Anthropic released Claude Opus 4.6 on February 5, 2026. The headline features look impressive: a 1-million token context window at launch, 128K max output, and benchmark scores that leave competitors behind. The pricing sheet looks identical to Opus 4.5—$5 per million input tokens, $25 per million output.
What the pricing sheet does not show—and what developers are discovering the hard way—is that Opus 4.6 consumes tokens at nearly five times the rate of its predecessor.
The Verbosity Problem
Artificial Analysis benchmarked Opus 4.6 across their Intelligence Index and found something striking. In non-reasoning mode, the model generated 11 million tokens during evaluation—compared to an average of 3.8 million across other models. In Adaptive Reasoning mode, that number increased to 58 million tokens against an average of 11 million.
The cost to evaluate Opus 4.6 on the Intelligence Index alone: $1,451.04. That number is worth sitting with for a moment.
This is not a flaw in the benchmarks. It is a fundamental characteristic of how the model operates. Opus 4.6 generates invisible "thinking" tokens to verify outputs and plan actions. Anthropic's researchers call this tendency "overthinking"—the model checks its conclusions more frequently, even on straightforward queries. The capability improvement is real. So is the cost increase.
The practical result: identical queries consume dramatically more tokens than they did on Opus 4.5. Same price per token, much larger bill.
The Context Window Trade-Off
The 1-million token context window is genuinely useful. Long-context retrieval performance increased from 18.5% on Opus 4.5 to 76% on Opus 4.6. For enterprise use cases involving large codebases or extensive documentation, this represents a real capability gain.
But there is a pricing detail that catches developers off guard, and it is worth understanding before running up a bill.
Once context exceeds 200K tokens, premium rates apply to the entire request—not just the tokens above the threshold. Input pricing doubles from $5 to $10 per million. Output pricing increases from $25 to $37.50.
A request that lands at 199K tokens costs one rate. A request at 201K tokens costs premium rates on all 201K tokens, not just the extra 1K. The math is punishing: slightly exceeding the threshold can meaningfully impact cost.
For codebases and AI-assisted development workflows that regularly push context limits, this creates a cliff in the pricing model that requires active management. The 200K boundary is not arbitrary—it is a decision point.
New Levers for Token Control
Anthropic anticipated the cost concerns. Opus 4.6 ships with several new mechanisms for controlling token consumption.
The Effort Parameter
Now generally available, the effort parameter directly controls how much the model "thinks" before responding. Low effort means minimal thinking, potentially skipping reasoning entirely. Medium provides a balanced approach for routine tasks. High, the default, enables deep reasoning. Max, available only on Opus 4.6, provides maximum capability for the hardest problems.
Anthropic's own documentation recommends reducing effort from "high" to "medium" when the model overthinks simple tasks. That recommendation exists because the problem is common enough to warrant official guidance. When the vendor tells you to use less of their product, it says something about the product.
Adaptive Thinking
The previous budget_tokens approach is deprecated. Adaptive thinking replaces it with a system where Claude evaluates each request's complexity and decides whether to think at all. At low effort levels, the model may skip thinking entirely for simple problems.
This helps—but developers report inconsistent results. Even with Adaptive Thinking enabled, the model sometimes reasons deeply for questions that do not require it. The "adaptive" part is not fully reliable yet.
Context Compaction API
Currently in beta, this feature addresses a different problem: long-running agent workflows that accumulate massive context over many tool calls. The API automatically summarizes older conversation parts when context approaches a configurable threshold.
The use case is clear—agentic coding assistants that run extended sessions. Without compaction, these workflows either hit context limits or incur premium pricing. With it, extended conversations become possible without manual truncation.
Practical Strategies for Cost Control
Developers and teams using AI coding tools have started converging on several optimization approaches.
Model Switching. The "opusplan" strategy uses Opus during planning and architecture decisions, then switches to Sonnet for implementation. This captures Opus-quality reasoning for complex analysis without paying Opus rates for every line of code.
Prompt Caching. Cache writes cost $6.25 per million tokens. Cache reads cost $0.50—a 90% reduction. For repetitive contexts like system prompts, codebase summaries, or documentation, the math works quickly.
Batch Processing. For non-time-sensitive workloads, batch processing offers a flat 50% discount: $2.50 per million input tokens, $12.50 per million output.
Context Awareness. For Claude Code users specifically, the /compact command helps when the model loses track of conversation context. The /clear command resets context when switching to unrelated work. Both reduce unnecessary token accumulation.
What This Means for Context Management
The broader pattern here is worth stepping back to see. As models become more capable, their token consumption characteristics become less predictable. A model that "overthinks" simple queries creates variable costs even for identical request types. Budgeting becomes guesswork.
This shifts the burden toward systems that can intelligently manage what goes into context and when. Retrieving only relevant code chunks rather than entire files. Understanding when a query needs deep reasoning versus a straightforward response. Knowing which model tier matches the complexity of the task.
Token efficiency was already important. With Opus 4.6, it becomes a cost management requirement.
The model's capabilities are impressive—68.8% on ARC AGI 2 (versus 37.6% for Opus 4.5), leading scores on Terminal-Bench 2.0 and Humanity's Last Exam, dramatically improved long-context retrieval. For hard problems, Opus 4.6 delivers. Nobody disputes that.
The question is whether "delivers" justifies "costs." The days of treating tokens as a simple input-output calculation are over. The same query can consume vastly different token amounts depending on the model's internal reasoning patterns. "Same price per token" no longer means "same price per query."
Teams building AI-assisted development workflows will need to factor this into their architecture. Token budgeting, model selection, and context curation are not optimization nice-to-haves. They are operational necessities. The alternative is surprise bills and difficult conversations about why the AI spend tripled without a corresponding improvement in output.