Milan (@MilanGlow) header
Milan (@MilanGlow) profile picture

Never lose your love for life

Joined October 2026
Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)Photo by Milan (@MilanGlow)
I'm getting a little tired I think... I think I just want to take it easy for a while
131
Run Qwen4-Max in Claude Code using QwenCloud 1. Configure your API key and endpoint (macOS / Linux) export ANTHROPIC_AUTH_TOKEN="YOUR_QWENCLOUD_API_KEY" export ANTHROPIC_BASE_URL="ap-southeast-1.maas.qwenclou…" export ANTHROPIC_MODEL="qwen4-max" export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen4-max" 2. Test the connection claude -p "Hello" 3. Start coding claude The endpoint above uses QwenCloud's Singapore region and a pay-as-you-go API key Official Claude Code setup guide: docs.qwencloud.com/developer… Qwen4-Max model documentation: docs.qwencloud.com/developer…
41
This fucking world!
1
2
133
Your coding agent keeps forgetting your project, and you're paying for it in tokens Every new Claude Code or Codex session starts from scratch, rereading the same files, rediscovering previous decisions, and asking questions you've already answered Claude-Mem is an open-source tool that gives your coding agents persistent memory across sessions Here's how it works: • Automatically records useful information while your agent works • Saves notes locally in your project • Loads a compact index of relevant context when you start a new session • Reduces the need to repeatedly explore the same codebase It supports 15 coding agents, including Claude Code, Codex, Cursor, and OpenCode Installation is simple: "npx claude-mem install" The installer automatically detects supported agents on your computer Start a second session in the same project, and the saved memory becomes available Instead of making your agent rediscover everything each time, you can let it pick up where it left off Your tokens should go toward writing code, not repeatedly figuring out what your code does
4
1
137
Google just launched an AI API that can answer up to 1,000 questions in a single call, with ZERO output token costs 🤯 Meet Gemini Decisions API, a lightweight model designed to make fast, low-cost decisions without generating text Instead of asking an LLM to explain its answer, you simply give it an input and a list of yes/no questions It returns probabilities and confidence scores, ready for your app to use directly 💰 Pricing & Performance • Input: $0.10 per 1M tokens • Output: Completely free • Latency: 150–250 ms per call • Up to 1,000 questions per request • Supports text, images, audio, video, and PDFs The best part? Your shared input is only billed once, even when you're asking hundreds of questions about it Google says it's roughly 5× cheaper than Gemini 3.5 Flash-Lite on its benchmark suite 🚀 What can you build with it? • Model routing — Decide whether a request needs a powerful or lightweight model • Support automation — Categorize tickets and determine which need escalation • Content moderation — Check whether content violates specific policies • Search filtering — Score results before passing them to an expensive LLM • AI agent workflows — Make hundreds of small decisions without generating unnecessary text ⚙️ A few technical details • Model: gemini-decision-1 • Free tier available in Google AI Studio • Yes probabilities max out at 0.99 • No prompt caching or batch discounts Here's where it gets interesting: You can build an AI workflow where Gemini Decisions handles thousands of routine decisions, automatically processes high-confidence results, and sends only uncertain cases to a more powerful model or a human Not every AI task needs a model that writes an answer Sometimes, all you need is a fast, cheap decision
1
1
209
🚨 Claude Pro's free Fable 5.2 access ends Monday at 11:59 PM PT After that, using Fable on Pro will cost extra, and the charges can add up quickly Here's what changes: 💰 Claude Pro ($20/month) Fable 5.2 will only be available through paid usage credits, billed at API rates: • Input: $8 / 1M tokens • Output: $40 / 1M tokens ⚡ Claude Max ($100/month) • Fable 5.2 is included in your subscription • Usage is capped at up to 50% of your weekly limit • Beyond that cap, you'll also need paid usage credits Which plan makes more sense? Max costs $80 more per month than Pro If you're spending over $80/month on Fable credits, Max could be the better deal, provided your usage fits within its included limits ⚠️ Here's the catch: Max also comes with $100/month in API credits, but those credits only work on the Claude Platform They cannot cover additional Fable usage in your Max subscription Staying on Pro? Go to Settings → Usage and set a monthly spending cap before enabling usage credits The free access ends Monday, November 16 Don't let your next Claude session turn into an unexpected bill
1
1
2
362
