AgentRank weekly: free models to try, Copilot model cutoffs, and which AI game tools you can actually sell with
Plus a new game-dev category and the catch behind every free hosting and database plan.
Hey builders,
Here's what changed this week for people shipping with AI. Everything below was checked against the official changelog or pricing page.
Top new releases
- Claude Code 2.1.304 adds /auto-mode-setup (Oct 19). A guided setup reads your project and recent sessions to configure auto mode, which now also works on Bedrock, Vertex and Foundry. Pro and Max users also get a warning before resuming an old conversation would eat a big share of the 5-hour limit.
- GitHub Copilot removes six models (Oct 19). Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5 are gone from Chat, inline edits, agent mode and completions. The suggested replacements are Gemini 3.8 Flash, GPT-6.1 Sol, GPT-6 Luna and Grok 4.7. A second round retires Opus 4.8, Opus 5 and GPT-5.6 Terra on Nov 16, so switch to Claude Opus 5.5, GPT-6.1 Sol or GPT-6 Luna before then.
- Two models you can try free on Vercel AI Gateway. MiniMax M3.1 Flash (1M-token context, built for coding and agents) is free until Oct 22 when MiniMax serves the request, and Reflection AI Beam (Apache 2.0, low-latency reasoning) is free for a limited time. Both work in Claude Code, Codex and Cursor through AI Gateway.
- Neon Realtime public beta (Oct 16). Subscribe to any SQL query and the result stays live as data changes, with no polling. It works with branching and scale to zero, in 4 AWS regions for now.
- Python on Vercel is GA (Oct 15). FastAPI, Flask and Django are auto-detected, with uv and pyproject.toml support. Vercel also opened Daemons in public beta for long-running workers like queue consumers and self-hosted CI runners.
- Claude Code Review is GA (Oct 15). Automatic PR reviews for Claude Team and Enterprise left research preview, about 3x faster, with Quick, Standard or Deep levels per repo.
- OpenCode 2.0 (Oct 15). A rewrite with a unified background server, a code-mode JS interpreter, background sub-agents and 80+ plugins.
Also worth a look: Lovable now lets your apps use GPT-Live voice, the latest Claude models, Gemini 4 Argon, Perplexity and Firecrawl, Supabase can sync live Stripe data into Postgres, and Meta released Muse Spark Open, an open-weight 816B model with a 1M-token context. The full list is on the releases feed.
Ranking changes
No tool moved up or down in the existing categories this week. What's new:
- A new category: AI tools for making games. Six tools, scored with commercial rights and licensing clarity weighted highest: Scenario (#1), Meshy, Tripo Studio, ElevenLabs, Cascadeur and Ludo.ai. The page includes a table answering "can you sell what the free plan makes?" Short version: Meshy's free plan is fine with attribution (CC BY 4.0), while the free plans of Scenario, Tripo, ElevenLabs and Cascadeur are non-commercial. Ludo.ai has only a free trial.
- The catch behind every free plan. Hosting and databases now show a dated line on each free plan's limits. A few examples: Vercel Hobby is non-commercial only, Render free services spin down after 15 minutes idle, Supabase free projects pause after a week of inactivity, and Fly.io and PlanetScale have no free plan.
- Prices filled in from official pages. Lovable Pro is $25/mo for 100 credits, and Cloudflare Workers Paid starts at $5/mo per account. No listed price went up or down this week.
Price heads-up: Gemini 4 Argon's intro API price doubles on Dec 1, 2026, and Gemini 3.8 Flash's doubles on Jan 1, 2027.
Tool of the week: Neon
Neon is #2 in Databases & backend: serverless Postgres with instant branching, so every preview branch or agent can get its own database. This week's Realtime beta adds live query results without polling, which makes it a stronger pick for realtime apps.
- Pricing: free plan; paid Launch plan is $0.106 per CU-hour.
- The catch: free compute always scales to zero after 5 minutes idle (you can't turn that off), and each project gets 100 CU-hours before it's suspended until the next period.
- Best for: Postgres per preview branch or per agent, on Vercel or any serverless stack.
That's the week. Reply and tell me what you're building or which tool you'd swap in your stack. Rankings on AgentRank are never for sale.