Hey from Tobias,

Claude Opus 5 landed this week at $5 per million input tokens, half of what the top-end Fable 5 costs. What's the actual headline: Anthropic deleted more than 80% of Claude Code's system prompt for these models and their coding scores did not drop.

Five things below: Opus 5 on price, the new rules for context engineering, the model layer being commoditized, why your next customer is an agent, and Jensen Huang's letter on X.

Opus 5 Costs Half of Fable 5

Fable 5 is still the most capable model Anthropic sells, at $10 per million input tokens and $50 per million output. Opus 5 gets close to it on harder tasks (long coding jobs, deep research, hours-long runs) and charges $5 and $25. Same 1 million token context window. Same feature set as the Opus you were already using.

The behavior shift matters more than the price. Opus 5 thinks before it answers by default, checks its own work without being asked, and pushes back when a request is wrong. Its judgment is genuinely better. That sounds like marketing until you look at what it means for your prompts.

Stop Overprompting Claude

For two years the winning move was to write longer instructions. Every failure got a new rule bolted on. Anthropic calls removing those rules "unhobbling", and Thariq Shihipar from their team published the new rules this week. Four that changed how I write instructions:

• Describe outcomes, not bans. Drop "never write long comment blocks". Write "match the comment style of the code around it" and let it read it.

• Delete your examples. Sample calls limit the model into copying your one pattern. Name your tools and inputs clearly instead and it explores further.

• Delete every "double-check". It already verifies itself. Telling it to verify again makes it slower and more expensive for no gain. This one surprised me most, because "ask it to check its work" was standard advice a month ago.

• Stop front-loading everything. Split your instruction files into small skills it can pull in when the task calls for it.

You do not have to do the cleanup by hand. Open Claude Code and run:

It reads your CLAUDE.md files and skills and flags what is now dead weight. Mine came back full of rules I wrote to fix problems the model no longer has. Deleting them made it faster.

The Model Layer Is Getting Cheap Fast

Look at the pattern. Near-frontier quality dropped to half price inside a few months. Open models like Kimi and GLM are a handful of points behind on coding tests and cost a fraction. Every few weeks the thing you were paying a premium for gets commoditized.

Which means the model is not your advantage. It is the cheapest, most replaceable part of what you build. Value is moving to the two ends: the product people use, and the infrastructure underneath. If your edge is access to a slightly better model, you have only a few weeks of tailwind.

Your Next Customer Is an Agent

Raoul Pal put it sharply this week: agents are becoming the main users of the internet, not people. That breaks the model most business software is built on. SaaS charges per seat, and a seat assumes a human who logs in, clicks around, and needs a pretty interface. An agent does none of that. It wants an endpoint, clear documentation, and a price per execution.

The market has already noticed. The S&P 500 software index is down roughly 20% this year, and Bloomberg has taken to calling it the SaaSpocalypse. My honest take: the products that survive are the ones an agent can operate with nobody watching. If yours only works when a person is looking at the screen, that is the thing to fix this year.

Jensen Huang's First Post Was About Open Models

On July 24 the NVIDIA CEO posted on X for the first time ever. He used it to share an open letter NVIDIA signed alongside Meta, Microsoft, and more than 20 other companies and organizations, arguing that open models are a requirement for American AI leadership.

Mark Zuckerberg backed it the same day, calling open source a force against centralization. Meta is a signatory too, which is no surprise given Llama models.

The timing is the tell. Washington is debating what to do about strong Chinese open models, whether export controls should cover model weights, and whether letting anyone copy a frontier model's behavior is a security risk. This letter is a push against rules that have not been written yet. The practical read: open weights are not going away, and running a small focused model on your own hardware stays a real option.

Delete more than you add this week.

Tobias