Skip to content

Claude Opus 4.5 just dropped.

Last week, it was Gemini 3 and GPT‑5.1 Codex Max.

2 min read

Originally posted on LinkedIn · 20 reactions · 14 comments · View original →

Claude Opus 4.5 just dropped. Here's what engineering leaders need to know.

Last week, it was Gemini 3 and GPT‑5.1 Codex Max. This week, Anthropic fires back with a frontier model that’s explicitly optimised for production engineering, agents, and computer use.

📈 The numbers that matter:

SWE-bench verified: 80.9% — the highest score ever recorded. For context, GPT 5.1 Codex Max sits at 77.9%, Gemini 3 Pro at 76.2%.

Terminal Bench 2.0: 59.3% — #1 for agentic terminal coding.

OSWorld: 66.3% — state-of-the-art for computer use (OpenAI and Google didn't release scores on this one).

  • ️ But here's what actually caught my attention: Anthropic gives their performance engineering candidates a notoriously difficult take-home exam. 2-hour time limit.

Opus 4.5 scored higher than any engineer they've ever hired.

Let that sink in for a minute. 🫠 (still think coding models just produce 💩?)

  • The efficiency story:

Opus 4.5 introduces an effort parameter (low, medium, high), letting you control how hard the model thinks per request.

At medium effort, Opus 4.5 matches Sonnet 4.5's best SWE-bench score using 76% fewer tokens.

At high effort, it beats Sonnet 4.5 by 4.3 percentage points while still using 48% fewer tokens.

  • ️ Intelligence per token matters when you're running agents at scale.

🛠️ Claude Code upgrades:

Plan Mode now asks clarifying questions upfront, generates an editable plan.md file, then executes. More precision, less rework.

🔓 The usage signal that matters most:

Anthropic removed Opus-specific caps. Max and Team Premium users now get roughly the same token volume as they had with Sonnet. (Can I get a Hooyah?)

Translation: Opus 4.5 is meant to be your daily driver, not a special-occasion model. 🚀

  • The cost reality:

$5 input / $25 output per million tokens. A 67% reduction from the previous Opus pricing.

Still roughly 2x Gemini 3 Pro, but if you're measuring value by output quality an d token efficiency, the math works differently.

We're past the point of debating whether AI Engineering transforms software development. Are we not?