Sonnet 5.5, Deepseek Flash Coder, $1 AI VSLs — AI Daily Sep 28
620 messages · 77 active members
Overview
Topics
Anthropic released Sonnet 5.5 same day as Opus 5.5 momentum peaked, with early benchmarks close to Opus but mixed reports on speed after a few hours. Kling 4.0 Flash, ElevenLabs v4 (10-sec voice cloning), and rumors of Astra 6.1, Gemini 4 Pro, and OpenAI's Grok-bot competitor all landed as builders frontrun OpenAI dev day. Community read: OpenAI looks more compute-constrained than Anthropic right now.
@thewildzeno benchmarked coding models and crowned Deepseek Flash on low reasoning for best cost/performance — beating Luna (weak below medium reasoning), Grok 4.7 (2x cost, 2x time, TSC errors), and Muse with top cache hits. @expadz's parallel tests showed Luna 6 fixing bugs 10x faster than Opus 5.5 at ~100x lower cost. @Anonymoushat hit 30k-70k TPS sustained on Baseten. Emerging pattern: plan with Opus/Fable, execute with Deepseek Flash or Luna.
Builders are generating full 3-minute VSLs for ~$1 using Opus 5.5 in Claude Code with SVG/React + Remotion/Hyperframes — no editors needed. @thewildzeno built a 90-sec kawaii onboarding video first-try on ~1% of the $20 quota; @iammikevineyard shipped playable 2D games in 30 min via Opus 5.5 + Codex sub-agent + Hyperframes + ElevenLabs. @yangthegoat: 'A one-person 8-figure ecom brand might be very feasible soon.'
After motherboard deaths and WSL memory leaks killing agent runs, builders are moving to Hetzner, Vultr, and EC2. @thewildzeno recommends 4 vCPU VPS to start, forcing portable setups with git-committed secrets. @geilt runs per-dev EC2 boxes with dynamic domain routing. Ubuntu on a $20/mo VPS consistently outperforms WSL and beefy local PCs for parallel agent orchestration.
Codex CLI revokes OAuth tokens the moment you log in from another machine, making multi-account VPS setups brutal. Workaround: run isolated terminal sessions per account so tokens persist for weeks. Oh My Pi and Claude are emerging as smoother alternatives for managing 10+ OAuth sessions cleanly. New claudex-loop skill also landed for four-phase plan hardening with adversarial cross-model review.
Words worth knowing
Three terms from the glossary.
Key Takeaways
- Deepseek 4.1 Flash on low reasoning is the new grunt coder king — best cost/performance, TPS, and cache hits, beating Luna, Grok 4.7, and Muse.
- Luna 6 fixes bugs ~10x faster than Opus 5.5 at ~100x lower cost — pair with a high-reasoning planner like Opus 5.5 or Fable 5.1.
- Full 3-minute VSLs can be generated for ~$1 using Opus 5.5 + Remotion/Hyperframes with SVG/React — no video editors required.
- Opus 5.5 via oauth ($200 sub) costs ~$1 for tasks that run $15 on API; OpenAI killing the 20x plan and reinstating $200 at half usage signals pricing convergence.
- Go remote-first for agent work: WSL leaks memory under parallel agents; Ubuntu on a $20/mo Hetzner/Vultr/EC2 box handles it far better and keeps you portable.
Hot Threads
Deepseek Flash benchmark results and grunt coder testing
Luna 6 benchmarks vs Opus and Deepseek Flash for bug fixing
Video editor agency pricing and SOP-driven creative workflows