Simon Willison published a detailed analysis of the differences between Claude Opus 4.6 and 4.7's system prompts. It's on Hacker News right now with 158+ points.
The changes are real and significant. The model behaves differently. Prompts that worked on 4.6 may not work the same way on 4.7.
And I genuinely don't care. Here's why.
The system prompt treadmill
If you're paying $20/month for direct Claude API access, you've probably experienced this:
- You spend 4 hours engineering a perfect system prompt
- Anthropic releases a new model version
- The system prompt needs tuning again
- Repeat every 3-4 months
This is the hidden cost of raw API access. The token prices are visible. The time cost of prompt re-engineering isn't.
Anthropologists of AI pricing track three costs:
- Subscription cost: $20/month (visible)
- Token inflation: 45% increase from 4.6→4.7 (semi-visible, requires monitoring)
- Prompt maintenance cost: 4+ hours per model update (invisible, never shows up in invoices)
What actually changed in 4.7's system prompt
Willison's analysis shows changes in how Claude handles:
- Uncertainty and hedging language
- Refusal patterns
- Formatting preferences
- Response length calibration
For developers building on raw API: you need to read this analysis, test your prompts, and potentially rewrite your system prompts. This is legitimate engineering work.
For SimplyLouie users: none of this is exposed to you. The same ✌️$2/month interface. The same conversation experience. We absorb the prompt engineering changes so you don't have to.
The abstraction value proposition
This is the part that gets lost in AI pricing discussions: the sticker price is never the real price.
When Anthropic changes their system prompt:
| User type | What happens |
|---|---|
| Raw API user | Must audit prompt changes, test, re-engineer |
| Claude.ai user | May notice behavior changes, no control |
| SimplyLouie user | Nothing changes in your experience |
We run the same underlying model but our interface layer is stable. When Anthropic makes changes that would disrupt your workflow, we absorb them.
The real cost comparison
Let's be honest about total cost of ownership:
Claude API direct access:
- $20/month subscription
- ~$5-15/month in tokens (varies wildly with new model versions)
- 2-4 hours/quarter in prompt maintenance (worth $100-400 if you bill hourly)
- Total real cost: $125-435/quarter
SimplyLouie:
- $2/month, flat
- Zero tokens to track
- Zero prompt maintenance
- Total real cost: $6/quarter
The 3x price gap at the sticker level is actually a 20-70x gap when you factor in maintenance overhead.
For developers in markets where this hits harder
This math is even more stark outside the US.
| Country | ChatGPT cost | SimplyLouie | Hours of work to pay ChatGPT |
|---|---|---|---|
| Nigeria | N32,000/month | N3,200/month | ~2 full days of work |
| Philippines | P1,120/month | P112/month | ~1.5 days of work |
| India | Rs1,600/month | Rs165/month | ~1 day of work |
| Kenya | KSh2,600/month | KSh260/month | ~1.5 days of work |
| Indonesia | Rp320,000/month | Rp32,000/month | ~1 full day of work |
Every time Anthropic releases a new model, the direct API user in these markets pays twice: once in subscription fees, once in prompt re-engineering time. Neither is cheap.
The Simon Willison test
Here's a quick test for whether you need to track system prompt changes:
You need to track system prompt changes if:
- You're building a product on raw API calls
- Your prompts depend on specific Claude behaviors
- Your users will notice if Claude's tone/formatting shifts
- You charge customers and need predictable model behavior
You don't need to track them if:
- You're using Claude as a personal assistant
- You're doing writing, coding help, research, Q&A
- You want AI to work the same way it did last week
- You have better things to do with 4 hours
If you're in the second category: SimplyLouie exists for you. $2/month. No token tracking. No prompt engineering. No model update anxiety.
One more thing
Simon's analysis is excellent and you should read it if you're building on the API. Understanding how Claude's underlying behavior shifts is real engineering knowledge.
But there's a category of developers — probably most developers — who just want to use Claude, not maintain a relationship with its system prompt.
For that category: the abstraction layer is the product.
7-day free trial. Card required, not charged for 7 days. Then $2/month.
50% of every subscription goes to animal rescue. The other 50% keeps the lights on.
Top comments (0)