The headline numbers

Claude Opus 4.6 launched on February 5th 2026 and Anthropic is, predictably, very excited about it. Here are the stats they're leading with: 200K context window (1M in beta), 128K max output tokens, 80.8% on SWE-bench, 65.4% on Terminal-Bench 2.0, and the number one spot on Text Arena.

These are impressive numbers. But numbers on benchmarks and daily experience are different things, so I've been using Opus 4.6 as my primary AI for two weeks to see what actually changed in practice.

What's genuinely better: coding

Let me start with the unambiguous win. Opus 4.6 is measurably better at coding tasks than its predecessor.

Code review is where I noticed the biggest improvement. Paste in a pull request and Opus 4.6 catches things that previous versions missed. Not just syntax issues, but architectural concerns, potential race conditions, and subtle logic errors that would have made it to production. It caught a concurrency bug in my code last week that I'd been staring at for an hour without seeing. The explanation was clear enough that I understood not just what was wrong, but why it was wrong and why my original approach was flawed.

Debugging long codebases is another clear improvement. The 200K context window (and the 1M beta, which I've been testing) means you can feed it an entire project and it holds the relationships between files in a way that smaller context models can't. I gave it a full Express API with about 40 route files and asked it to trace a specific data flow. It nailed it, including a middleware that was silently modifying the request object three layers deep.

The 128K max output is also quietly significant. Previous models would truncate long responses, which was frustrating when you needed a complete file rewrite. Opus 4.6 can generate an entire large file without cutting off. Sounds small, saves a lot of "continue from where you left off" gymnastics.

What's genuinely better: agent teams

The "agent teams" feature is the headline capability for power users. You can ask Opus 4.6 to split a complex task into sub-tasks and spawn parallel sub-agents to work on each one independently.

In practice, this means you can say "refactor this authentication system to use JWTs, update all the tests, and update the API documentation" and it'll create three separate work streams, each with their own context and approach. The results are then merged.

I've been using this primarily through Claude Code, and it's impressive. A refactoring task that would have taken me an afternoon of back-and-forth with the AI took about 20 minutes. The sub-agents occasionally step on each other's toes (editing the same utility file in incompatible ways), but the overall time saving is substantial.

This is genuinely new. Not "incrementally better." Actually new. No other model I've used can decompose a task and parallelise it like this.

What's genuinely better: research

Long document analysis was already a Claude strength, and Opus 4.6 pushes it further. The 1M context beta lets you feed in entire research papers, financial reports, or legal documents and ask specific questions about them.

I tested it with a 300-page company annual report. Asked it to identify all the risk factors that could affect revenue in the next fiscal year. It found fourteen, including three that were mentioned only in footnotes. That's not something you can do reliably with any other model I've tested.

The data residency controls are worth mentioning for enterprise users. You can now choose between global routing and US-only routing for your data. For companies with compliance requirements, this removes a genuine blocker to adoption.

What's the same old story: safety hedging

Claude's safety training is still occasionally frustrating. Opus 4.6 is marginally better about not refusing reasonable requests, but you'll still hit moments where it gives you a careful, disclaimed, thoroughly balanced response when you just wanted a direct answer.

Ask it a question about a controversial topic and you'll get a miniature essay exploring all sides of the issue when sometimes you just want it to answer the question. This is a philosophical choice by Anthropic, and I respect the reasoning behind it, but in daily use it's still the most common source of friction.

It's better than it was. The refusal rate for reasonable requests has dropped noticeably. But it hasn't gone away entirely, and if you're coming from ChatGPT (which is generally more willing to engage directly with edgy topics), the difference is still noticeable.

What's the same old story: no internet

Claude still can't browse the internet. In February 2026. This is baffling.

Every time I want to ask about something that happened last week, I have to switch to ChatGPT or Perplexity. Every time I need to verify a current fact, I have to leave Claude and look it up. This is a significant limitation that Anthropic seems content to leave unaddressed.

Yes, the context window is enormous and you can paste in web content manually. But that's a workaround, not a solution. Copy-pasting from a browser into an AI chat in 2026 feels like using a floppy disk.

How it compares to GPT-5.2

The elephant in the room. OpenAI's GPT-5.2 launched with its own set of impressive capabilities, and comparing the two is unavoidable.

For coding: Opus 4.6 has a slight edge, particularly for complex, multi-file tasks. GPT-5.3-Codex (their coding-specific model) is competitive on single-file tasks.

For writing: Opus 4.6 is still better. The prose is more natural, more varied, and requires less editing.

For research with current information: GPT-5.2 wins by default because it has internet access.

For creative tasks: roughly equal, with different strengths. Opus 4.6 is better at nuanced, longer pieces. GPT-5.2 is better at structured, templated content.

For the agent teams / parallel task feature: Opus 4.6 is ahead. ChatGPT doesn't have an equivalent feature at this quality level yet.

The pricing reality

Claude Pro is still $20/month. The same price as ChatGPT Plus. For the quality of model you're getting, this is genuinely good value. The 1M context beta doesn't cost extra (yet), which feels like a limited-time gift.

If you're already on Claude Pro, the upgrade to Opus 4.6 is automatic and free. If you're considering subscribing, the coding and research improvements make the Pro tier more justifiable than it's ever been.

My honest assessment

Opus 4.6 is the best version of Claude yet, and it's the best model available for coding and writing tasks. The agent teams feature is genuinely innovative. The context window improvements are practical, not just impressive on paper.

But the lack of internet access is a real problem that keeps me from recommending it as a standalone tool. I still need ChatGPT or Perplexity for anything requiring current information, which means Claude is my "best tool for specific tasks" rather than my "one tool for everything."

If you code or write for a living, Opus 4.6 is worth your attention. The improvements are real and they compound. Two weeks in, I'm doing better work faster, and that's the only metric that actually matters.