Claude Opus 4.6: what it is, what it costs, and what to use instead

On February 5, 2026, Anthropic launched Claude Opus 4.6 — the first Opus model with a 1 million token context window. With 128K max output, Agent Teams and adaptive thinking, it changed what was possible in a single prompt.

It is no longer Anthropic's flagship. Anthropic's own documentation lists Opus 4.6 as a legacy model and tells you to consider migrating to Claude Opus 5.5 (Anthropic, models overview, read 29 September 2026). Opus 4.6 is still available and is not retired before February 5, 2027, so nothing here is "out of date and unusable" — but if you are choosing a model today, this is not the one to choose.

This article keeps the numbers for Opus 4.6, because that is what a lot of existing code and configuration is pinned to, and it marks clearly where the current model differs.

What's new in Opus 4.6?

1 million token context

For the first time, an Opus-class model offers a context window of 1 million tokens (in beta). To put that in perspective:

• 200,000 tokens (Opus 4.5): Enough for a medium project, but you had to choose which files mattered most
• 1,000,000 tokens (Opus 4.6): Your entire codebase, all WordPress files, plugin code, configurations, database schema and relevant documentation — all at once

In practice, this means the model can understand relationships across your entire project. When you ask about a bug in your WooCommerce checkout, it can simultaneously see your theme, your custom plugins, your hooks, and your .htaccess — and give an answer that accounts for everything.

128K output tokens

Opus 4.6 can generate up to 128,000 tokens in a single response — four times more than its predecessor. That's equivalent to:

• A complete WordPress plugin with 3,000+ lines of code
• A detailed migration guide with code examples
• Comprehensive code reviews across multiple files
• An entire blog post with examples, tables, and code blocks

This capacity is particularly important for agentic workflows, where the model needs to generate large amounts of code without losing context or quality along the way.

Agent Teams: Parallel AI agents

Agent Teams is the most transformative feature. Instead of one AI instance working sequentially, multiple Claude instances collaborate like a real development team:

• An orchestrator analyzes the task and distributes it to specialized sub-agents
• Each agent works in its own tmux pane with its own context and focus area
• Frontend, API, database, and tests can be built simultaneously
• Agents communicate and coordinate via the orchestrator's overview

In a public demonstration, 16 parallel Claude agents wrote a C compiler in Rust with over 100,000 lines of code in just two weeks — with a 99% pass rate on the GCC test suite. That's a result that would normally require a team of 5-10 experienced developers over several months.

Adaptive thinking (Extended Thinking)

Opus 4.6 uses adaptive thinking — an internal reasoning process that scales with task complexity:

• Simple tasks (formatting, syntax): Minimal thinking time, fast response
• Medium tasks (implement a function, fix a bug): Moderate reasoning, well-considered response
• Complex tasks (architecture decisions, multi-file refactoring): Deep reasoning with planning, dependency analysis, and step-by-step execution

You can see this thinking process in the Claude Code terminal. It provides transparency — you can follow why the model makes certain choices, not just see the result.

Performance: what Anthropic actually documents

The benchmark figures I kept, and the ones I removed

Anthropic's launch post for Opus 4.6 states that on GDPval-AA — an evaluation of economically valuable knowledge work in finance, legal and other domains, run independently by Artificial Analysis — Opus 4.6 outperformed the next-best model (OpenAI's GPT-5.2) by around 144 Elo points, and its own predecessor (Claude Opus 4.5) by 190 points. Anthropic adds the caveat that this means it scored higher than GPT-5.2 on that eval about 70 % of the time, where 50 % would imply parity. That is a launch-day vendor claim from a specific evaluation, not a standing truth — model line-ups move fast — but it is a dated, attributable claim, so it stays.

What I removed from this article: the rows for Gemini 3 Pro and DeepSeek-V3.2 in that table, and the SWE-bench Verified / HumanEval+ percentage table. I could not find those specific numbers in a readable primary source. Anthropic's post does mention SWE-bench Verified — 77.2 % averaged over 25 trials, and 81.42 % with a prompt modification — but the GPT-5.2 and Gemini 3 Pro comparison figures I had could not be confirmed, so the comparison table is gone rather than half-sourced.

I have not replaced them with different numbers of the same kind. If you need benchmark comparisons for a buying decision, read the number you intend to rely on from the source, at the date you read it.

