Claude Opus 5 Didn't Just Get Smarter — Anthropic Got Out of Its Way
Site Owner
Published on 2026-07-26
Opus 5 is half the price of Fable 5 and tied with it on SWE benchmarks. The bigger story is that Anthropic deleted 80% of its own system prompt for the new model and the coding evals did not move. Here is the new context-engineering playbook and what to do with your own CLAUDE.md on Monday.
Claude Opus 5 Didn't Just Get Smarter — Anthropic Got Out of Its Way
On Friday July 24, 2026, Anthropic released Claude Opus 5. By the numbers it's a pricing ambush: $5 per million input tokens, $25 per million output — the same price as Opus 4.8, half the price of Fable 5, and within 0.5% of Fable 5 on CursorBench 3.2's max-effort tier (qbitai). Simon Willison noted that it currently leads the Artificial Analysis leaderboard, ahead of Fable 5 itself.
The louder story, buried in a separate engineering blog post by Claude Code's Thariq Shihipar the same day, is different. Anthropic deleted 80% of the system prompt for Claude Code, and the team's internal Coding evaluations didn't move. Not "barely moved." Didn't move.
That second fact is the one worth your weekend.
<!-- 配图：封面 -->
The benchmark gauntlet, in plain English
The release lines up against GPT-5.6 Sol more than against Fable 5. Anthropic put three numbers in the same table, on purpose (qbitai):
Frontier-Bench v0.1: Opus 5 at 43.3%, GPT-5.6 Sol at 34.4%. Opus 5's score is 2x Opus 4.8 at lower per-task cost.
CursorBench 3.2 max-effort: Opus 5 within 0.5% of Fable 5's peak, at roughly half the price.
GDPval-AA v2: 1861 vs 1736 (against GPT-5.6 Sol).
BrowserComp: 90.8% vs 90.4%.
The non-coding evals are where it gets weird:
: Opus 5's score is .
<!-- 配图：fig1 - context composition -->
<!-- 配图：fig2 - the six shifts -->
#Anthropic#Claude#上下文工程#AI工程#AI模型
Claude Opus 5: The 80% System Prompt Cut Behind It
ARC-AGI 3
3x the next-best model
Zapier AutomationBench: at equal cost, the pass-rate is ~1.5x the next-best.
OSWorld 2.0: beats Fable 5's best result at roughly one-third the cost.
Anthropic's life-science eval: organic chemistry +10.2 points, protein +7.7 points over Opus 4.8.
Independent reads roughly agree. Epoch's ECI for Opus 5 is 159, vs Fable 5's 161 — but on SWE-ECI (software engineering) they're tied at 161 (Latent Space). The headline number says "near Fable." The head-to-head on coding says "tied with Fable." Pick the framing you prefer; both are defensible.
Anthropic's own marketing language stays cautious — "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price" (Simon Willison quoting the announcement). Independent testers are not cautious. The qbitai user am.will built a working Rocket League clone on 27% of a 5x Max subscription. Noema, after running the same prompt at lower effort, summed it up in three words: "是个怪物" — a monster (qbitai).
Two of those three numbers — the price cut and the 80% prompt cut — are about the same thing. Anthropic is no longer trying to make the model smarter by shouting more instructions at it.
The 80% that disappeared
Thariq's post is about "the new rules of context engineering for Claude 5 generation models." The core data point is the cut itself. Claude Code used to ship a long system prompt. For Opus 5 and Fable 5, Anthropic removed more than 80% of it, and internal Coding evaluations showed no measurable loss (Anthropic engineering blog, qbitai).
What did the missing 80% look like? Thariq's own example is the cleanest window into the old code. The system prompt used to include lines like:
"In code, do not write comments by default. Never write multi-paragraph docstrings, never write multi-line comment blocks — at most one short comment line. Unless the user explicitly asks, do not create plans, decisions, or analysis documents; work from conversational context, do not rely on intermediate files."
That kind of rule, multiplied across every model failure mode Anthropic ever saw, is what 80% looks like. The instinct was reasonable: every time a model misbehaved, add a rule. Over a year of "the model did X, so forbid X in the prompt," the system prompt turned into an employee handbook (qbitai, AINLP Chinese report).
The new rule for the new model is the opposite. From Thariq's CLAUDE.md guidance:
"If you remove this rule, Claude can still get the answer from the code, tool definitions, or tests — it probably shouldn't sit in the context permanently."
That single sentence inverts two years of muscle memory. The default is no longer "add a rule." The default is "remove a rule."
Why more instructions hurt a smarter model
A model that gets confused by conflicting rules is not lazy — it's doing the math. Thariq gives the diagnostic in one paragraph: when the same request triggers a system prompt, a Skill, a CLAUDE.md, and a tool description, the model has to spend inference cycles deciding who wins. qbitai's Chinese paraphrase is sharper than the English original: five product managers standing behind a programmer, each saying "one important thing." The dev spends all their attention figuring out who to listen to.
The mechanism has three pieces:
Token cost. Every rule in the system prompt is a permanent subtraction from the context window you actually care about. Anthropic's own research note on effective context engineering for AI agents (linked in Thariq's post) treats this as a tax, not a feature.
