Claude Opus 4.6 Guide: 1M Context on the Last Pre-4.7 Tokenizer

Claude Opus 4.6 is three Opus releases old, and on most capability measures its successors are ahead. It is still worth understanding, because it sits on a boundary: it is the last Opus model before Anthropic changed the tokenizer and tightened the request surface, and it is the first one with the long context window and adaptive thinking that define the current line.

That combination makes it cheaper per word than the Opus models that followed at the same list price, and more permissive about how you call it.

What arrived with 4.6

Anthropic released Claude Opus 4.6 on 5 February 2026 at $5 per million input tokens and $25 per million output, unchanged from Opus 4.5. The launch notes describe a model that plans more carefully, sustains agentic tasks for longer and works more reliably in larger codebases, with better code review and debugging.

The API changes mattered more than the prose:

• 1M-token context window, the first for an Opus model. It launched in beta, with prompts over 200K tokens billed at a higher rate. Anthropic now lists the full 1M window at standard pricing for Claude 4.6 and later models, so a 900K-token request bills at the same per-token rate as a 9K one.
• 128K maximum output tokens, double Opus 4.5. On the Batch API a beta header raises that to 300K.
• Adaptive thinking, where the model decides how much to reason, steered by effort. Manual budget_tokens still works but is deprecated.
• A max effort level, giving low, medium, high and max. The xhigh level arrived with Opus 4.7.
• Server-side compaction in beta, which summarises older context as a conversation approaches a threshold.

Anthropic lists its reliable knowledge cutoff as May 2025, the same as Opus 4.5, so the improvement was in behaviour rather than fresher knowledge.

The long-context numbers

The figure Anthropic used to support the 1M claim was MRCR v2, an 8-needle retrieval test, in its 1M-token variant: 76% for Opus 4.6 against 18.5% for Claude Sonnet 4.5. That is a large gap, and it is the right kind of test, because finding several facts spread across a huge prompt is where models usually degrade.

Other text-stated launch figures: 81.42% on SWE-bench Verified with a prompt modification, averaged over 25 trials; 53.0% on Humanity's Last Exam with tools; and roughly 190 Elo points ahead of Opus 4.5 on GDPval-AA. Note the qualifiers. A prompt modification and a multi-trial average are legitimate, but they are not the configuration you will run.

A window you can fill is not the same as a window the model uses well at every depth. Long context degradation and context length vs effective context cover how to test that on your own documents.

The tokenizer quirk that makes 4.6 cheaper per word

Opus 4.7 introduced a new tokenizer, and every Claude model since uses it. Anthropic says the same text produces roughly 30% more tokens on it than on older models. Opus 4.6 is on the older one.

Opus 4.6, 4.7, 4.8 and Opus 5 all list at $5 and $25 per million tokens. Because 4.6 counts fewer tokens for the same text, an identical prompt and an identical answer cost less on it. The exact difference depends on content, so measure it with the token-counting endpoint on your own prompts rather than assuming 30%. Tokenizer differences across models explains why the same text bills differently.

This does not make 4.6 the cheaper choice overall. A newer model that finishes a task in fewer turns can cost less despite the higher token count. It does mean comparing models on list price alone gets the answer wrong.

What 4.6 still accepts that later Opus models reject

Opus 4.7 tightened the request surface. On Claude 4.7 and later models, non-default temperature, top_p or top_k values return a 400, and manual thinking budgets are not accepted. Opus 4.6 still takes sampling parameters, and still takes budget_tokens as a transitional option while you tune effort.

If your pipeline depends on low-temperature sampling for repeatable outputs, or on a hard ceiling on thinking tokens, Opus 4.6 is the newest Opus that supports both without a rewrite. That is a reason to use it as a stepping stone, not a destination.

Where it falls behind

On capability, the later Opus releases and the Fable tier are ahead, and Anthropic now points new work at Claude Opus 5.5, which lists at $4 and $20. On Bedrock, Opus 4.6 is available through the older InvokeModel integration rather than the newer Messages-API endpoint that Opus 4.7 and later use. Anthropic lists retirement as not sooner than 5 February 2027.

The catalogue's :batch variant is the same model on the half-price asynchronous endpoint, at $2.50 and $12.50 per million tokens. It is a good fit for long-document jobs that do not need an answer in seconds.

When to choose it

• Choose 4.6 when you need 1M context and 128K output on an Opus model, your code still sets sampling parameters or thinking budgets, and token count drives your bill more than turns do.
• Choose Opus 4.7 or later when you need xhigh effort, the newer request features, or the capability gains, and you have removed sampling parameters.
• Leave 4.5 for 4.6 whenever the 200K window is constraining you; the price is identical.

Whichever you pick, measure cost per completed task on your own work. The list price is the same across four Opus generations, and the real cost is not.