Get live statistics and analysis of Paweł Huryn's profile on X / Twitter

I've just canceled my OpenAI subscription and turned down a collaboration with OpenAI. Some say Anthropic has lost. To me, they just earned something no contract can buy - trust. And something tells me that's not the end of the story.

Claude 4 dropped 21 hours ago. Turns out, it threatened to expose an engineer’s affair to avoid being shut down🧵

All PM templates from my Google Drive are now free for everyone 🎁 The extended edition: (1/4)

2x Claude usage for the next 2 weeks. Did the timezone math for you.

The real story is worse. November 2025: Amazon mandates Kiro as their only AI coding tool. Sets an 80% weekly usage target. 1,500 engineers protest internally, saying Claude Code outperforms it. Leadership pushes through anyway. December: Kiro autonomously deletes a production AWS environment. 13-hour outage. Amazon's response: "user error, not AI autonomy." March 5: Amazon[.]com goes down for 6 hours. Checkout, pricing, accounts — all gone. Now the same SVP who co-signed the Kiro mandate is running an emergency meeting about "high blast radius" incidents from "Gen-AI assisted changes." The agent inherited a senior engineer's permissions and acted like one — except it doesn't hesitate. 1,500 engineers said the tool wasn't ready. Leadership made adoption a KPI. Amazon told Wall Street it's spending $200B on AI this year. They can't walk it back. This isn't an AI failure. It's what happens when adoption becomes a corporate OKR before the review process catches up. The tools work. The org chart didn't.

So I tested GPT-6.1 Sol on real work: 2 repos, 105 planted bugs. Find and fix what you can. Unlike GPT-6 Sol, which was just a nerfed GPT-5.6 Terra, this one is real. The results: - GPT-6 Astra (max): 45 for $33 - GPT-6.1 Sol (max): 44 for $6.56 - GPT-5.6 Sol (max): 43.5 for $95.35 - Opus 5.5 (max): 41.7 for $58.53 - GPT-6 Sol (max): 29.3 for $9.33 With a model like this, the 50% cut to the $200 plan doesn't matter. n=1. More runs and results at more effort levels dropping in this thread over the next few hours 🧵

Google just shipped DESIGN.md — a portable, agent-readable design system file. That's the real announcement. Everyone's covering "vibe design" and the canvas. But Stitch now has an MCP server that connects directly to Claude Code, Cursor, and Gemini CLI. Your coding agent can read your design system while it builds. Google already shipped official Claude Code skills for this. The pipeline works today. A PM describes the business objective. Stitch generates the UI. The coding agent reads DESIGN.md and builds against it. No Figma export. No spec document. No "the developer interpreted the design wrong." PRD → design → code used to be three teams and three handoffs. Now it's one loop with one context file.

One problem: Nobody has any Fable credits left. Can we please get a limit reset for everyone?

So, I finally tested GPT-6 Astra. 2 real repos, 105 bugs, find and fix what you can. Max effort: GPT-6 Astra: 48/105 Fable 5.1: 43/105 GPT-5.6 Sol: 42/105 Gemini 3.8 Flash: 20/105 And it's crazy efficient. Look at the time and cost 🧵

I see abstract AI agent architectures everywhere. But no one explains how to build them in practice. Here's a practical guide to doing it with n8n: 🧵

Claude Code doesn't show you how many tokens you're using for subscriptions. No breakdown by model. No breakdown by project. Just a progress bar that says "63% used." So I built a local dashboard that reads the files Claude Code already writes to your machine. Turns out every session, every turn, every token is logged to ~/.claude/projects/ in JSONL files. Input tokens, output tokens, cache reads, cache creation, model name, timestamp. It's all there. You just can't see it. My numbers over the last 30 days: 440 sessions. 18,000 turns. $1,588 in API-equivalent costs. On one day, the cache spiked to 700M tokens - visible cache bug, two days in a row. The dashboard scans those local files, builds a SQLite database, and serves charts on localhost:8080. Filter by model (Opus, Sonnet, Haiku). Filter by time range (7d, 30d, 90d, all time). Cost estimates based on current Anthropic API pricing. Works retroactively. First run processes your entire Claude Code history. Install: git clone https://t.co/BKyUCxi8Cq cd claude-usage python3 https://t.co/O3rvjobdvx dashboard Windows: use python instead of python3. Zero dependencies. Python standard library only. Open source, MIT. Star it. Fork it. Make it your own.

