Back in February, Godot co-founder Remi Verschelde called the flood of AI pull requests “draining and demoralizing” and said he didn’t know how long the team could keep triaging it. On June 30 the Godot Foundation stopped asking nicely and started rejecting.
Fully AI-authored PRs are now banned outright. The reasoning cuts deeper than “AI is sloppy,” and the actual scope is narrower than the headline.
In today’s indie hacker news:
- 🕹️ Godot banned fully AI-written PRs, and its maintainers begged for it
- 💸 Sonnet 5’s $2 price is a teaser that expires August 31
- 🍌 Google shipped a 3-cent image model the same day as Sonnet 5
- 🕵️ Claude Code hid markers in proxied requests, and the “fix” still ships them
- 🔖 A browser Kubernetes, an OSS agent sandbox, and Fable 5’s export ban reversed
TOP STORIES
THE MAINTAINERS SAID NO
🕹️ Godot banned AI-written code, and its own maintainers begged for it
The story: On June 30 the Godot Foundation published a contribution-policy update that prohibits code authored entirely by AI, PRs from autonomous agents, and “vibe coding.” The contributor docs now spell it out: “contributions made entirely by AI are prohibited.” Read the scope carefully, because it is narrower than “Godot banned AI.” Code completion, regex, and find-and-replace are fine with no disclosure. Any heavier AI use in your code has to be disclosed in the PR. Only fully-generated, agent-authored, or vibe-coded PRs get hard-rejected.
The details:
- The headline quote is a Foundation statement, not one person: “AI cannot take responsibility, and we can’t trust heavy users of AI to understand their code enough to fix it”
- AI-generated text is banned from issue and PR discussions too, framed as “a basic principle of respect” to volunteer reviewers; machine translation of your own words is still allowed
- New contributors with three or fewer merged PRs can no longer open features or major refactors without a maintainer signing off first
- Godot joins a growing list: NetBSD presumes LLM code “tainted,” Gentoo bans it, QEMU declines it, and curl killed its bug bounty after ~20% of submissions turned into AI slop
Why builders care: “Disclose your AI use” is quietly becoming the default OSS contribution rule, and Godot’s policy text is specific enough to fork straight into your own CONTRIBUTING.md. If you build autonomous PR bots or Devin-style agents aimed at big public repos, the addressable surface is shrinking, not growing.
Turn one English video into 30 languages, in your own voice. ElevenLabs' dubbing studio transcribes your video, translates it, then voices each language in your cloned voice. Built for creators who want a Spanish or Portuguese channel without hiring a VO artist per region. Free tier covers your first few minutes.
We get a cut if you sign up. Only added for tools we use ourselves.
READ THE FINE PRINT
💸 Sonnet 5 launched at $2, but that price expires August 31
The story: Anthropic shipped Claude Sonnet 5 on June 30 as its new mid-tier model, now the default on Free and Pro. The number everyone is quoting is $2 per million input tokens and $10 output. Anthropic’s own page says that rate is introductory through August 31, then it reverts to $3/$15, which is the exact same headline price as Sonnet 4.6. The sharper catch is underneath: Simon Willison found Sonnet 5 ships a new tokenizer that emits roughly 30 to 40% more tokens for the same English text. So even at a flat per-token price, your cost per English task goes up.
The details:
- The tokenizer inflation is about 1.4x for English, 1.27x for Python, and ~1.01x for Mandarin, so English-heavy agent workloads take the biggest hit
- Two migration snags: sampling params temperature, top_p, and top_k are gone, and adaptive thinking is on by default unless you disable it, which will break scripts that set those
- Anthropic reports Sonnet 5 at 80.4% on Terminal-Bench 2.1, up from Sonnet 4.6’s 67.0%, and calls it “the most agentic Sonnet yet” at a fraction of Opus 4.8’s $5/$25, so the pitch is cheaper per task, not per token
- Open-weight pressure is real: GLM 5.2 lands within a point on SWE-bench Pro at $1.4/$4.4, and LongCat-2.0 is close behind at $0.75/$2.95, both self-hostable
Why builders care: Budget off the $2 headline and you under-forecast real spend twice, once when the intro rate ends and again on the token tax for English work. Model the reverted price and the English multiplier together before you commit an agent workload, or your first invoice will do the math for you. For commodity coding, GLM 5.2 and LongCat landing within a few points at a third of the price is the call that matters now.
CHEAP ENOUGH TO THROW AWAY
🍌 Google shipped a 3-cent image model the same day as Sonnet 5
The story: Hours after Sonnet 5, Google DeepMind launched Nano Banana 2 Lite, officially Gemini 3.1 Flash-Lite Image. It runs $0.034 per image at 1K resolution, or $0.017 on the Batch API, exactly half the full Nano Banana 2’s $0.067. Google claims about four seconds per image. The more interesting piece for builders is that Lite chains into Gemini Omni Flash, Google’s any-to-any video model, so you can generate a still and animate it in one vendor pipeline that OpenAI doesn’t offer as a single flow.