DeepSeek just made AI agent web searches 13× cheaper than Claude and OpenAI 🤯 100,000 web searches could cost you $1,000 with Claude or OpenAI, but just $75 with DeepSeek DeepSeek's new Search API offers a massive price difference: 💰 Cost per 1,000 searches • Claude API: $10 • OpenAI API: $10 • OpenAI preview tool: $25 on non-reasoning models • DeepSeek Search API: $0.75 And DeepSeek doesn't charge for empty results, failed requests, or timeouts Here's what makes it interesting for AI agents: 1. Keep your existing AI model The API returns structured JSON, so Claude or GPT can continue handling reasoning while DeepSeek handles web searches No need to replace your main model 2. Get up to 50 results per search You can request up to 50 results in a single API call without paying extra per result 3. Filter results before sending them to your model Each result includes a rerank_score DeepSeek considers scores of 1.5 or higher to indicate high-quality results, so you can filter out anything below that threshold before passing the results to Claude or GPT The cost difference at scale: 100,000 searches/month: • Claude / OpenAI: $1,000 • DeepSeek: $75 That's $925 saved every month on search requests alone You don't always need a cheaper AI model to cut agent costs Sometimes, you just need a cheaper tool behind it
🚀 The full DeepSeek App experience, now via API Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API. Enable built-in features with a few parameters, or call standalone APIs such as OCR, ASR, and TTS. Build richer conversations or recreate the full DeepSeek App in your own application. Try our open-source frontend demo, with support for importing DeepSeek conversation history. 🔗 Demo: cdn.deepseek.com/chat-demo 📦 Code: github.com/deepseek-ai/chat-… 📖 Details: api-docs.deepseek.com/guides… 1/4
3
3
318
Claude Code has a hidden experimental feature that can cut token usage by up to 37% 🤯 It's called Code Mode, and you can enable it with a single environment variable Instead of handling every tool call separately, Claude can write a JavaScript script to orchestrate multiple operations, keeping unnecessary tool output out of your context The results are impressive: In Anthropic's API tests on complex research tasks: • Standard tool calling: 43,588 tokens • Code Mode: 27,297 tokens • 37% fewer tokens consumed 🚀 How to enable it in Claude Code 1. Update Claude Code claude update Requires version 2.1.323 or later 2. Launch with Code Mode enabled CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude 3. Give Claude a multi-step task For example: • Refactor code across multiple files • Search and filter results from a large codebase • Run several independent commands in parallel • Coordinate related tool operations 4. See what's happening Press Ctrl+O to inspect the JavaScript Claude generates ⚙️ A few important details • Add the environment variable to ~/.claude/settings.json to enable it permanently • Project-level settings won't enable this feature • Resume sessions with the variable still enabled • Currently works only in the terminal, not the VS Code panel or -p mode The 37% reduction comes from Anthropic's API research, not a separate Claude Code benchmark Less tool output filling your context, more room for actual reasoning and coding
2
1
4
316
Claude Code's max effort costs 15× more than medium, yet performs worse than high 🤯 Artificial Analysis tested Claude Fable 5.2 in Claude Code across four effort levels 📊 Benchmark Results Cost per task | Score • Medium: $2.89 | 53 • High: $7.51 | 58 • XHigh: $15.47 | 58 • Max: $44.22 | 56 The surprising part: • High scores 5 points above Medium at 2.6× the cost • XHigh costs more than twice as much as High, with no score improvement • Max costs over 15× as much as Medium and scores lower than High 💡 How to optimize Claude Code Both Fable 5.2 and Opus 5.5 default to Medium effort When Claude struggles with a difficult task, try increasing effort before switching to a more expensive model Useful commands: • /effort high → Increase reasoning effort • /effort → Press s to apply the change to the current session only Google's Addy Osmani recommends the same approach: increase effort first, switch models second Garry Tan's 52-task test also found that Fable 5.2 cost 2.8× more per solved task than Opus 5.5, despite achieving a similar success rate More thinking doesn't always mean better results High delivered the best score in this benchmark, while Max mostly delivered a much bigger bill
3
1
3
238