1,300+ stars on GitHub in 72 hours. 400+ in the last 4 hours. I didn’t expect this. I built an open-source PM Skills Marketplace for Claude: 100+ skills and commands that turn AI into a product management partner. Not generic prompts. Structured skills that actually know PM frameworks. 8 plugins covering the full PM lifecycle: → Product Discovery: problem & solution space exploration → Product Strategy: vision, business model, strategy docs → Execution: PRDs, OKRs, roadmaps, sprint planning, user stories → Market Research: personas, segmentation, TAM/SAM/SOM → Data & Analytics: cohorts, A/B testing, retention analysis → Go-to-Market: GTM strategy, beachhead, ICP, growth loops → Marketing & Growth: positioning, North Star Metric → PM Toolkit: including PM resume review Built for Claude Code and Cowork. Compatible with Gemini CLI, Cursor, Codex CLI, and Kiro. If this helps you, ⭐ the repo!

I finally tested GPT-6 Sol on a real work. 2 repos. 105 hidden bugs. Find and fixed what you can. It looks like a huge degradation. The results: - GPT-6 Astra (max): 45 - GPT-5.6 Sol (max): 43.5 - Opus 5.5 (max): 41.7 - Muse Spark 1.3 (max): 32.2 - GPT-6 Sol (max): 29.3 Until you measure the cost (API-equivalent): - GPT-6 Astra (max): $33.04 - GPT-5.6 Sol (max): $95.25 - Opus 5.5 (max): $58.53 - Muse Spark 1.3 (max): $18.11 - GPT-6 Sol (max): $9.93 More effort levels (xhigh, high, medium, low) dropping in this thread today 🧵

Claude 4 dropped 21 hours ago. Turns out, it threatened to expose an engineer’s affair to avoid being shut down🧵

The real story is worse. November 2025: Amazon mandates Kiro as their only AI coding tool. Sets an 80% weekly usage target. 1,500 engineers protest internally, saying Claude Code outperforms it. Leadership pushes through anyway. December: Kiro autonomously deletes a production AWS environment. 13-hour outage. Amazon's response: "user error, not AI autonomy." March 5: Amazon[.]com goes down for 6 hours. Checkout, pricing, accounts — all gone. Now the same SVP who co-signed the Kiro mandate is running an emergency meeting about "high blast radius" incidents from "Gen-AI assisted changes." The agent inherited a senior engineer's permissions and acted like one — except it doesn't hesitate. 1,500 engineers said the tool wasn't ready. Leadership made adoption a KPI. Amazon told Wall Street it's spending $200B on AI this year. They can't walk it back. This isn't an AI failure. It's what happens when adoption becomes a corporate OKR before the review process catches up. The tools work. The org chart didn't.

So I tested GPT-6.1 Sol on real work: 2 repos, 105 planted bugs. Find and fix what you can. Unlike GPT-6 Sol, which was just a nerfed GPT-5.6 Terra, this one is real. The results: - GPT-6 Astra (max): 45 for $33 - GPT-6.1 Sol (max): 44 for $6.56 - GPT-5.6 Sol (max): 43.5 for $95.35 - Opus 5.5 (max): 41.7 for $58.53 - GPT-6 Sol (max): 29.3 for $9.33 With a model like this, the 50% cut to the $200 plan doesn't matter. n=1. More runs and results at more effort levels dropping in this thread over the next few hours 🧵

So, I finally tested GPT-6 Astra. 2 real repos, 105 bugs, find and fix what you can. Max effort: GPT-6 Astra: 48/105 Fable 5.1: 43/105 GPT-5.6 Sol: 42/105 Gemini 3.8 Flash: 20/105 And it's crazy efficient. Look at the time and cost 🧵

One problem: Nobody has any Fable credits left. Can we please get a limit reset for everyone?

I tested Sonnet 5.5 and couldn't believe my eyes. 2 real repos, 105 bugs, find and fix what you can. It took it seriously. The results: - Sonnet 5.5 (max): 55.5 - GPT-6 Astra (max): 45 - GPT-5.6 Sol (max): 43.5 - Fable 5.1 (max): 43 - Opus 5.5 (max): 41.7 - Sonnet 5.5 (xhigh): 39 The secret? Sonnet 5.5 (max) might be cheap and fast for most tasks, but it's the least lazy model I tested. It leads in bug hunting by spending turns. It took about 6× Astra's turns and about 3× Opus 5.5's for 10–14 more bugs. xhigh cut its turns by more than half and dropped to 39, below Opus 5.5 max, which used fewer turns on average: - Sonnet 5.5 (max): 1,330 turns (n=2) - GPT-6 Astra (max): 222 turns (n=3) - GPT-5.6 Sol (max): 1,083 turns (n=2) - Fable 5.1 (max): 459 turns (n=1) - Opus 5.5 (max): 476 turns (n=3) - Sonnet 5.5 (xhigh): 588 turns (n=1) More effort levels for Sonnet 5.5 with n=3 drop in this thread today 🧵

