Lucas Souza
{AI Engineer} — apaixonado por Laravel, arquitetura de software e construir produtos com impacto. Compartilho aqui tutoriais, descobertas e reflexões sobre o dia a dia de engenharia.
GPT-6.2 Astra launched at $8/$40, the exact price of Claude Fable 5.2, and tied it at 174 on the Epoch Capabilities Index. But cache reads cost twice as much and a surcharge kicks in above 272K tokens. The math for a real agent session, what is lab data and what is independent, and why there was never a 6.1 Astra.
Anthropic launched Claude Meetings in beta: Claude joins your Google Meet call, takes notes and builds docs and whiteboards live. How to turn it on, what gets stored, what it costs and the gotchas.
Gemini 4 Hydrogen costs 2.7x less than Claude Opus 5.5 per token and has a free tier. But where your coding agent lives, Opus still holds. Benchmarks side by side (Google vs Anthropic), the math on a cached agentic session, which Gemini 4 to use after Carbon and Argon double in price on 12/1, and a Laravel fallback with the list of what breaks when you switch.
Anthropic measured what actually works in AI agent memory: four layers, telling the agent what memory is for instead of rules, and dreaming, an offline step where the agent rereads transcripts and fixes its own memory. CL-Bench, Pokémon and wiki vs search numbers, plus Dreams API code and how to apply it in Claude Code.
Claude Fable 5.2 launched today 20% cheaper than 5.1, but it still costs twice as much as Opus 5.5 per token. Official benchmarks against Opus 5.5, Fable 5.1 and GPT-6 Astra, the cost of an agentic session with caching, what Mythos 5.2 is, and a decision tree for when each model is worth it.
55% of the improvements an agent proposed for its own harness fixed one failure and broke something that worked. A PHP tutorial for a self-improving harness with the piece that makes it engineering: the regression gate. Git-versioned rules, failure mining, minimal rules, protected cases, provisional promotion and pruning.
Across 54 cycles, the agent said it improved every time. In 56%, nothing changed or it got worse. The progress mirage measured in a paper, why a bigger LLM judge doesn't fix it, and how to build a Laravel loop with an external verification gate: checkable exit, hard budget, and a guard that stops the agent from editing its own exam.
An agent that runs across several context windows forgets everything each session. The pattern Anthropic, Manus and LangChain converged on: move state out of the window and into files. A JSON feature list that only accepts status changes, a progress log, a session start ritual and offloading large results with a restorable pointer, all with PHP code.
A subagent isn't a job title, it's a context window. When to use subagents (and when one agent is enough), the difference between isolated and fork modes, the numbers from Anthropic and Cognition, and how to implement both in a PHP harness straight on the Claude API, no framework.
Anthropic confirmed two sentences in Claude Code's prompt that stop it from recreating existing games. What AI has already done with game reverse engineering, the damage to the market, and why the brakes came now.
GPT-6.1 Terra costs a quarter of Opus 5.5 per token, and OpenAI says it comes out 9x cheaper per task. We ran the numbers per solved task, separated lab numbers from independent data, and show which agent the swap makes sense for (and which it doesn't).
Max 20x became Claude Ultra and lost the 5-hour limit, but only on verified accounts. The weekly cap is the same: 10x Pro. What changed, why "20x" vanished from the name, and how to use /usage to find out which limit is actually hitting you.