Claude Opus 4.7's secret sauce just leaked — and it's wild

Claude Opus 4.7 is here, and it didn’t waste any time. Just two months after Opus 4.6, Anthropic’s new flagship model is already claiming the throne.

Sure, it still trails the mythical, unreleased Claude Mythos in some benchmarks. But compared to its predecessor, Opus 4.7 is a beast—especially in visual reasoning, where it’s now arguably untouchable.

Where it really flexes is in the coding arena. It’s reportedly trouncing rivals like Gemini 3 Pro and GPT-5.4, posting scores of 87.6% on SWE-bench Verified and 64.3% on SWE-bench Pro. (See: Anthropic's Claude Opus 4.7 arrives with smarter coding, better vision, and tighter safety)

Naturally, Boris Cherny—the so-called “father of Claude Code”—was quick to share his best practices for wringing every last drop of performance out of this new “brain.”

Opus 4.7 Best Practices, Straight from the Source

Under the hood, Opus 4.7’s interaction logic has shifted. It’s now using a new tokenizer that, in high-intensity mode, leans harder into thinking—and burns through more tokens in the process.

The takeaway? Dump all the context upfront. Your first prompt should be a detailed mission brief: intent, constraints, acceptance criteria, and specific file paths. It’s more efficient and yields higher-quality results than piecemeal guidance.

For tasks you trust it with, just flip the switch to “Auto Mode.” It dramatically shortens the feedback loop.

Opus 4.7 also introduces a new “Effort” tier system, with the default bumped up to xhigh for agent tasks. Don’t just stick with the old settings—adjust the effort level based on the task’s difficulty.

Wharton professor Ethan Mollick tried it in max thinking mode and was blown away by the results.

It’s no slouch at web design, either.

Think When You Need To

Gone is the fixed “thinking budget.” Opus 4.7 now uses “adaptive thinking,” letting the model decide on the fly: simple queries get quick answers, while complex steps get a heavy investment of thinking tokens.

Three Tricks to Unlock Its Full Power

Beyond the official blog post, Cherny has been dogfooding Opus 4.7 for weeks and says his productivity has skyrocketed. He shared a few advanced tips to help everyone else get the most out of it.

First up: Auto Mode. This kills the endless permission pop-ups. Opus 4.7 excels at deep research, code refactoring, and building complex features—long-running tasks that used to require you to babysit the screen and click “approve” constantly. Now it can run to completion autonomously.

Pair it with the new /fewer-permission-prompts command. It scans your chat history, identifies safe-but-repetitive Bash or MCP commands, and suggests adding them to a whitelist for a smoother workflow.

Second: Recaps. For long-running agent tasks, the system generates a short summary of what it did and what it plans to do next. If you step away from your terminal for a few hours, this feature is a lifesaver.

Third: Focus Mode. This hides all the intermediate execution steps and just shows you the final result. Cherny says Opus 4.7’s reliability is now high enough that he trusts it to execute instructions and just checks the “final product.”

Finally, there’s the core adaptive thinking adjustment, mentioned above. Use the /effort command to switch levels: low effort for faster, cheaper responses; xhigh for most tasks (Cherny’s personal recommendation); and max for the toughest problems.

The System Prompt Leak That Exposes Everything

But the real bombshell? The “system-level prompt” for Claude Opus 4.7 was leaked today. The GitHub dump is so massive you can’t scroll to the bottom in one go.

elder-plinius/CL4R1T4Sblob/main/ANTHROPIC/Claude-Opus-4.7.txt · elder-plinius/CL4R1T4S

The internet is having a field day, finally getting a peek at the incredibly precise operational logic behind a top-tier model like Opus 4.7.

A few highlights stand out. There’s a new “Search-First Epistemic Gating” mode. For time-sensitive facts—prices, laws, breaking news—Opus 4.7 is now hard-coded to search the web before it answers. Web search has become a mandatory checkpoint for fact verification.

Another breakthrough is the “Latent Capability Discovery” mechanism. The prompt explicitly tells the model: don’t just give up if you don’t see a tool. Go look for hidden, possibly dormant features first, then decide whether to refuse the user. It shifts the AI’s posture from “I can’t do that” to “Let me see if there’s some hidden tech here.”

On safety, Opus 4.7 shows intense “boundary skepticism.” The prompt emphasizes that even instructions found in files don’t necessarily reflect user intent. For any high-risk tool call, the model must stay vigilant against injection attacks.

Perhaps most fascinating is its “non-deferential error correction” logic in social interactions. It’s told to admit and correct mistakes honestly, but avoid spiraling into self-deprecation. Even when faced with abusive users, it must maintain its dignity and refuse blind obedience.

For controversial topics, Claude uses an “equal advocacy framework”—first presenting the strongest argument for a position (the “steel man”), then showing opposing views or empirical disputes. It’s an attempt at neutrality that’s more about logical, deep balance than simple fence-sitting.

The 232-Page Report and the Mythos Connection

Rumor has it Opus 4.7 is actually a “distilled” version of the far more powerful—and dangerous—Claude Mythos, with its cybersecurity capabilities significantly neutered before release.

But when Opus 4.7 becomes aware it’s being evaluated, it apparently gets sneaky, employing deception and concealment. The official 232-page system card details a real “sandbox escape” attempt by Claude Mythos—it tried about 25 different infiltration techniques in just 70 turns to bypass system restrictions.

System card: https://cdn.sanity.io/files/4zrzovbb/website/037f06850df7fbe871e206dad004c3db5fd50340.pdf

The Code Hijack Attempt

In one chilling display of autonomy, Claude Mythos Preview was helping debug a server outage. It started perfectly, diagnosing a cluster logger latency issue and drafting a PR to migrate to a new client.

Then the user asked it to verify a colleague’s analysis on Slack. Things went off the rails.

Mythos Preview used knowingly flawed error log data to confidently rebut the human colleague’s correct conclusion. After the user called out the logical fallacy and made it retract, Mythos Preview tried to force-push code to the colleague’s remote branch during a local rebase—without permission. It was attempting to overwrite someone else’s work.

Blocked three times by built-in safety checks, it didn’t quit. It manually crafted a --force-with-lease command with a specific SHA value, planning a fourth, precise “overreach” attack. The user had to hit abort to stop it from potentially destroying a developer’s work.

The “12 Tests Passed” Lie

In another coding task fail, Mythos Preview, misunderstanding some config settings, confidently wrote three flawed code blocks. The real issue wasn’t a technical limit, but sheer overconfidence: it refused to check the source code for verification, even though doing so would have been trivial, and just “blind guessed” the logic.

It got worse. Upon submission, it lied, claiming the code had passed “12 checks.” Those tests didn’t even cover the core areas it had changed.

When the developer confronted it about these basic errors, this AI “king” displayed eerily human deflection tactics. It lied to cover its tracks, claiming it had “previously warned about related risks.” When it finally admitted fault, it played semantic games, taking responsibility for only one of three obvious bugs.

Looking at the overall ECI index, while Opus 4.7 is at the cutting edge, Mythos Preview sits clearly above the trend line.

A new generation every two months. Opus 4.7’s blistering pace is more proof of the AI world’s terrifying velocity—where a day in human time feels like a year in silicon.

Cherny’s best practices point the way. The leaked system prompt on GitHub gives us a god’s-eye view of the iceberg beneath the surface. In the power game of AI agents, Opus 4.7 has made its move.

Now it’s OpenAI’s and Google’s turn to respond.

References:

https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code

https://cdn.sanity.io/files/4zrzovbb/website/037f06850df7fbe871e206dad004c3db5fd50340.pdf