Product Hunt
Kilo Code has become one of the fastest-climbing tools on Product Hunt this month — an open-source agentic coding assistant that meets you in VS Code, JetBrains, or the CLI, connects to 500+ models with zero markup on top of the provider's own rate, and needs no API key to start. The pitch that's landing: pick a model, switch mid-task, pay only what the model costs. Reviewers keep calling out the same thing — non-professional coders using it to ship real internal tools (expense trackers, warehouse systems) without a dev team.
Also climbing this week: Catenary, a spatial canvas IDE built specifically for orchestrating AI coding agents visually rather than through a chat window — worth watching if you find yourself running more than one agent at a time and losing track of what's doing what.
Frontier models
Two major model drops landed almost on top of each other:
- OpenAI shipped GPT-6 Astra on September 3 — its first model to cross what OpenAI calls its internal "Critical" cybersecurity threshold, which is why initial access is gated to vetted partners in its Daybreak program before wider rollout. On benchmarks it saturates ARC-AGI-3 (99.9%) and hits 100% on ExploitBench, and on computer-use tasks it's reportedly ~47% faster than GPT-5.6 Sol at a higher success rate (72.6% vs 65.7% on OSWorld 2.0). Pricing is $10/$50 per million tokens with a 1M token context window.
- Anthropic answered with Claude Fable 5.1 and Claude Mythos 5.1 on September 1 — stronger agentic coding and research, better computer use and spreadsheet/document handling, and a 75% cost cut on Fable cache reads. The other headline: Anthropic is now watermarking text generated by models released after August 2, with a detection API in private preview for orgs that need it under EU disclosure rules.
Translation for anyone building on top of these APIs: agentic and computer-use capability is now the actual battleground, not raw chat quality, and cache-read pricing is quietly becoming a real cost lever worth checking before you re-architect a pipeline around one vendor.
GitHub Trending
The single most notable trending repo this week isn't a framework — it's a folder of Markdown. mattpocock/skills ("Skills for Real Engineers, straight from my .agents directory") sat at #2 on GitHub's daily trending page on September 6, picking up over 2,200 stars in a single day on top of a base already past 200k. It's a plain-language skill pack for Claude Code / agent harnesses — install any skill individually with npx skills@latest add mattpocock/skills/. The trend underneath it: the fastest-growing category on GitHub right now isn't code, it's structured instructions for code-writing agents, and teams are treating well-written skill packs as reusable infrastructure the same way they'd treat a shared lint config.
Worth your time
If you're picking one thing to actually try this week: grab a couple of skills from mattpocock/skills and drop them into your own agent setup — it's a five-minute experiment with an immediately visible payoff, versus waiting on GPT-6 Astra access or re-plumbing a pipeline around Fable 5.1's cheaper caching. The model wars matter for planning; the skills repo is the thing you can use today.