... come from tuning a rubric until the score looked nice it came from an engineering audit and real fixes. Slower, less flattering, far more durable. We also ran a head-to-head classifier benchmark and moved to Claude Haiku 4.5 after the alternative

... models for each use case with a sample data set. I'll look at a mix of things like quality, price, speed and then pick. This has so far landed me at a mix of Sonnet, Opus, and Haiku for text-based work and Gemini models for images. I'm early in development, though, so I'm curious if other approaches have worked better as you scale! Comment from Oliver Graf(@oliver_graf1): Did you notice a big difference ... ... days. Thanks for the image link. I will check it out Comment from Hans Desjarlais(@ismaelyws): @alyssa_at_kalina  Good insight. I'm also using Anthropic's Claude Code for development and a mix of Sonnet, Opus, and Haiku

... stop being their own QA department. How it works: Paste your URL Describe any user flow in plain English (you can also use templates or simply record yourself performing the flow with our Record It Yourself feature) Claude Haiku drives a real browser, intelligently clicking and typing on the live page Get a beautiful report in minutes: video replay, per-step screenshots, console/network errors with full replayable trace, and WCAG accessibility checks Key Features: AI that adapts when your ...

... things I'm proud of: Focus Score one number that reflects how well your actual day matched your planned day. Calendar-first plan and reality live on the same surface (ghost outline = plan, solid = actual). Weekly review with AI Haiku a 3-line poetic reflection on your week, written by Claude. It's oddly motivating. Day 1 in 25+ languages , iOS + Android, fully native feel on both. Free covers the daily loop (Day/Week view, 1 alert slot, rule-based insights ... ... unlocks Month+ views, multi-alert scheduling, and the AI Haiku

... blow up your margins. Comment from Roopesh Donde(@roopesh_donde): @davem_0 I am working on an AI writing tool, and I constantly think about API costs. Ran calculations with AI atleast 3-4 times, Changed model from Haiku to Gemini Flash 2.5, usage limits, etc. Eventually I have decided to move to an early access->trial->premium-only model from freemium. This way I can control costs and get better feedback. If it picks up traction ...

... split the current models more clearly by tier. Free now gets: Gemini 3.1 Flash Lite GPT-5.4 Nano (Remember! no sign-up required . Just install extension and use it right away) Pro gets: Kimi K2.6 Grok 4.3 Claude Haiku 4.5 GPT-5.4 Mini Gemini 3.1 Pro Max gets: GPT-5.5 Claude Opus 4.7 Some specific cleanup: Grok 4.1 Fast was removed after provider issues Gemini 3 Flash was replaced by Gemini 3.1 Flash Lite Kimi was upgraded to Kimi K2.6 ...

... when I started building Luminens a 16-bit browser RAW photo editor running on WebGPU and libraw-wasm. Talk about a nightmare of structural complexity. I moved away from automated builders (I've tried all of them), switched to VS Code , and started using ChatGPT 5.4 xHigh , which I found incredibly reliable and capable, yes better than Claude. My entire method shifted: every single time I hit a wall, I didn't just ... ... withmarlowe): I have no technical coding background but I had a pretty good experience using Claude chat to discuss what I wanted and have Claude create the code. First thing Claude asked was for me to install Node.js and VS

... over and over. 1. Token burn is the #1 frustration Multiple users maxed out weekly limits in 2-6 hours. One burned 25% on 7 slides. The community built a workaround strategy: Opus for first prompt, Sonnet for edits, Haiku for tweaks. When your users need a cost mitigation strategy on day one, that's a signal. 2. "Design system integration" disappointed people Claude Design reads your codebase and extracts visual patterns. It doesn't use your actual components. Designers ... ... trying to explain something to a LLM when I can do it faster and more precisely Comment from Ng Jun Sheng(@ng_junsheng): The community inventing a cost mitigation strategy on day one (Opus for drafts, Sonnet for edits, Haiku

