Claude Opus 4.6

Claude Opus 4.6 is the last Opus on the previous tokenizer, which gives it a quietly useful property: its 1M-token context window holds roughly 750,000 words, against about 555,000 on Opus 4.7 and later. Same nominal window, materially more text.

That is the one real reason to still be on it.

The fast mode trap

Model	speed: "fast"
Opus 5, Opus 4.8	Works, at premium pricing
Opus 4.7	Returns an error
Opus 4.6	Runs at standard speed, billed at standard rates

Opus 4.6 is the dangerous row. No error, no speed-up, no unexpected bill — just a request that quietly does not do what your code asked for. If you have speed: "fast" in a codebase that also targets 4.6, it is doing nothing there.

Specifications

Property	Value
Claude API ID	claude-opus-4-6
Context window	1M tokens (~750K words on the older tokenizer)
Max output	128K tokens
Adaptive thinking	Yes
Extended thinking parameter	Yes (deprecated)
Reliable knowledge cutoff	May 2025
Pricing	$5 in / $25 out per MTok

It is also the first generation where dateless model IDs appear — and, despite the appearance, they are still pinned snapshots.

Lifecycle

Active but legacy. Tentative retirement not sooner than 5 February 2027.