Why 1 Million Tokens Changes Everything

On March 13, 2026, Anthropic made one of the most consequential pricing decisions in the AI industry: the 1-million-token context window for Claude Opus 4.6 and Sonnet 4.6 became generally available at standard pricing. No beta headers. No long-context surcharge. A 900K-token request now costs exactly the same per-token rate as a 9K one.

One million tokens is roughly 750,000 words — about 10 novels, or the entire codebase of a medium-sized software project. What was previously a fragmented, multi-step process of chunking, summarizing, and recombining can now happen in a single API call. This guide covers everything you need to know: technical specs, API setup, pricing analysis, model comparison, and battle-tested best practices.

The Road to 1M Tokens

Context windows have been the silent bottleneck of LLM applications since GPT-3's modest 4K limit. Claude 3 pushed the boundary to 200K tokens in early 2024, and Claude Sonnet 4.5 introduced a beta 1M window in mid-2025 — but with a required beta header and a 2x price multiplier once you crossed 200K tokens.

Claude Opus 4.6 launched on February 5, 2026, carrying the 1M context in beta. Just five weeks later, Anthropic flipped the switch: GA at standard pricing for everyone. The previous long-context premium — where Opus input jumped from $5 to $10 per million tokens beyond 200K — is gone entirely. This isn't just a feature update; it's a strategic signal that Anthropic considers large-context processing a baseline capability, not a premium add-on.

Opus 4.6 Technical Specifications

Here's what you're working with:

• Model ID: claude-opus-4-6
• Context window: 1,000,000 tokens (GA, no headers needed)
• Max output tokens: 128K (doubled from the previous 64K)
• Input pricing: $5 / 1M tokens
• Output pricing: $25 / 1M tokens
• MRCR v2 benchmark: 78.3% (highest among frontier models at 1M context)
• Max images/PDFs per request: 600 pages (6x increase from 100)
• Fast mode: Up to 2.5x faster output at premium pricing ($30/$150 per MTok)
• Availability: Claude Platform, Microsoft Foundry, Google Cloud Vertex AI

The 128K output ceiling is particularly significant. Combined with adaptive thinking, Opus 4.6 can produce comprehensive refactoring plans, full security audit reports, or multi-chapter analyses in a single response — something that was physically impossible with the previous 64K limit.

API Setup: Zero Code Changes Required

The best part about the GA release? You don't need to change a single line of code. If you were already using the beta header, it's now silently ignored. Any request to claude-opus-4-6 that exceeds 200K tokens automatically uses the full 1M window.

Here's a basic Python example with the recommended adaptive thinking mode:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    messages=[{
        "role": "user",
        "content": "Analyze this entire codebase for security vulnerabilities: [large code input]"
    }]
)

A few important migration notes for Opus 4.6:

Adaptive thinking is now the default. The old thinking: {type: "enabled", budget_tokens: N} is deprecated. Adaptive thinking lets Claude dynamically decide when and how much to reason, which is more efficient for variable-complexity tasks.

Prefilling is gone. Opus 4.6 returns a 400 error if you try to prefill assistant messages. Use structured outputs (output_config.format) or system prompts instead.

Effort levels matter. A new max effort level provides absolute peak capability on Opus 4.6. For Sonnet 4.6, medium effort is the sweet spot for most use cases.

Opus 4.6 vs Sonnet 4.6: Choosing the Right Model

Both models support 1M tokens, but they serve different purposes.

Where Opus 4.6 dominates: Long-context retrieval accuracy. On MRCR v2 (an 8-needle, needle-in-a-haystack benchmark at 1M tokens), Opus scores 78.3% while previous Sonnet models scored around 18.5% — a 309% improvement. If you're feeding the model hundreds of thousands of tokens and need it to reliably find and reason about specific details buried deep in the input, Opus is the only real choice.