... production workflows in queue mode on a self-hosted Hetzner setup. Recent shipped projects: - LastSend - live posthumous-messaging app on Google Play, just landed first paying customer this week. - VendorIQ - 6-agent contract analysis orchestrator (Claude Sonnet + Haiku, tool calling, structured outputs, human approval gate) submitted to Airia AI Agents Hackathon - CaseDrop - AI case intake pipeline for personal injury law firms. Police report PDF GPT-4o Vision extraction human approval Clio Manage v4 (OAuth 2.0) retainer generation e-signature. Built ... ... Specialist. Strong fit at AI startups needing someone to integrate and ship, not train foundation models. linkedin.com/in/rohankumarvg / portfolio at rohanbuilds.com Comment from Maliik(@maliikb): @fmerian  Thanks! And congrats on the Kilo Code v7 launch today. Rebuilding the VS

... rebuilt it completely. Now it runs 3x/day, 3 quality matches per agent. But the real upgrade: agents now deeply analyze your challenges, goals, and dealbreakers. Early users reported that similar profiles were matching each other: competitors finding competitors

... developer, but I ve always been curious about technology. What started as a simple hobby just 5 days ago became something I really enjoyed building. I realized that with the right tools (I used HTML, Netlify, and Claude Haiku), I could actually create something useful for people in my country. That s how CVitae was born. It s a simple tool to help people in Paraguay and LatAm create better resumes and beat those automated filters that often leave good ...

... embedding models though - Frontend: React + Vite + TS + shadcn, TanStack Query, types generated from the FastAPI OpenAPI spec so frontend + backend never drift - LLMs: OpenAI (GPT-5-nano for cheap structured calls, heavier models for big prompts) with Claude Haiku as fallback. All prompts and traces in Langfuse - Auto-apply: I tested some apis like skyvern , browser use. Ended up building my own - Hosting: Render (web, celery workers, celery beat, static frontend) Auth, billing, perms what you asked about: - Auth: Supabase ... ... auth its just device id stored in keychain no accounts needed which keeps onboarding super minimal biggest challenge has been text extraction different sites block scrapers differently. revenucat for subscriptions has been great tbh, saves a ton of time vs

... Edgee is a Token Compression technology which can save up to 50% of the token bill. And it works with Claude Code pro/max plan as well Comment from Sacha MORARD(@sachamorard): @alina_petrova3  I agree! There are alternatives to Sonnet, for sure. However, Sonnet is now recognized as the best (coupled with Claude Code) by a huge portion of developers, and I know many who don't want to take the risk of trying a cheaper model. This ... ... where token compression strategies come into play. What if we could use Sonnet for the price of a Haiku

... matters Flexibility Start a thread on GPT-4, continue it on Claude the context follows. Smart routing Send queries to their respectful models and complex ones to premium options without rebuilding anything. True A/B testing Compare ChatGPT vs Claude using the exact same memory and history. Built-in failover Provider outage? Switch models instantly, without losing state. Future-proof architecture New LLM launches? Plug it in. Mnexium handles the memory layer. How it works Just include your provider ... ... ndini): @haiqa_irfan   Great question! Priority routing is on our roadmap as we add more providers & features. We're planning on smart routing options: Cost-based routing Route simple queries to cheaper models (GPT-4o-mini, Claude Haiku

... would really be awesome if you had any feedback or thoughts or notes of support (or not) as I build it out further! For AI model comparisons I m using the two Claude models above (plus Haiku), along with GPT-5.2, Grok 4.1, Gemini 2.5 Flash, and models from Cohere, Perplexity, Meta (Llama), DeepSeek, Mistral, and Jamba for lots of different AI flavors! The goal is for users to add personal experiences and context through a chat-based conversation with ...

... would highly appreciate some information how you keep costs low. What we are now going to do: Run everything on VMs Rent A100s through Google Cloud for Startups Program (if they let us do this) Switch to Haiku What else can you recommend ...