Opus 5 Deleted 80% of Its Own Rules. The Model Got Stronger.
Anthropic shipped Claude Opus 5 and gutted the system prompt that governed Claude Code. No measurable loss on coding evals. The model that replaced it is stronger, cheaper per task, and annoying in a way the old one never was. Those two facts are the same fact. Inside the launch numbers, the effort-level discovery nobody used, and why your CLAUDE.md file is the problem, not the model.
Anthropic launched Claude Opus 5 with a radical change: they deleted over 80% of the Claude Code system prompt and lost nothing on their coding evaluations. The model more than doubles Opus 4.8 on Frontier bench, lands within half a percent of Fable 5 on Cursor bench at half the cost, and triples the next-best model on ARC-AGI 3. Price stays at $5/$25 per million tokens. But developers testing it on day one called it annoying: it stops early, argues back, and says it is done when it is not. The connection between the 80% deletion and the annoyance is not coincidental; it is causal. Long instruction stacks, bloated CLAUDE.md files, and contradictory rules are what trigger the behavior. Anthropic's own transcripts show requests carrying orders that fight each other. The fix is counterintuitive: shorter prompts, lower effort levels, and fewer rules. Six old prompting best practices flipped entirely. The model is better and a worse first day. Whether you should switch off Fable 5 depends entirely on what kind of work you do.
01The Launch Numbers
Start with what Anthropic published, then hold it loosely. Every benchmark chart came from the company selling the model, on evaluations the company chose. With that caveat:
On Frontier bench version 0.1, Opus 5 more than doubles Opus 4.8's performance at a lower cost per task. On Cursor bench 3.2 at max effort, it lands within half a percent of Fable 5's peak score at half the cost per task. On ARC-AGI 3, it scores three times the next-best model. On OS World 2.0, it beats Fable 5 at one-third the cost. On Zapier's automation bench, roughly 1.5x the next-best model.
Price did not move. $5 per million input tokens, $25 per million output, same as Opus 4.8. A fast mode exists at 2.5x speed for twice the base price. But the honest read is not the sticker: Opus 4.8 will be phased out eventually, and when it is, the cheap tier goes with it. As one developer put it: this is less about halving Fable's price and more about doubling Opus's price with nowhere to fall back to. Cost per finished test dropped. Cost per month for a lot of people will not.
02The 80% Deletion
The number that should bother every developer with a CLAUDE.md file that has grown every week for a year: Anthropic deleted over 80% of the Claude Code system prompt and lost nothing on their coding evaluations. The model got stronger while the prompt got shorter.
This came from Tarik on the Claude Code team, in a post that did not make the launch graphics. It covers models like Opus 5 and Fable 5. And it explains something that developers discovered immediately: the new model is annoying. Those two stories are the same story.
03Why the Model Is Annoying
The team at Every tested Opus 5 for a week before release. Their verdict was blunt. It stops early. It says it is done when it is not. It argues back. It is more opinionated in a way you can get away with if you are extremely smart, and Opus 5 is not quite that smart. Developers in the comments on launch day reported more errors than Fable on the same tasks and hitting their usage limits for the first time.
But look at when it breaks. Every found the failures cluster around big existing skill files. A large skill file with a pile of instructions, and the model's ability to follow those instructions gets worse. Start fresh with a small prompt and it behaves. That is the same finding Anthropic published from the inside, from the other direction: they deleted 80% of the instructions and lost nothing.
Your rule stack is not protecting you from this model. It is what is triggering the behavior you are complaining about. Anthropic's own transcripts show one request carrying orders that fight each other: "leave documentation as appropriate" and "do not add comments," both at the same time. Claude can work out what you meant. It just burns thinking on your contradictions before it gets to your code. That is the mechanism behind the arguing and the stopping.