Claude Code can cost 2× more than a free, open-source coding agent running the exact same Claude model And the difference in coding performance is surprisingly small A study from Arena, involving Ion Stoica and Matei Zaharia, compared Claude Code, Codex CLI, and Pi using identical tasks, models, and API pricing Here's what happened with Claude Fable 5: Claude Code • Task success rate: 97.8% • Cost per task: $1.33 Pi (open-source) • Task success rate: 96.7% • Cost per task: $0.67 Almost the same results, but Claude Code costs nearly twice as much Across models on SWE-bench Lite, Claude Code cost roughly 2× as much as Pi, while average success rates differed by less than 2 percentage points Why such a big cost difference? Part of the answer is context overhead Before doing any actual work, Claude Code sends over 10× more initial context than Pi, including longer system instructions and larger tool definitions That overhead consumes tokens before the agent even starts solving your problem And there's another interesting comparison The recently launched Reason agent advertises savings of "up to 20×" But in its own benchmark, comparing the same GPT model against Codex, the actual saving was 22% per successfully solved task Want to test this yourself? Pi is free and open-source, and you can run it in your own repository with your preferred model and API key Installation instructions in the comments 👇 Before switching to a cheaper model, try switching the agent running it The agent itself might be costing you more than you think
1
1
2
311
Try Pi, the open-source coding agent Install: npm install -g --ignore-scripts @earendil-works/pi-coding-agent Run it in your repository: pi Then configure your provider and model: /login — Authenticate with your provider /model — Select the model you want to use Compare the token usage and task costs against Claude Code using the same model
1
64
Google Antigravity can now turn a single coding task into a team of AI agents working in parallel And all it takes is one command: "/run-workflow" For example, paste this into any Antigravity chat: /run-workflow List the Python files changed since main, review each one in parallel for unhandled exceptions using the flash model, and give me a one-line verdict per file Here's what happens: 1. Antigravity generates a short Python script for the workflow 2. Shows you the script before execution 3. Runs it in the background, distributing tasks across multiple agents 4. Each agent works in its own chat, keeping file reads and test logs out of your main conversation 5. You get a concise result for each file A few tips to get the most out of it: • Use cheaper models — "flash" or "flash_lite" for routine checks, "pro" for complex reasoning • Isolate parallel edits — Ask for a branch workspace when multiple agents modify code, giving each its own Git copy • Keep batches small — If one item fails, the remaining parallel tasks get canceled • Reuse successful workflows — Save the generated script as a ".py" file and rerun it instead of generating it again The feature works with Google AI Pro and Ultra, through both the Antigravity app and CLI The best part? You're not just asking AI to do more work You're turning a natural-language request into a reusable, inspectable workflow that coordinates multiple agents A script you can inspect and rerun is far more useful than a long conversation that slowly loses direction
2
1
224
Google is doubling Gemini 4 Carbon and Argon's API prices next month, but its new Gemini 4 Hydrogen model might be the cheaper alternative 👀 Here's the price change per 1M tokens: • Carbon / Argon input: $2 → $4 • Carbon / Argon output: $10 → $20 Meanwhile, Gemini 4 Hydrogen stays at $1.50 input / $7.50 output It's also slightly ahead of Argon on the Artificial Analysis Intelligence Index, scoring 51 vs 50 Sounds like an easy switch, right? Not so fast — Hydrogen uses significantly more output tokens According to Arena's measurements, Hydrogen generates: • 41% more output tokens per task than Carbon • 2.8× as many output tokens as Argon And reasoning tokens are billed as output tokens So what happens to your actual costs after the price increase? • Carbon → Hydrogen: Output costs per task drop by roughly 50% • Argon → Hydrogen: Output costs per task stay about the same The biggest opportunity is input-heavy workloads Hydrogen charges just $1.50 per million input tokens versus $4 for Carbon and Argon, making it especially attractive for long-context tasks and document processing For Argon-based coding agents: Try Hydrogen with thinking_level: "medium" first, then compare the total billed output tokens per completed task A cheaper model doesn't always mean a cheaper task Compare cost per completed task, not just the API price sheet
7
1
8
2,125
Cloudflare just open-sourced a browser built specifically for AI agents, using up to 7× less memory than Chromium 🚀 Meet Kitesurf, a lightweight, stateless browser engine that runs on Cloudflare Workers without launching Chrome or managing browser containers Unlike traditional browser automation, Kitesurf is designed around what AI agents actually need What can it do? • Load web pages and parse HTML and CSS • Execute JavaScript and interact with the DOM • Capture screenshots and generate PDFs • Connect to Puppeteer and other CDP-compatible tools • Scale browser workloads across Cloudflare's global Workers network The performance numbers are impressive: According to Cloudflare's benchmarks: • 3× lower CPU time for screenshots • 3.8× lower CPU time for HTML extraction • 4.7× lower memory usage for screenshots • 6.9× lower memory usage for HTML extraction Built with Rust and WebAssembly, it combines a custom browser engine with existing web technologies AI agents need browsers, but do they really need Chrome? Most AI agents just need to read web pages, execute some JavaScript, or extract information, not run an entire Chrome browser Kitesurf provides those capabilities without the overhead of a full Chromium instance The project is open-source under Apache 2.0, supports local development, and includes a live playground GitHub and demo links in the comments 👇