Conflict cost. When two rules disagree, the model picks one. The choice is not visible to the user. Worse, the model's "obedient" answer often costs more reasoning than the model would have spent on the actual task.
Stale-knowledge cost. A rule written for Opus 3 hangs around to constrain Opus 5. As Thariq's team discovered, "the model already knows the better way, but the old rule is so specific that it has to spend extra reasoning to humor it" (qbitai).
None of these costs are new. What changed is the model. A weak model needs to be told not to delete files; a strong model has internal priors that already cover the rule. The rule then becomes overhead, with the model performing the work of being told twice.
This is also why the 80% cut is so weird on the marketing surface. If Anthropic just wanted to lower price, they could. Cutting the prompt is an admission that Anthropic's own past advice to itself was holding the model back.
What replaces the deleted 80%
The cut doesn't leave a hole. Anthropic moved the deleted 80% into a different structure. The new context composition, drawn from Thariq's diagram (qbitai), looks like this:
System prompt: small. Holds hard guarantees, the model's name, and the things that must be true for every request.
CLAUDE.md: ≤200 lines, project-specific. Real build commands, the architecture decisions that aren't visible in code, the library choices that deviate from the default.
Memory: automatic, cross-session. The model accumulates what it learned last time.
Reference materials: code, tests, screenshots, the actual broken test case, the real example page. Concrete, not described.
Tools: lean. Tool names, parameters, return shapes, and one example — not three walk-throughs of "how to press the button."
Two of the shifts are the most counterintuitive. First, front-load is dead. Instead of pasting every relevant rule at the start of the conversation, you put the rule where it will be used. Skills get loaded when the user triggers the relevant workflow. Tools get discovered when the agent decides to look for them. Anthropic's tool search tool documentation (linked in Thariq's post) is built around this idea — don't pre-load 200 tool definitions, let the model pick the three it needs.
Second, concrete beats abstract. Thariq writes the example directly: telling the model "the page should look premium" is worse than giving it a screenshot of a real page from your codebase. "Write ten rules saying tests must be comprehensive" is worse than giving the model the existing test suite and one example that breaks the implementation. The model is a much better reader of evidence than a reader of adjectives.
Where the new shape breaks
There's a hard edge to this. The 80% cut is a story about soft constraints — comments, docstrings, plans, formatting, file creation. It is not a license to delete your security guardrails.
Anthropic is explicit: the system prompt is guidance, not enforcement. Anything that must hold 100% of the time — production deploys, dangerous commands, secret rotation, compliance checks, irreversible deletes — has to live outside the model. It goes in allow-lists, CI checks, Git hooks, and dedicated review tools. Thariq's word for it: "you can delete 80% only because of where you put the guardrails."
This is the part most teams will get wrong on the first try. They will read the 80% headline, delete their safety sections, and ship a regression. The right read is more boring: move the safety out of the prompt and into infrastructure that the model cannot bypass. The prompt can say "be careful with this command." A hook that asks a human before the command runs is the difference between an audit and an incident.
There's a second limit. The 80% number is a Claude Code number, measured on Opus 5 and Fable 5, on the team's own Coding evaluations. It is not a universal law. Older models, domain-specific workloads, and high-stakes workflows still need more rules, not fewer. The portable lesson isn't "delete 80%." The portable lesson is the deletion heuristic — if the model can derive the rule from code, tests, tools, or context, the rule doesn't belong in the prompt.
What to do with your own CLAUDE.md on Monday
Four operations, in this order. None of them are optional, none of them are clever.
1. Trim to 200 lines. Anthropic's official number. Every line above that has to earn its place against the rule "if I remove this, can Claude still figure it out from somewhere else?" If yes, remove it.
2. Move procedures to Skills. Anything that's a multi-step workflow (release, migration, code review, test generation) doesn't belong in CLAUDE.md. It belongs in a Skill, loaded when the workflow is invoked. CLAUDE.md stays for facts; Skills stay for procedures.
3. Replace adjectives with artifacts. Where you used to write "tests must be thorough", link the test file. Where you used to write "the UI must be consistent", link the design screenshot. Where you used to write "we follow a specific commit message style", link the contributing guide. The model reads real things better than described things.
4. Run /doctor. Claude Code ships with a /doctor command that audits your Skills and CLAUDE.md. Thariq recommends it as the closing step of any context trim. Use it.
The result should feel wrong. A trimmed CLAUDE.md feels underwritten next to the one it replaced. That feeling is the signal. The model is doing the work the old prompt was doing, in inference cycles you were already paying for.
The default flipped
For two years the prompt-engineering default has been: the model is dumb, write more rules. Claude Opus 5 is the release where Anthropic stopped believing that. They deleted 80% of their own instructions to their own model, and their own benchmarks didn't notice. They moved the rules out of the prompt and into Skills, Memory, hooks, and CI. They told the model the destination and the guardrails, and let it pick the road.
This is the contract Anthropic is signing with Claude 5. More capability buys you less micromanagement. The teams that internalize that will pay for the model once. The teams that don't will pay for the model and for the 80% of context they made it carry — twice, in tokens and in reasoning.
The first real casualty of Opus 5 is going to be the rule list in your repository. Most of it never needed to be there.
