Opus 5.5 & GPT-6 Sol/Luna Drop, Jev Browser Automation — Sep 22
798 messages · 89 active members
Overview
Topics
Anthropic launched Opus 5.5 (~30% faster, ~40% cheaper than Opus 5, Fable 5.1-level performance) with banked resets and a rumored full 5.5 series. OpenAI countered with GPT-6 Sol at half Opus 5.5's price and Luna at $0.10/M input tokens — ~10x cheaper than Haiku. Early testers say Opus 5.5 fixes Opus 5's overthinking; OMP users needed a client update first.
@tounano and @WhiskeyATX debated architecture: for coding, orchestrators should be smart because bad plans are the top failure mode — smart planner, dumb implementer (Fable orchestrating Opus). For retrieval/dispatch tasks like Hermes, speed matters more so Gemini or Grok 4.5 wins. Model selection debates around Luna, Astra, and Terra reinforced fitting the model to the job.
Kieran wanted to automate monthly Xero bookkeeping and the group converged on Jev — it decides which buttons to click while a cheap LLM (Gemini, DeepSeek Flash) handles fetches. @tounano stressed Jev needs explicit rubrics to shine. @leewardbound shared a jev-ultrafast browser-use repo, and Jev is now accessible via Vercel's AI Gateway beyond Typesafe.
@rockdm shared how replying to FB ad comments with the destination URL generated an extra $30K/month, with the thread digging into attribution challenges via the FB API and post-ID matching. Separately, @Biundini's $90 CPM / 9% CTR campaign was diagnosed as a funnel problem — high CPMs often signal qualified audiences; fix drop-offs with PostHog before touching ads.
@jarvisballer detailed moving his Orca backend fully onto GitHub Actions with role-specific OMP profiles (Dev, QA, Review), abandoning Linear as an unnecessary layer. Others are following, using a single orchestrator workspace that manages sub-workers and worktrees. Took months to refine but is now considered the cleanest approach for agent-driven development.
Words worth knowing
Three terms from the glossary.
Key Takeaways
- Opus 5.5 ships ~30% faster and ~40% cheaper than Opus 5 with banked resets, while GPT-6 Sol undercuts it by ~50% and Luna hits $0.10/M input tokens — model economics keep collapsing.
- For coding agents, make the orchestrator smart and workers dumb: bad plans are the primary failure mode, not bad implementation.
- For browser automation, pair Jev (decides clicks) with a cheap fetcher LLM — but give Jev explicit rubrics, not open-ended reasoning tasks.
- Replying to FB ad comments with your destination URL is a major unlocked revenue lever — one builder attributed $30K/month to it.
- High CPMs with strong CTR usually mean qualified audiences; if conversions are missing, fix the funnel with drop-off tracking (PostHog) before touching ads.
Hot Threads
FB ad comment reply strategy and attribution workarounds
High CPMs, low conversions — is it ads or funnel?
Most token-efficient LLM for browser tasks (Xero bookkeeping)