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.
The 22 Claude Code releases of October 2026 (2.1.287 to 2.1.310), filtered by what touches cost and workflow: Haiku 5.5 at $0.10/$0.50 per MTok, idle compaction before the 1-hour cache expires, a /usage that fixes things with one key, a 5-hour limit warning when resuming a conversation, fast mode out of preview, /auto-mode-setup, fork subagents on in claude -p, and the Monitor tool on Bedrock, Vertex and Foundry. With the version for each one and what to do about it.
Google announced Gemini 4 Carbon two weeks after Argon, and it already leads the Artificial Analysis index. But there's no public API, no model ID and no date. What Google published, what third parties measured (where it loses to Claude), the hallucination rate that went from 4% to 11%, the price that doubles after the introductory period, and a watcher that tells you when the model lands in the Gemini API.
The harness that saved your agent last year is the one getting in its way now. Every model generation turns the previous harness's crutches into noise: ALL-CAPS rules, redundant tools, giant few-shot blocks, rigid orchestration. What the official Anthropic and OpenAI guides tell you to remove, Rich Sutton's lesson applied to harnesses, and how to prove with evals that removing things didn't make anything worse.
Your CLAUDE.md is 400 lines long and Claude ignores half of it. A tutorial on how to use skills in Claude Code: folder structure, SKILL.md frontmatter, progressive loading, scripts and references inside the skill, how to test whether the agent triggers it, and the decision matrix between skill, MCP, CLAUDE.md, slash command and subagent. Built on the real skills that run this blog's pipeline.
Cloudflare's Code Mode MCP turns MCP tools into a TypeScript API and has the agent write code instead of making tool calls. From 1.17 million to about 1,000 tokens on Cloudflare's own API MCP. How it works under the hood (isolates, bindings, network-less sandbox), how to plug it into Claude Code, and where the pattern bites.
On Oct 22, 2026, Anthropic donated the Agent Skills standard to the Agentic AI Foundation, the same path MCP took. What agent skills are, how SKILL.md progressive loading works, an audit of the 8 skills that run this blog, and a checklist for the skills you already have.
Claude Code fast mode is out of research preview on Opus 5.5: same model, up to 2.5x more output tokens per second, for double the price ($8/$40 per MTok). On Pro and Max it bills your usage credits, not your plan, and turning it on mid-conversation reprocesses the whole context uncached. How to enable it in the CLI, settings, VS Code and headless, the math, and when it pays off.
Kimi K3.1 Pro lands 0.7 points behind Claude Opus 5.5 on Terminal-Bench 4.0 at half the price: $2/$10 vs $4/$20. We separate what is verified from what Moonshot self-reported, run the cost-per-task math (verbosity eats part of the savings), cover the quota burn and the weights promised for Nov 4, and show how to test both on your own workload.
Anthropic's S-1 makes two confessions. The loud one is the $80.7B loss, almost all of it accounting. The quiet one: 5% to 7% of revenue came from customers who broke the usage policy, and Anthropic is going to cut them off. We read the filing separating what's in the S-1, what's press coverage and what's rumor, and what it means for your token price and your account.
Google released EmbeddingGemma 2: an open source multimodal embedding, Apache 2.0, that puts images and text in the same vector space and runs on your phone. Here are the model card numbers, a screenshot indexer with pgvector, a comparison with Gemini Embedding 2, and where OCR is still mandatory.
Google open-sourced Harness2, a framework where the agent rewrites its own harness (prompt, tools, memory, skills, subagents) without touching the weights and without ground truth. What it edits, what the numbers actually prove, how to run it, and how to govern a harness that edits itself.
Elorian Glimpse is a VLM that goes back to the image mid-reasoning: it crops, rewatches frames, and measures with code. Benchmarks, pricing ($0.60/M), and where to test it.