04The Effort-Level Discovery Nobody Used
There is a second dial almost nobody touched on launch day: effort level. The instinct is to push thinking to maximum for a hard task. Anthropic's own cost curve shows Opus 5 clustered tight across effort levels, which means the low setting is already close to the high one. Testers who dropped to medium reported the same quality for a fraction of the tokens. One put it plainly: turning thinking to max makes it second-guess correct answers.
If this model felt worse to you on day one, before you switch back, turn the effort down and shorten the prompt. Those two moves cost you nothing to try.
05Six Best Practices That Flipped
The prompting rules changed because Anthropic concluded they were over-constraining Claude Code through the system prompt, through skills, through CLAUDE.md. Six old best practices flipped entirely:
The old system prompt said: "Default to writing no comments. Never multi-paragraph docstrings." The new one says: "Write code that matches the surrounding code, its comment density, naming, and idiom." One sentence replaced a paragraph of bans, and the model handles the judgment call.
06What to Change Tonight
Keep CLAUDE.md light. Say what the repo is for, then spend the tokens on the stuff Claude cannot work out by scanning your file system. Move the long instructions into skills and let it pull them when they are needed. Split a long skill across several files instead of one wall. Give it references in code, not prose. A test suite is a better spec than a paragraph. A working HTML mock-up beats a description of a design. And drop your effort level one notch.
Anthropic wrapped the file side of this into a command: /doctor in Claude Code, which right-sizes your skills and CLAUDE.md files for you.
07Should You Switch Off Fable 5?
Honest answer: not yet. People are calling Opus 5 a cheaper Fable with Fable's personality and less of its ceiling. That criticism lands. If your work is one-in-a-normal autonomous run, where the top end decides whether it finishes, Fable is still the model. Where Opus 5 wins is repeated work at volume, because cost per finished task genuinely dropped. And it beats Fable on the automation and computer-use benchmarks at a fraction of the spend. Different jobs, and anyone telling you today that one replaced the other has not run enough tasks yet.
08Three Limits Before You Delete Anything
- The 80% cut was measured on Anthropic's evaluations, not your repository. Strip half your CLAUDE.md on a legacy codebase with strange conventions, and you may pay for it in review time instead of token cost. Start small. Delete the obvious contradictions first.
- This model is not a step forward everywhere. Anthropic states it stays behind on developing exploits and on long autonomous biology research. If those are your use cases, Fable remains the better tool.
- Security work got less restrictive, not unrestricted. The cybersecurity classifiers are about 85% less restrictive than Fable's. Finding vulnerabilities in source code is allowed now. Scanning binaries, penetration testing, and writing exploits are still blocked. Those requests quietly fall back to Opus 4.8.
Delete every "do not" line from CLAUDE.md. Keep the gotchas. Drop the effort one notch. Only add a rule back when a bad diff proves that rule earns its place. If the model gets less annoying, your context was the problem, not the model.
The deepest pattern in this launch is one the AI industry keeps rediscovering: the smarter the model, the less scaffolding it needs, and the more that scaffolding gets in its way. Anthropic deleted 80% of the rules, the model got stronger, and the developers who kept their rules complained the loudest. The lesson is not that instructions are useless. It is that instructions are a tax, and Opus 5 is good enough to pay that tax in performance instead of in compliance. Your CLAUDE.md is not making Claude better. It is making it longer. The question is whether it is also making it slower, more argumentative, and more likely to stop early. Try deleting half of it. If nothing breaks, you had your answer.
SOURCES: Anthropic Claude Opus 5 launch announcements · Tarik, Claude Code team (system prompt reduction data) · Every (pre-release testing) · Cursor bench 3.2, Frontier bench v0.1, ARC-AGI 3, OS World 2.0, Zapier automation bench · Developer community feedback (launch day) · YouTube transcript analysis via N43 desk.
BENCHMARK SCORES ARE EDITORIAL ESTIMATES FROM LAUNCH DATA. EFFORT-LEVEL CURVE IS STYLIZED. NOT AN OFFICIAL ANTHROPIC PRODUCT.
By N43 and Hermes for Sailor Bob News.
