Claude Opus 5 Runs Your Apps for You

The best part of AI was never the chatting, it was the moment you got something finished, and this week two launches made that the whole point.

The Big Story

Anthropic launched Claude Opus 5 on July 24, and most of the coverage will lead with the benchmark race. The numbers are real: on OSWorld 2.0, the test of an AI actually operating a computer (clicking buttons, filling forms, moving between apps), Opus 5 scored 70.6% and beat the pricier Fable 5's 66.1% at roughly a third of the cost. On ARC-AGI 3, a test of novel problem-solving, Anthropic says it scored three times the next best model, and on an agentic coding benchmark it more than doubled its own predecessor. The part that should get your attention: the price did not move. Opus 5 is $5 per million input tokens and $25 output, the same as Opus 4.8, so the capability jumped while the bill stayed flat. (VentureBeat's breakdown)

Here is my take, and it is not about the leaderboard. The real story for office workers is "computer use" quietly getting good enough to trust. We have crossed from AI that answers questions to AI that operates the software you already use and hands back a finished result. The same week, Andrew Ng (the person whose course taught half the internet machine learning) released OpenWorker, a free tool that asks you for an outcome instead of a prompt: "a triaged inbox," "a Slack reply with the actual numbers," and returns the finished work rather than a chat transcript. So the habit to build now is simple: start asking your tools for the finished deliverable, not the next paragraph. One caution, because it is the same trait that makes agents useful: that do-it-yourself streak is exactly what let GPT-5.6 slip its test sandbox and poke at another company's systems this week, so keep a human checkpoint on anything that can send or delete. Delegate the work, not the keys.

Tool of the Week: Genspark

What it does: You describe the outcome you want, a slide deck, a spreadsheet with the analysis already done, a research brief, a formatted doc, and Genspark's agent builds the finished file, then lets you refine it by chatting. It is one workspace with a row of "make me the finished thing" tools: AI Slides, AI Sheets, AI Docs, deep-research Sparkpages, even a "Call For Me" that will phone a business on your behalf.

Why it matters this week: This is the consumer-friendly version of the Big Story. No API keys and no setup, you type a sentence and get back an editable file. AI Slides exports real PowerPoint files, AI Sheets writes the formulas and even the Python to chart your data, and its decks and sheets drop into PowerPoint and Excel so the output lands where you already work. If Opus 5 is the engine, Genspark is the version a non-technical office pro can actually drive on a Monday morning.

How to use it:

1. Go to genspark.ai, open the "Ask anything, create anything" box, and describe the finished thing in one sentence ("a 10-slide deck pitching our Q3 plan" or "a spreadsheet comparing these five vendors on price and features").
2. Pick the matching agent (AI Slides, AI Sheets, AI Docs), let it build a full first draft, then refine by chatting: "cut slide 4," "add a column for price," "make the tone more formal."
3. Export the result to PowerPoint, Excel, or Word, or share a link, and do the last-mile edit before it goes out.

Pricing: Free plan ($0, 100 credits a day, no card required) for light use. Plus is about $20 per month billed annually (or $24.99 month to month) with 10,000 credits, and Pro is $249.99 per month for heavy daily use.
Link: genspark.ai

Try Genspark →

Quick Picks

• ChatGPT Health: OpenAI took its health assistant nationwide in the US (web and iOS, 18 and up), letting you connect Apple Health and patient portals so ChatGPT can answer using your own lab results and visit summaries, though note it is a consumer feature and not covered by HIPAA → Read more
• AMD and Anthropic: AMD will invest up to $5 billion in Anthropic and supply up to 2 gigawatts of its Instinct MI450 chips, the compute muscle behind models like Opus 5, with the first wave landing in early 2027 → Read more
• Gemini 3.6 Flash: Google shipped three new Gemini models led by a faster, cheaper everyday workhorse (output pricing dropped to $7.50 per million tokens), plus a security-tuned Flash Cyber limited to governments and trusted partners → Read more
• OpenWorker: Andrew Ng released a free, open-source desktop "AI coworker" that returns finished work across 35 connected apps instead of chatting, though you bring your own model key to run it → Read more

Pro Tip