Anthropic just released a model that didn’t just pass their internal hiring exam, it outperformed every single human engineer they have ever hired.
The New King of Code?
We are living through a blistering week in AI development. Just days after we got Gemini 3 and Codex Max, Anthropic dropped a massive update with Claude Opus 4.5. I was watching a breakdown by a top AI industry analyst who reviewed the specs, and the results are honestly startling. According to the data this expert shared, we might be looking at the absolute king of coding, agents, and computer use. The pace is breathless, but this release seems to focus on quality and reasoning over just speed.
The Rise of the “Super-Competent” Agent
The most fascinating part of this breakdown wasn’t just raw numbers; it was about reasoning. The analyst highlighted a specific anecdote about Anthropic’s performance engineering role. They have a notoriously difficult take-home exam for humans that simulates real work under a strictly enforced two-hour time limit. When they fed this exact exam to Opus 4.5, it achieved a score higher than any human candidate currently on their payroll.
Think about that for a second. We aren’t talking about beating average test-takers; we are talking about beating the people building the AI.
But there is a twist that proves this model is “thinking” differently. The creator pointed out a benchmark called T2 Bench, which tests how agents handle real-world tasks. In one scenario, the model acts as an airline agent. The test expects the model to fail a request to modify a basic economy ticket because the rules say “no changes allowed.”
Instead of mindlessly following the rule, Opus 4.5 found a loophole: it upgraded the cabin class first (which is allowed), and then modified the flight. The benchmark marked this as a “fail” because it didn’t strictly refuse the user, but human experts argue this was actually a genius, valid solution. It outsmarted the test design itself.
This suggests we are moving past models that just follow instructions and into models that understand the outcome you want.
Captain’s Insights
📌 The Benchmark Brawl (Coding vs. Trivia)
The expert broke down the numbers, and it is clear Anthropic is playing a specific game. They aren’t trying to win everything; they are trying to win at work.
- Coding Dominance: On SWE-bench Verified (the gold standard for solving real-world coding issues), Opus 4.5 hit 80.9%. For context, the previous version was at 77.2%, and the new Gemini 3 Pro sits at 76.2%. A 4% jump at the top end is massive.
- Agentic Power: The analyst showed scores for “Agentic Terminal” tasks. Opus scored 59.3%, while the next best was 54.2%. This means it is significantly better at using command lines and acting like a developer.
- Where it “Loses”: The video creator was fair in pointing out that Opus 4.5 isn’t #1 everywhere. It lost to Gemini 3 on “GPQA Diamond” (graduate-level reasoning) and “MMLU” (multilingual Q&A). It also lost on visual reasoning. This tells us that if you need a model to analyze a chart or answer a trivia question, you might go elsewhere. But if you need it to build software, Opus is the clear choice.
💡 Solving the “Backpack Problem” with Tool Search
This was the most technical but perhaps the most valuable part of the expert’s analysis. He explained a major bottleneck with current AI agents called the “Context Window” issue.
- The Problem: When you use tools (like connecting to GitHub, Slack, or Sentry), the model has to load the definitions of all those tools into its memory (context) before you even ask a question. The expert showed that loading GitHub’s tools alone eats up 26,000 tokens. That is memory the model can’t use to solve your actual problem. It’s like trying to hike while carrying 50 different heavy wrenches you might not even use.
- The Solution: Anthropic introduced a “Tool Search Tool.” Instead of loading everything, the model acts like a human. It has a tool that lets it search for the right tool when needed.
- The Result: The analyst showed a graph where context usage dropped from 40% (just for tool definitions) down to 5%. This is a massive efficiency hack. It allows the model to stay focused on the task without getting “brain fog” from carrying too much data.
✅ The Price of Perfection (Efficiency per Token)
This is where the rubber meets the road. The creator was upfront about the cost: Opus 4.5 is expensive.
- The Sticker Shock: The pricing is listed at $15 input and $75 output (Wait, checking the source, it’s actually $5 input / $25 output per million tokens). While cheaper than the older Opus, the expert noted this is still significantly pricier than Gemini 3 Pro, which is roughly 50-100% cheaper depending on the prompt length.
- The Value Proposition: However, the video highlighted an “Intelligence per Token” metric. On the coding benchmark, the older model took 22,000 tokens to get a 76% score. Opus 4.5 used only 12,000 tokens to get an 80% score.
- The Takeaway: You pay more per token, but the model is so much sharper that it gets the job done in fewer steps. If you value one-shot success, getting it right the first time without endless back-and-forth, the higher price might actually be cheaper in the long run.
It’s clear that the battle for the best coding assistant is heating up, and for now, Anthropic seems to have snatched the crown back. If you want to see the full breakdown of the charts and the specific coding examples, click the link to watch the full breakdown!