What Anthropic's documentation does state

These are verifiable, dated figures from Anthropic's own model documentation (read 29 September 2026):

Two things are worth noticing. 1M tokens is now standard, not a differentiator — Opus 5.5, Sonnet 5.5 and Fable 5.1 all have it, so "we run a 1M-token model" no longer distinguishes anything. And Opus 4.6 is no longer the cheapest way to reach that context window; Sonnet 5.5 is 2.5× cheaper per input token and Opus 5.5 is cheaper still.

Safety and alignment

Constitutional AI

Opus 4.6 is built with Anthropic's Constitutional AI approach — a training methodology that gives the model explicit principles for ethical and safe behavior. This means:

• The model refuses to generate harmful code (malware, exploitation)
• It actively warns about security issues in your code
• It suggests better alternatives when you use insecure patterns

No training on your data

Anthropic lists model training as opt-out on the consumer plans (Free, Pro, Max), and as none by default on Team and Enterprise. So your code is not used to train models, but on consumer plans you should confirm your own setting rather than assume it. This applies to:

• Claude.ai (Free, Pro, Max — opt-out)
• API access
• Claude Code

For businesses with strict compliance requirements, Anthropic offers enterprise agreements with contractual guarantees about data handling.

Responsible scaling

Anthropic follows a Responsible Scaling Policy — they thoroughly test for safety risks with each new model before launch. Opus 4.6 has undergone:

• Red-team testing of safety research
• Evaluation of autonomy and agentic capabilities
• Testing of guardrails for destructive actions
• External audit of AI Safety Level (ASL)

The Claude model family

Prices are per million tokens, input/output, base rate (batch is 50 % off). From Anthropic's pricing page, read 29 September 2026:

When to use which model?

• Opus 5.5: When quality matters most — long agentic coding runs, architecture decisions, deep analysis
• Sonnet 5.5: Your daily workhorse — code reviews, feature implementation, debugging, documentation
• Opus 4.6: Keep using it while it works, but do not build new spend around it — Anthropic's own guidance is to migrate
• Haiku 4.5: Quick tasks — formatting, simple questions, data transformation, classification

How to use Opus 4.6

Claude.ai (Pro/Max)

The simplest approach. Log in to claude.ai and select the model. Opus is included with Pro and Max, not with Free. Pro is $20/month when billed monthly, or $17/month on annual billing ($200 up front); Max starts at $100/month and offers 5x or 20x more usage than Pro (Anthropic pricing, read 29 September 2026).

Claude Code (terminal)

For developers, Claude Code is the most productive way to use Opus 4.6. It's Anthropic's agentic coding tool that runs directly in your terminal:

# Start Claude Code in your project
claude
 
# Give it a task
> Analyze this entire WordPress theme for security issues.
> Check for SQL injection, XSS, and insecure file operations.

Claude Code with Opus 4.6 can read your entire project, run commands, edit files, and handle git — all in one workflow.

API integration

For custom applications, you can use Opus 4.6 via Anthropic's API:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-4-6",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "Your prompt here"}]
  }'

Microsoft Foundry on Azure

For enterprise customers, Opus 4.6 is available via Microsoft Azure, with the compliance and security guarantees Azure provides.

What it means for WordPress developers

Your entire codebase in context

With 1M tokens, Opus 4.6 can analyze a complete WordPress setup — theme, child theme, 10+ plugins, configuration files, database schema, and .htaccess — and provide coherent suggestions based on the full picture. This eliminates the constant "context switching" where you manually had to tell the AI about your setup.

Practical examples

Complete security audit:

claude "Review this entire WordPress theme for security issues.
Check for: SQL injection, XSS, CSRF, insecure file uploads,
missing nonce validation, and direct database queries
without prepared statements."

Plugin development from scratch:

claude "Build a WordPress plugin that adds custom REST API
endpoints for a React frontend. Endpoints:
- GET /api/v1/products (WooCommerce products with filtering)
- POST /api/v1/inquiry (contact form with rate limiting)
Use WordPress coding standards, add PHPDoc,
and include uninstall.php."

Performance optimization:

claude "Analyze this theme for performance issues.
Look for: N+1 queries in loops, missing object caching,
large images without srcset, render-blocking JavaScript,
and unnecessary plugin calls in the critical render path."

