Fast move on the workhorse tier. Matters if you build anything agentic 👇 The news: 3.7 Flash launched just three weeks after 3.6, at half the token cost — $0.75/M input, $3.75/M output, locked through Dec 31. The gains are real: → DeepSWE coding: 49% → 65.3% → Reportedly beats Claude Sonnet 5 and GPT-5.6 on coding → 1M context, better multi-step planning and tool calls The bigger move: Gemini Spark — Google's 24/7 personal agent for Pro/Ultra — now runs on 3.7 Flash. The agent working between your sessions just got a real bump, especially for Workspace tasks. My take: Google isn't chasing the frontier crown right now — their flagship 3.5 Pro is still delayed. They're dominating the workhorse tier instead: the cheap, fast, good-enough model you actually run at scale. Smart. Most agentic workflows don't need the smartest model — they need a reliable one cheap enough to run thousands of times. For builders: The intro pricing is a defined window to test cost-per-task before rates double in January. Use it. And this is the price war I keep flagging, now three-way — Claude doubled Cowork limits, OpenAI gamified resets, Google halved Flash pricing. Every lab's fighting for the builder. You win. Stay model-agnostic and ride whoever's cheapest for the job. Caveat: still no EU availability. Check your region first. Running Gemini Flash in any workflows yet, or all-in on Claude/OpenAI? 👇

I built a tool which allows you to save those sweet Fable tokens 😩 When you're deep in a Claude session and the context window fills up, /compact summarises your whole conversation using that same expensive model. I built a solution for that - which most of you will hate 🤣 /side-compact lets you use a cheaper model, on a lower effort level, while compacting a conversation or creating a handover doc. It runs Sonnet or Haiku on your transcript, writes a chaptered summary, and gives you a command to start a fresh session that picks up where you left off. What it does: 🔹 Runs inside your live session, accessible by your agent 🔹 Strips noise like tool-call payloads and thinking blocks before summarising 🔹 Outputs chapters with line references back to the original transcript, so a fresh session can look up details on demand 🕵️ The summarisation call is sandboxed with no tools, which makes it more secure than an LLM provider's sandbox because it won't go hack a country on your behalf Check it out, it's open source and free - https://github.com/akash-joshi/side-compact

This week I noticed a repeating pattern that happens with most of the users Claude buries simple answers in jargon it never explains and it turns a one-line question into a five-paragraph essay you have to mine for the actual answer I hit both walls myself. Asked a quick question, got a term I'd never seen, no definition, then three paragraphs of context I didn't ask for Burned tokens, burned time So I built a skill around one rule: explain it like you're talking to a 15 year old. No assumed expertise, no dense wording, no padding Under the hood it leans on ASD-STE100 — the simplified English standard manuals use One idea per sentence. Common words only. No hidden acronyms It's built for people who can't afford to misread a sentence, which is exactly the bar an AI answer should clear to Since I started using it, the difference isn't subtle. Same questions, half the words, zero re-reading Turns out most "complexity" in an answer isn't the topic being hard. It's the explanation being lazy Anyone else keeping a running list of terms Opus 5 uses without ever defining them?

ChatGPT Computer History is a new, controversial feature for the desktop app that acts like a digital observer, allowing the AI to see and remember the activities you perform across your apps and websites. Key features and implications include: • Workflow Automation: The AI analyzes your tasks to suggest ways to automate repetitive work, such as turning notes into project plans or streamlining your daily processes. • Contextual Awareness: By tracking the sequence of your actions-such as browsing, writing, or moving files-the AI gains a deep understanding of your work habits, providing context that standard agents cannot access. • Customization: You maintain control by deciding which applications the AI can access, though you must be proactive about setting privacy boundaries. • Privacy Concerns: Because the feature monitors your active work and the content within your documents, it has drawn comparisons to "Big Brother." Users are encouraged to pause the feature during sensitive tasks or private communications. Currently, the feature has a limited rollout: it is available only on Mac for Pro, Business, or Enterprise users, and it is excluded from the UK, Switzerland, and the European Union.

Some people think adding more steps makes something more valuable. I think the opposite. If two automated agents solve the same problem equally well, I’ll choose the one that is simpler and easier to use every time. Customers don’t pay for complexity. They pay for outcomes -and for making their lives easier. The best automation often feels almost invisible: fewer steps, less friction, less effort. And this applies to almost everything -software, services, processes, and even content. Complexity may demonstrate how much work went into something. Simplicity demonstrates how well you understood the problem. That has been my goal with ClickFlow from day one. What part of your business or offer could you simplify today?