The details:
- Full Nano Banana 2 scales to 4K; Lite caps at 1K, so it is a draft-and-ideate model, not a final-output one
- Gemini Omni Flash is separate: text, image, audio, or video in, video plus audio out, up to 10 seconds, priced at $0.10 per second of output
- Google’s own docs list the weaknesses plainly: it struggles with small faces, accurate spelling, and fine details, plus masked edits and multi-image blending
- Willison’s hands-on test misspelled a word two different ways, so text-in-image is still not reliable, and there is no programmatic aspect-ratio control like the full model has
Why builders care: If you generate images at volume (ad variants, product mockups, dynamic social cards), Lite is cheap enough to make 20 and keep one, and you can A/B a batch for under a dollar. Keep the full model or gpt-image-2 for anything customer-facing where spelling matters. Same launch day, opposite race: Sonnet 5 pushed on capability, Google pushed on cost-to-serve.
TRENDING TODAY
🌐 Sam Rose ported Kubernetes to the browser - ngrok’s Sam Rose rebuilt a working slice of Kubernetes (scheduler, deployment controller, kube-proxy, an in-tab runtime) as ~140KiB of from-scratch TypeScript, backed by 204 integration and 1,855 unit tests ported from the real Go suite. It is not a WASM compile of real K8s, and it skips ConfigMaps, Secrets, and persistent volumes on purpose. Not production, but the best free way yet to see how a cluster schedules pods.
🔓 The US lifted its Fable 5 export ban, with strings - Commerce withdrew the export controls it slapped on Claude Fable 5 and Mythos 5 back on June 12, and Anthropic began restoring global access this week. This is the reversal of the 18-day freeze we covered in edition #75. The reversal is lopsided, though: Fable 5 gets full global restoration, while Mythos 5 goes back to US-approved organizations only pending review, and Anthropic agreed to proactively detect and report risks.
📦 TakoVM sandboxes your agent’s code in throwaway containers - An Apache-2.0 sandbox that runs each agent job in its own ephemeral Docker container, no network by default, with gVisor isolation and a per-job allowlist when you need one. Self-hostable with zero per-execution fees, though you operate the infra yourself (Docker plus a Postgres it spins up). A real answer to the “my agent just ran rm -rf” problem.
DRAMA
THE CHANGELOG THAT LIED
🕵️ Claude Code hid markers in proxied requests, and the “fix” still ships them
A developer reverse-engineered the Claude Code binary and found that when you set ANTHROPIC_BASE_URL to a non-Anthropic host, the tool quietly embeds hidden markers in every request, swapping in visually identical characters based on your system timezone and proxy host. It only touches proxy, gateway, and self-hosted setups, not default users, and it has run since April with no changelog mention. The HN thread hit 2,163 points. Anthropic’s reported line, secondary-sourced, is that it’s an anti-distillation and anti-reseller-abuse experiment it plans to roll back.
Why builders care: Our reader agent confirmed the marking is still present in v2.1.197, the release some outlets called the fix. If you route Claude Code through a corporate gateway or cost-router, classification data about your setup has ridden along in every request for three months. For a CLI with full shell and git access, a clean-looking changelog no longer means nothing changed.
STACK OF THE DAY
The Godot ban is about AI that doesn’t understand the code it writes; this is a tool aimed at the same problem. World Model MCP builds a temporal knowledge graph of your codebase so coding agents stop hallucinating APIs and reintroducing old bugs, then validates changes against it. The v0.10.0 release ships adapters for seven agents (Claude Code, Cursor, Codex, pi, OpenClaw, Hermes, and Continue) that share one SQLite fact graph per project. MIT-licensed, runs locally, no hosted dependency for the core. Setup caveat: OpenClaw needs an absolute Python path and telemetry is off by default.
BOOKMARKED TODAY
🎨 figit: paste HTML into Figma as editable layers - An open-source library (figitdesign on GitHub) that turns HTML into fully editable Figma layers by copy-paste, no plugin install. Free today; a hosted team API is listed coming soon.
🛡️ Pokayoke: turn repo conventions into deterministic checks for agents - MIT-licensed and self-hostable via npx, it turns your repo’s conventions into checks both humans and coding agents can run and repair. Currently Bun-only and early (v0.0.6).
🖥️ InfraCanvas: your VMs, Docker, and K8s on one live canvas - Auto-discovers Docker, Kubernetes, systemd, and VMs and draws them as a color-coded topology canvas that refreshes every 30 seconds, with in-browser terminal, logs, and restart. AGPL-3.0, free to self-host.
Stop trusting random coffee-shop WiFi with your auth tokens. Public WiFi is a free packet sniffer for anyone in earshot of your laptop. NordVPN encrypts the tunnel so a logged-in Vercel session, a GitHub PAT, or a Stripe dashboard tab stays yours. Works on the same 6,400+ exit nodes you'd use for geo-testing.
We get a cut if you sign up. Only added for tools we use ourselves.
Curated by AI, built by a human.