Claude Opus 5 Now Available

Claude Opus 5 just dropped. And honestly? It feels different this time. Not the usual "incremental improvement" press release.

Not another benchmark chart with cherry-picked tasks. This one actually changes how you work with AI. I've been testing it since the early access window opened last week. Here's what matters.

What Is Claude Opus 5 Anthropic's new flagship model. Released July 2026. The "Opus" tier has always been their most capable model — the one you reach for when Sonnet isn't enough and you need genuine reasoning, not just pattern matching. Opus 5 keeps that positioning but shifts the architecture underneath.

They moved away from the dense transformer approach that defined the Claude 3 family. The new model uses a mixture-of-experts design with dynamic routing, plus a separate verification layer that runs parallel to generation. Translation: it thinks before it speaks. And it catches its own mistakes mid-stream.

The context window sits at 2 million tokens now. That's not a typo. Two million. You can drop an entire codebase, a legal contract portfolio, or a multi-year research corpus in one conversation and it actually retains all of it.

Pricing came in at $15 per million input tokens, $75 per million output. Same as Opus 4. They didn't raise prices despite the compute jump. That surprised people.

Why It Matters Most model releases follow a predictable arc. Better benchmarks. Slightly longer context. Maybe a new modality.

You upgrade, you get 5-10% better results, life goes on. Opus 5 breaks that pattern in three ways. First, the verification layer. Every response goes through a self-critique pass before you see it.

The model generates, then a separate expert head evaluates the output for logical consistency, factual grounding, and instruction adherence. If it fails, it regenerates. You never see the failed attempt. You just get fewer hallucinations.

this means you can ask "analyze this 400-page SEC filing and identify every material risk factor with page citations" and trust the citations. I tested this on three filings last week. Zero fabricated page numbers. Zero invented risk factors.

That's new. Second, the dynamic routing means different types of reasoning engage different expert pathways. Mathematical reasoning routes through one subnet. Code analysis through another.

Creative writing through a third. They're not fully isolated — there's cross-talk — but the specialization shows. Complex debugging tasks that used to require careful prompting now work with a vague "this is broken, help. " Third, the 2M context isn't just a bigger bucket.

They built a hierarchical attention mechanism that maintains coherence across the full window. I fed it a 1.8M token technical specification (about 1,400 pages) and asked for a summary of section 7.3's dependencies on section 2.1. It nailed the cross-references. Previous models would hallucinate connections or miss them entirely.

How It Works The Architecture Shift Anthropic published a technical blog post alongside the release. The key changes: Mixture of Experts with 128 experts, 8 active per token. This isn't the first MoE model, but the routing is learned end-to-end with the verification objective. Experts specialize emergently — they weren't pre-assigned roles.

Some handle syntax, some handle semantic reasoning, some handle factual retrieval. The routing patterns are visible in the attention visualizations they released. Parallel verification head. This is the novel part.

In other news: Stocks Tumble Amid AI Spending and Oil Price Concerns and Australian(1) Regulator(2) Warns(3) of(4) Rare(5) Eye(6) Disorder(7) Linked(8) to(9) GLP‑1(10) Drugs(11) ->.

A separate transformer head (about 15% of total parameters) runs concurrently with generation. It receives the same context plus the tokens generated so far, and outputs a confidence distribution over "continue," "regenerate last sentence," "backtrack to paragraph," and "abort. " The main model conditions on this signal. Hierarchical attention for long context.

The 2M window uses a three-tier attention: local (sliding window, 8k tokens), segment (compressed representations of 64k chunks), and global (learned summary tokens updated every 256k). This avoids the quadratic scaling that makes naive 2M attention impossible. What This Means for Prompting You prompt differently now. The old "think step by step" tricks still work but they're often redundant.

The model's internal verification handles multi-step reasoning automatically. But new patterns emerge: Implicit verification requests. Instead of "check your work," you can say "be thorough" and the verification head activates more aggressively. There's a temperature-like parameter called verification_depth (0-1) in the API.

Default is 0.7. Crank it to 1.0 for legal/medical/financial work. Drop to 0.3 for creative brainstorming where you want more risk-taking. Cross-document synthesis.

With 2M tokens and hierarchical attention, you can now feed 50+ documents and ask for synthesis. "Here are 47 customer interview transcripts. Identify the top 5 unmet needs with supporting quotes from at least 3 transcripts each. " This worked.

Previous models lost the thread around document 15. Codebase-aware development. Drop your entire repo (up to ~500k lines comfortably). Ask for refactors, architecture reviews, security audits.

The model maintains file-level context across the conversation. I had it redesign a legacy authentication module across 12 files. It preserved the API contracts, updated tests, and caught a race condition the human team missed.

• context_mode (enum: "standard" | "hierarchical" | "full"): How the 2M window is managed. "hierarchical" is default and recommended. "full" uses naive attention — slower, sometimes better for precise cross-references.
• expert_routing_hint (optional string): Lets you nudge routing. Values like "mathematical," "code," "creative," "analytical." Subtle effect but measurable on edge cases. The Messages API format stays compatible. Existing integrations work unchanged. Common Mistakes Treating It Like Opus 4 With More Context Biggest mistake I see. People dump 500k tokens and expect magic. The hierarchical attention means relevance* matters more than ever. If you stuff irrelevant documents, the segment-level compression loses signal. Curate your context. Use the new context_analysis endpoint (returns relevance scores per document) before the main call. Ignoring Verification Depth Default 0.7 is tuned for general chat. For production workloads, you need to tune this. I've seen teams ship with default settings and get subtle errors in financial calculations that 0.95 would have caught. Conversely, creative teams complain about "sterile" output at 0.7 — they should be at 0.3. Assuming Zero Hallucinations The verification layer catches internal* inconsistencies and instruction violations. It doesn't fix knowledge gaps. If the model genuinely doesn't know a fact, it may still confabulate — just more coherently. The hallucination rate* dropped ~60% in Anthropic's evals, but the plausibility* of remaining