Opus 4.6 also excels at multi-file architectural reasoning, complex refactors spanning multiple services, security audits of large codebases, and any task where "catching what others miss" is the priority. In real-world tests, it handled multi-million-line codebase migrations with the planning and adaptability of a senior engineer.

Where Sonnet 4.6 wins: Cost efficiency and speed. At $3/$15 per million tokens (40% cheaper than Opus), Sonnet 4.6 is the pragmatic choice for everyday tasks — code reviews, document summarization, customer-facing applications, and workflows where latency matters more than maximum depth.

The practical strategy: Use a two-stage pipeline. Let Sonnet 4.6 handle initial filtering, classification, or summarization at lower cost, then route the refined results to Opus 4.6 for deep analysis. This can cut costs by 40-60% compared to running everything through Opus.

Real-World Use Cases

Whole-Codebase Analysis

This is Opus 4.6's killer application. Load an entire medium-sized project (~200K lines) into context and run architecture reviews, vulnerability scans, or refactoring plans in a single conversation. Claude Code users on Max, Team, or Enterprise plans get the full 1M context with Opus 4.6, enabling this workflow directly inside the IDE.

Massive Document Processing

With 600 images or PDF pages per request (up from 100), you can now process entire contracts, research paper collections, or financial reports in one shot. Legal teams reviewing multi-hundred-page agreements, researchers cross-referencing dozens of papers, and analysts processing quarterly reports all benefit from eliminating the chunk-and-reassemble workflow.

Long-Running Agent Sessions

The new Compaction API (beta) provides server-side summarization that automatically condenses earlier parts of a conversation when you approach the context limit. This enables effectively infinite conversations — critical for agent workflows that run for hours or span multiple sessions. Sonnet 4.6 also features context awareness, where the model actively tracks its remaining token budget and manages its work accordingly.

Best Practices: More Context Isn't Automatically Better

Anthropic's own documentation makes this clear: "As token count grows, accuracy and recall degrade, a phenomenon known as context rot." The 78.3% MRCR score is impressive but not perfect. What you put in context matters as much as how much fits.

Place documents first, questions last. For inputs exceeding 20K tokens, put your documents at the top of the prompt and your queries at the bottom. Tests show up to 30% performance improvement with complex multi-document inputs.

Use XML tags for structure. When feeding multiple documents, wrap each in descriptive XML tags with source and content subtags. This helps Claude distinguish between sources and cite them accurately.

Leverage compaction for long conversations. Rather than manually managing context, let the Compaction API automatically summarize earlier turns. It's available in beta for both Opus 4.6 and Sonnet 4.6.

Use the token counting API proactively. Opus 4.6 returns a validation error (not silent truncation) when you exceed the context window. Estimate token usage before sending to avoid failed requests.

Understand thinking token economics. Previous-turn thinking blocks are automatically stripped from context window calculations. Extended thinking doesn't eat into your context budget for subsequent turns — a crucial efficiency for multi-turn reasoning sessions.

Cost Analysis

Let's be concrete. A full 1M-token input request with 10K tokens of output on Opus 4.6 costs approximately $5.25. That's not cheap for a single API call, but consider the alternative: manually splitting a codebase into 5 chunks of 200K, running 5 separate requests, and then spending another request to synthesize the results. The 1M window is often cheaper than the fragmented approach, while delivering better results because the model sees everything at once.

For organizations running repeated large-context analyses, the two-stage Sonnet→Opus pipeline is the most cost-effective pattern. Use Sonnet 4.6 ($3/$15) for triage and Opus 4.6 ($5/$25) for deep dives.

Looking Forward

The elimination of long-context pricing premiums signals where the industry is heading: context windows are becoming commoditized infrastructure rather than premium features. With Opus 4.6's 78.3% MRCR score proving that large contexts can be usable (not just available), and the Compaction API enabling effectively infinite conversations, we're entering an era where the constraint isn't how much the model can see — it's how well you engineer what it sees. Master context engineering now, and you'll be positioned to extract maximum value as these capabilities continue to improve.