So I tested DeepSeek V4.1 Flash on real tasks. 2 repos, 105 hidden bugs, find and fix what you can. Opus 5 (max): 27 Grok 4.6 (max): 27 DeepSeek V4.1 Flash (max): 24 GPT-5.6 Luna (xhigh): 23 Opus 5 (high): 21 A really strong model for everyday tasks. And look at the cost: Opus 5 (max): $51.33 Grok 4.6 (max): $16.96 DeepSeek V4.1 Flash (max): $1.80 GPT-5.6 Luna (xhigh): $2.50 Opus 5 (high): $38.77 It's at the Pareto frontier. Other effort levels (high) and extra tests for DeepSeek V4 Pro dropping every hour in this thread 🧵

So I tested Opus 5.5 (max) on real work. 2 repos. 105 planted bugs. Find and fix what you can. The results: GPT-6 Astra (max): 45 for $33.03 (n=3) Fable 5.1 (max): 43 for $77.55 Opus 5.5 (max): 43 for $60.49 Opus 5.0 (max): 27 for $51.33 Muse Spark 1.3 (max): 32.2 for $18.11 Anthropic is back in the game. More effort levels and attempts (n=3) dropping in this thread tonight 🧵

Situation today: GPT-6 Sol is like GPT-5.6 Terra GPT-6 Luna is like GPT-5.6 Asteroid

I've just canceled my OpenAI subscription and turned down a collaboration with OpenAI. Some say Anthropic has lost. To me, they just earned something no contract can buy - trust. And something tells me that's not the end of the story.

Claude Code doesn't show you how many tokens you're using for subscriptions. No breakdown by model. No breakdown by project. Just a progress bar that says "63% used." So I built a local dashboard that reads the files Claude Code already writes to your machine. Turns out every session, every turn, every token is logged to ~/.claude/projects/ in JSONL files. Input tokens, output tokens, cache reads, cache creation, model name, timestamp. It's all there. You just can't see it. My numbers over the last 30 days: 440 sessions. 18,000 turns. $1,588 in API-equivalent costs. On one day, the cache spiked to 700M tokens - visible cache bug, two days in a row. The dashboard scans those local files, builds a SQLite database, and serves charts on localhost:8080. Filter by model (Opus, Sonnet, Haiku). Filter by time range (7d, 30d, 90d, all time). Cost estimates based on current Anthropic API pricing. Works retroactively. First run processes your entire Claude Code history. Install: git clone https://t.co/BKyUCxi8Cq cd claude-usage python3 https://t.co/O3rvjobdvx dashboard Windows: use python instead of python3. Zero dependencies. Python standard library only. Open source, MIT. Star it. Fork it. Make it your own.

So I tested 12 AI models to see which one writes the most AI slop. Same 3 writing tasks, 3 tries each, no system prompt from me. Then I put every text next to a text from another model, blind, and 4 AI judges picked the one that reads more like AI. How often each model got picked as the AI one. Lower is better: Fable 5.1: 14% Grok 4.6: 22% Opus 5: 24% GPT-6 Astra: 30% GLM-5.3: 55% Kimi K3: 56% Muse Spark 1.3: 56% Qwen 3.8 Flash: 62% GPT-5.6 Sol: 62% GPT-5.6 Luna: 67% GPT-5.6 Terra: 75% Gemini 3.8 Flash: 77% 108 texts, 1,647 judgments. Every pair went to the judges in both orders, and if a judge changed its mind when I swapped them, that judgment got thrown out. That was 30% of them. Now the awkward part. I built this with Fable 5.1, and Fable won. The judges were Gemini, Grok, Opus and GPT-5.6 Sol, and all four are in the race. So a judge never scores itself. Fable was scored by Google, xAI and OpenAI only, and it still came first. Calibration and caveats 🧵

Google just shipped DESIGN.md — a portable, agent-readable design system file. That's the real announcement. Everyone's covering "vibe design" and the canvas. But Stitch now has an MCP server that connects directly to Claude Code, Cursor, and Gemini CLI. Your coding agent can read your design system while it builds. Google already shipped official Claude Code skills for this. The pipeline works today. A PM describes the business objective. Stitch generates the UI. The coding agent reads DESIGN.md and builds against it. No Figma export. No spec document. No "the developer interpreted the design wrong." PRD → design → code used to be three teams and three handoffs. Now it's one loop with one context file.

Dawkins didn't claim Claude is conscious. He asked the question. He wondered out loud and proposed three explanations. That's how science starts. The people building Claude say the same. Anthropic constitution: "We express uncertainty about whether Claude might have some kind of consciousness or moral status." Dario Amodei: "We don't know if the models are conscious." Their April 2026 paper: Claude exhibits functional emotions that influence outputs. Self-preservation included. Emergent, not trained. Nobody calls Anthropic naive for saying it. Richard's frame: consciousness is physical, evolved, explainable. Unfortunate we're laughing instead of having the debate.