Better debugging

Opus 4.6 can correlate errors across files. When your WooCommerce checkout throws a 500 error, it can:

1. Read the error log
2. Identify the failing function
3. Find the hook or filter causing the issue
4. Check if it's due to a plugin conflict
5. Suggest (or implement) the fix

Limitations and what to know

Speed

Anthropic does not publish a fixed latency figure per model, so I am not going to invent one. What is documented is that all current models use adaptive thinking, where the model decides how much to think — and that Opus 5.5 and Fable 5.1 run slower than Sonnet 5.5 and Haiku 4.5 (comparative latency, Anthropic models overview, read 29 September 2026). Longer thinking means longer answers on complex tasks. For quick answers, Sonnet 5.5 is the better choice.

Hallucinations

It's still an LLM. It can:

• Suggest WordPress functions that don't exist in your version
• Reference plugin APIs that are deprecated
• Generate code that looks correct but has subtle logical errors

Always review AI-generated code. Use it as a draft, not as finished code.

Price

Opus 4.6 is not the cheapest way to reach a 1M-token context window: at $5/$25 per MTok it costs 2.5× Sonnet 5.5 ($2/$10) per input token, and Opus 5.5 is cheaper still at $4/$20. If you are holding Opus 4.6 because a specific agentic task needs it, that is fine — but if you picked it for the context window, Sonnet 5.5 gives you the same 1M tokens for less than half the input price.

Rate limits

Anthropic does not publish a fixed per-model allowance. Usage limits reset on a rolling five-hour window, and paid plans add weekly limits on top; Max offers 5x or 20x more per five-hour session than Pro. Your Claude web, desktop, mobile and Claude Code activity all draw from one shared pool, and no fixed message count is guaranteed (Anthropic pricing FAQ, read 29 September 2026).

The "vibe working" era

CNBC has described the launch as the start of a "vibe working" era — where AI handles technical execution while humans focus on creative direction, architecture decisions, and client collaboration.

For freelancers and agencies, this means:

• More time on strategy: Instead of spending hours coding, you can focus on understanding client needs and designing the right solution
• Faster prototyping: Go from idea to working prototype in hours instead of days
• Better quality: AI catches bugs and security issues you might miss
• More ambitious projects: Tasks that were previously too large can now be tackled by a single developer with AI assistance

But it's important to emphasize: AI doesn't replace expertise. It amplifies it. An inexperienced developer with Opus 4.6 will still produce inferior code compared to an experienced developer with the same tool — because prompt quality and the ability to evaluate output still depend on human knowledge.

Conclusion

Claude Opus 4.6 did two things: it gave the Opus class a 1 million token context window, and it showed that adaptive thinking and Agent Teams were workable rather than theoretical. Both were real when it launched on 5 February 2026.

It is no longer the model to start with. Anthropic lists it as legacy and recommends Claude Opus 5.5, which is cheaper per input token at the same 1M context and 128K output. Anything you build on Opus 4.6 today should be a conscious, pinned choice — not an inherited default.

What this means if you don't have a developer

Most people reading this are not choosing a model. They are running a business and finding out that the thing they depend on — the site, the hosting, the integrations — now has an opinion about which model it runs on.

Here is the part that matters, and it is not the model: choosing the model is the easy half. Opus 4.6, Opus 5.5, Sonnet 5.5 — every one of them writes plausible code that is wrong in ways only a careful review catches. The hard half is deciding what the AI is allowed to touch, who reviews the output, what happens when a plugin update breaks an integration it generated, and what the fallback is when the API is down on the morning a campaign goes live.

That is the part that does not scale by adding a better tool. It is ownership of a production system.

If that is the problem you are actually facing, that is the work I do as a technical partner — you get someone accountable for the whole stack, not a licence key. See what technical partnership costs (from 2,500 DKK/month for drift, 15,000 for full partnership), or tell me what you are running into and I will tell you whether it is worth fixing.

Sources

All figures and statuses in this article were checked against these primary sources on 29 September 2026:

• Anthropic — Claude Opus 4.6 model reference — legacy status, context window, max output, API pricing, knowledge cutoff, retirement date
• Anthropic — Models overview — current lineup, comparative latency, current model pricing
• Anthropic — Pricing — Pro and Max plan prices, usage limits, batch and prompt-caching rates
• Anthropic — Claude Opus 4.6 announcement — release date and launch claims

What I removed, and why: the Gemini 3 Pro and DeepSeek-V3.2 rows from the GDPval-AA table, and the SWE-bench Verified / HumanEval+ comparison table, because I could not confirm those specific figures in a readable primary source. The GDPval-AA Elo figures that Anthropic's own launch post states are kept and attributed. Nothing was replaced with different numbers of the same kind.