desktop, and web. Good news. But the response wasn't universally positive. There's real frustration building over a separate change — the "claude -p" programmatic flag shifting to metered credits starting June 15. People who built pipelines around that flag are not happy.
Also on OpenRouter — they launched anthropic/claude-opus-4.7-fast. Full Opus 4.7 intelligence — but at roughly 2.5 times faster throughput, via model alias. This comes at an interesting moment — there are reports of companies like ServiceNow and Uber burning through their annual Anthropic budgets early. Speed at this level matters when you're burning through tokens at enterprise scale.
And here's something genuinely impressive from DatologyAI. They released vision-language models improved through data curation alone — same architecture, same training recipe, same compute budget. Their 2B model gained 11.7 percentage points on average across 20 evaluations — using roughly 150 times less training compute than the baselines. It beats InternVL3.5-2B by about 10 percentage points at 17 times less compute — with no SFT or RLHF involved. Their 4B model hits near-frontier accuracy at 3.3 times lower response FLOPs than Qwen3-VL-4B. The argument here is that the data is the lever — not the model size.
Now — the Claude Code billing story. This one's going to sting for some people.
@darkzodchi documented seven mistakes that burned through a $200-a-month Pro plan in 70 minutes — across just four prompts fixing a small bug. Here's the breakdown.
Number one: defaulting to Opus. It costs five times more than Sonnet and drains the 5-hour limit five times faster. Number two: waiting until 95% context capacity before running /compact — which means 232,000 tokens getting re-read every single turn in a 30-message session. Number three: loading five unused MCP servers — that's 90,000 tokens of overhead per turn, doing nothing. Number four: feeding raw 10,000-line log files directly, with no preprocessing — 99.5% of those tokens are waste. Number five: subagent fan-out on simple tasks — that's a 7x token multiplier for work one agent could handle. Number six: not using CLAUDE.md or /memory — so you're re-explaining the project at 3 to 5 thousand tokens every session startup. And number seven: not running /clear between tasks — carrying over irrelevant context from previous conversations.
The fixes are simple: default to Sonnet, run /compact early, unload your MCP servers. One developer reportedly cut their monthly spend from $1,200 down to $189. Worth reading the full breakdown.
On the automation side — @eng_khairallah1 walked through Claude Code Routines — Anthropic's cloud automations that run independently of your laptop. You configure a prompt, a repo, connectors like GitHub, Slack, Linear, or Google Drive, and a trigger — scheduled, API-based, or GitHub event. The use cases getting traction: automated 9am PR reviews with inline comments capped at three per file — approve only if no critical bugs. Post-deploy webhook smoke tests that ping Slack with a go/no-go verdict. 2am Linear issue triage. Friday doc updates triggered by merged PRs. Current limits in preview: 15 runs per day, claude/* branches only.
And Cursor shipped cloud agents running in engineer-ready dev environments — cloned repos, installed dependencies, toolchain credentials, multi-repo support — with per-environment version history, rollback, audit logs, scoped secrets, and egress controls. Cursor engineer @sjwhitmore highlighted internal VM use for trusted merges and demos. Customers already using it in production include Decagon, Amplitude, BILT, and Snyk.
One more quick one — OpenCode hit version 1.14.49. Key additions: auto prompt-caching for Anthropic and Bedrock across tools, system, and user boundaries. DigitalOcean OAuth and router discovery. TUI notifications, sounds, and session pinning. Stricter Zod and Effect Schema tool contracts. The macOS arm64 bundle comes in at 101.6 megabytes — down 0.3MB — and usage is now persisted per session.
That's your AI digest for 14 May 2026.