2
1
198
Claude Code flagged 5,980 issues across 200 real pull requests Only 435 were actual bugs 🤯 Alibaba just open-sourced Open Code Review, an AI-powered code review tool it has used internally for two years It's currently #1 on GitHub Trending And its benchmark results are pretty interesting Both tools were tested using the same Claude Opus 4.6 model: Claude Code: • 7% of reported issues were real bugs • 5.7M tokens consumed • 13 minutes per review Open Code Review: • 34% of reported issues were real bugs • 385K tokens consumed • Under 1.5 minutes per review That's nearly 5× better precision, 15× fewer tokens, and almost 9× faster reviews There's a trade-off, though Claude Code found 29% of the known bugs, while Open Code Review found 20% So you sacrifice some coverage, but get far fewer false positives and much faster reviews Getting started is simple: Install the CLI, configure your model provider, and run "ocr review" in your repository Don't have a separate API key? There's also a Claude Code plugin that lets you run reviews using your existing Claude Code plan The goal isn't to flag every possible issue It's to find the bugs that actually matter, without burying developers in noise Installation instructions in the comments 👇
10
1
10
453
Alibaba Open Code Review Install the CLI: npm install -g @alibaba-group/open-code-review Configure your provider and model: ocr config provider ocr config model Run a review: ocr review Using Claude Code without a separate API key? Install the Open Code Review Claude Code plugin, then run: "/open-code-review:delegate-review" This lets you run reviews through your existing Claude Code plan
89
The best coding agent workflow might start with the AI refusing to write code until it understands exactly what you want Matt Pocock has open-sourced his personal collection of coding agent skills, and the GitHub repo has already passed 300,000 stars 🔥 They're built for real software engineering, not just vibe coding His #1 recommendation: Start with a grilling session Before making changes, the agent interviews you, challenges your assumptions, and asks detailed questions until you're both aligned on what needs to be built The collection also includes: • "/grill-with-docs" — Clarifies requirements and documents shared terminology and design decisions • "/tdd" — Uses a test-first, red-green-refactor workflow • "/diagnosing-bugs" — Systematically investigates bugs before attempting fixes • "/improve-codebase-architecture" — Scans your codebase for opportunities to simplify and improve its design What makes these skills different? Unlike frameworks such as Spec-Kit that try to manage your entire development process, Matt's skills are small, composable, model-agnostic, and easy to customize You can use only the pieces you need without handing over control of your workflow They work with Claude Code, Codex, and other coding agents Most agent tools focus on making AI write code faster These focus on making sure it builds the right thing in the first place GitHub and installation instructions in the comments 👇
23
16
606
Matt Pocock's Skills for Real Engineers GitHub: github.com/mattpocock/skills Claude Code installation: claude plugin install mattpocock-skills@claude-plugins-official Then run once in each repository: "/setup-matt-pocock-skills" To start a requirements interview: "/grill-with-docs" For Codex and other agents: npx skills@latest add mattpocock/skills Select the skills you want, including "setup-matt-pocock-skills", then run the setup command in your agent
96
Claude Code has a hidden trick to avoid paying 4× more for Fable 5.2 👀 You don't need Claude Fable 5.2 running on every Claude Code turn It's 4× more expensive than Sonnet 5.5 per token • Fable 5.2: $8 input / $40 output per 1M tokens • Sonnet 5.5: $2 input / $10 output But Claude Code has a hidden flag that lets you use both models: "claude --model sonnet --advisor fable" Here's how it works: Sonnet handles the actual coding, while Fable acts as an advisor and only steps in when: • Claude needs to choose an approach • It gets stuck repeating the same error • It's about to declare a task complete And the results are interesting: In Anthropic's tests, Sonnet with an Opus advisor scored 2.7 points higher on SWE-bench Multilingual, while costing 11.9% less per task than Sonnet alone There's one catch: The advisor rereads the entire conversation without prompt caching, so a single consultation at 100K tokens costs around $0.80 in Fable input alone You can pause it anytime with "/advisor off" without breaking your main prompt cache, unlike switching models with "/model" Fable for the tough decisions, Sonnet for everything else
107