Opus 5 is Anthropic's downfall

what is up guys 👋,

I have spent part of this week talking to my phone more than I have talked to my friends.

Before you ask, no, I have not finally lost it.

ChatGPT added Voice to Codex, which means I can talk to my phone while it remotely controls my laptop and coordinates multiple AI agents.

My laptop stays open at home or in my bag, and I can use my phone to start tasks, check progress, redirect agents and keep projects moving while I am away from my desk.

I have been using it for client work on the go, and it is probably the closest thing I have found to a real Jarvis setup.

OpenAI is not sponsoring this; I just find it really cool and genuinely useful.

Opus 5 is the beginning of Anthropic's downfall

It looks incredible on benchmarks. Then people started using it and found that it ignores instructions, overcomplicates simple tasks and confidently attempts things it cannot actually do.

ChatGPT Voice can now direct Work and Codex

You can start tasks, check your agents, interrupt them and redirect the work without sitting there typing every instruction.

GPT-5.6 Luna got 80% cheaper

If you use Sol Medium for routine coding work, try Luna Max. Your weekly usage may start feeling suspiciously close to infinite.

Firecrawl open-sourced a smarter PDF parser

If you work with a lot of PDFs and regularly feed them to AI agents, this will be incredibly useful and could save you a LOT of tokens.

Opus 5 launched with ridiculous benchmark scores.

People were creating complete games, websites and graphics from one prompt. For about two days, it looked like the best AI model ever released.

Then people started using it for actual work.

The same complaints kept appearing: ignored instructions, unnecessary complexity and a model that takes huge swings at problems it does not fully understand.

Then it confidently tells you it was right.

Opus 5 can still be brilliant at difficult autonomous work. The problem is that its intelligence has become harder to control.

And this pattern started before Opus 5.

Every Claude release since Opus 4.6 has climbed the benchmarks while feeling more bloated, rigid and difficult to understand.

Anthropic's own prompting guide now explains how to stop Opus from over-engineering tasks, narrating everything and expanding the scope of your request.

You now need a prompting guide to tell the smartest Claude model to simply do what you asked.

I think Anthropic has become too focused on winning benchmarks.

Benchmark scores look great in announcements and investor decks but much worse at actually using it.

OpenAI has ChatGPT with the largest user adoption.

Google owns an entire ecosystem.

Kimi and DeepSeek compete aggressively on price.

Anthropic built its reputation on having the best model.

If Claude keeps feeling worse to use, people will leave.

And judging by how many developers are already switching back to older Claude models, Codex or Kimi, that may have already started.

ChatGPT Voice now works directly with Codex inside the desktop app.

The really cool part is that I can control it from my phone.

My laptop remains the host. ChatGPT can access the project, local files and apps I have given it permission to use.

Then I can leave my desk, open Voice on my phone and continue directing the work remotely.

For example:

→ "Check what the agents working on the client project have finished."

→ "Open the website, test it on mobile and ask an agent to fix anything broken."

→ "Check my gmail and tell me if I need to reply to anyone urgently"

I can hear what is happening, interrupt whenever I want and redirect the agents without opening five different chats.

That is the part that feels like Jarvis.

I am still deciding what should happen. I just spend far less time sitting at my laptop translating every decision into a perfectly written prompt.

This week I have been using it to keep client projects moving while travelling.

I can check progress from my phone, explain feedback as soon as I think of it and send the agents in a different direction before they waste an hour building the wrong thing.

This “Jarvis” also needs your laptop to remain awake and connected, so we have not completely escaped electricity yet.

But this is one of the first AI features in a while that has genuinely changed how I work.

OpenAI cut the API price of GPT-5.6 Luna by 80%.

It now costs $0.20 per million input tokens and $1.20 per million output tokens (which is insane in itself).

More importantly for Codex users, Luna and Terra now consume fewer credits.

If you are using Sol Medium for routine coding, try Luna Max instead. It handles most normal implementation work surprisingly well and barely touches your usage.

There is also a free Codex plugin called sol-advisor that uses:

You still get the expensive intelligence where it matters, without paying Sol prices for every button and database query.

Firecrawl open-sourced a tool called PDF Inspector.

It runs locally, detects which pages contain normal text and only sends the difficult scanned pages to OCR.

The result is clean Markdown with headings, tables and structure preserved.

Your AI still uses tokens when it reads the extracted content, but you stop wasting them trying to understand broken PDF formatting or repeatedly processing the entire document.

If you regularly give reports, research papers or client documents to AI agents, this could save you a lot of time and money.

Your agent can finally read a PDF without somehow spending $4 and still missing the table on page three.

I also released person-to-skill.

You give it someone's books, interviews, documents or YouTube videos, and it turns their knowledge into an advisor you can talk to inside Claude Code, Codex and other coding agents.

Instead of asking a generic model for business advice, you can build a skill grounded entirely in one person's real work.

For example run the command below:

It will study the sources and create a new skill called ask-hormozi.

You can then use /ask-hormozi in Claude Code or Codex whenever you want advice grounded in Hormozi's actual work.

It is free and open source.

I also published a video about using Record a Skill in Claude.

You record yourself doing a task, explain what you are doing and why, and Claude turns the process into a reusable skill.

It is much easier than trying to write a perfect instruction document from scratch.

Watch the full tutorial and breakdown here.

That’s it for this week.

See you next Sunday,