Claude Opus 4.6

Claude Opus 4.6 remains callable in August 2026. Full specs, how it compares with Opus 4.7 and Opus 5, prompt-caching differences, and when staying put is defensible.

Claude Opus 4.6 remains callable in August 2026 under the model ID claude-opus-4-6, at $5 per million input tokens and $25 per million output. Anthropic lists its retirement as no sooner than 5 February 2027. It has been superseded three times - by Opus 4.7, Opus 4.8 and now Claude Opus 5 - and it is the last Opus model whose world knowledge stops in 2025.

#What was Claude Opus 4.6?

Opus 4.6 launched on 5 February 2026 and introduced the 1M-token context window to the Opus line, initially in beta and with long-context pricing on requests over 200k input tokens. That premium has since gone: Anthropic's pricing documentation now states Claude 4.6 and later include the full 1M window at standard rates, so a 900k-token request bills at the same per-token rate as a 9k one. Opus 4.6 also shipped the compaction API in beta - server-side context summarization, still limited to 4.6-and-later models.

It is also a boundary in two API conventions. Opus 4.6 is the first model using the dateless ID format (claude-opus-4-6 rather than a snapshot date), and the last Bedrock ID to carry the -v1 suffix. And it is where assistant message prefilling was removed: on Opus 4.6 and everything after it, prefilling an assistant turn returns a 400.

The specification that ages worst is knowledge. Opus 4.6's reliable knowledge cutoff is May 2025, with training data to August 2025. Opus 4.7 jumped to January 2026 and Opus 5 to May 2026. For anything touching recent library versions, regulation or events, that gap is a real quality difference and no amount of prompting fixes it. Retrieval or web search is the workaround; the Claude API page covers the server-side web search tool.

#Opus 4.6 vs Opus 4.7 vs Opus 5

All three cost $5 in and $25 out per million tokens. Everything else has moved.

Note the tool-use overhead column. Opus 4.6 is cheaper per tool-bearing request than Opus 4.7 (497 tokens against 675), which is one of the few places the older model wins outright. Opus 5 beats both. Cost-per-task reasoning of this kind is worked through on the model comparison page.

#How do you migrate from Opus 4.6 to Opus 5?

One string, then four things to check. This is a bigger jump than 4.7 → 5, because two breaking changes landed between 4.6 and 4.7.

msg = client.messages.create(
    model="claude-opus-5",   # was "claude-opus-4-6"
    max_tokens=8192,
    messages=[{"role": "user", "content": prompt}],
)

#What actually breaks

• temperature, top_p and top_k now 400. This is the change most likely to break a working integration. Opus 4.6 accepts them; every model from Opus 4.7 onward rejects non-default values outright. The documented replacement is to delete them and control style through prompting - which usually means rewriting the part of your prompt that was compensating for a low temperature.
• Your token counts change. Opus 5 uses the tokenizer introduced with 4.7, which Anthropic says produces roughly 30% more tokens for the same text. Your max_tokens, your context budgeting and your cost model all need re-measuring against your own corpus. The Sonnet 4.6 page works through the arithmetic in detail.
• Extended thinking is gone. thinking: {"type": "enabled", "budget_tokens": N} is deprecated on Opus 4.6 and removed on Opus 5, where it returns a 400. Move to adaptive thinking and control depth with effort - low through max, defaulting to high. See thinking and effort.
• Thinking is on by default on Opus 5, and disabling it while effort is xhigh or max returns a 400. Expect more output tokens per task than an unconfigured Opus 4.6 produced.
• Prompt caching gets much easier. Opus 4.6's 4,096-token minimum is the highest of any live model; Opus 5's is 512. Prompts that could never be cached on 4.6 become cacheable on Opus 5 - see below.

#The 4,096-token cache minimum, and why it matters

Prompt caching only applies above a per-model floor. On Opus 4.6 and Opus 4.5 that floor is 4,096 tokens; on Opus 4.7 it is 2,048; on Opus 5 and Fable 5 it is 512. Below the floor, a cache_control breakpoint does nothing at all - no error, no saving.

The economics are otherwise identical across the Opus line: a 5-minute cache write costs 1.25× base input, a 1-hour write 2×, a cache read 0.1×. Caching pays for itself after one read on the 5-minute TTL, or two on the 1-hour TTL. So an agent with a 1,500-token system prompt gets no caching benefit at all on Opus 4.6, and roughly a 90% discount on that prefix on Opus 5. The multipliers stack with the Batch API's flat 50% discount; both are on the pricing page.

#When staying on Opus 4.6 is defensible

There is nearly six months of runway, and three arguments genuinely hold.

#You depend on sampling parameters

This is the strongest case, and it is specific to 4.6. If your system genuinely needs temperature=0 for reproducibility - evaluation harnesses, deterministic test fixtures, regulated pipelines where output variance is auditable - Opus 4.6 is the newest Opus model that will accept it. Everything after returns a 400. Migrating means replacing a parameter with a prompt, and you cannot assume equivalence. Budget real evaluation time for this, not an afternoon.

#A validated production prompt on a pinned snapshot

Dateless IDs are pinned snapshots, not moving pointers: Anthropic does not update weights under an existing model ID. If a prompt has been through review and sign-off against claude-opus-4-6, that behaviour is stable until February 2027. Anthropic's own caveat applies - weights are fixed, but router, safety classifiers and sampling logic can change, so pinned is not frozen.

#Evaluation baselines you cannot cheaply rebuild

Crossing the 4.6/4.7 boundary changes token counts, removes sampling control and changes thinking defaults simultaneously. If your regression scores were set against Opus 4.6, all three move at once and you cannot attribute a change in pass rate to any one of them. The clean approach is to hold 4.6 in production, run Opus 5 in shadow on the same suite, and switch on measured evidence. Claude Code's /claude-api migrate skill handles the mechanical part - model ID swaps and breaking parameter changes across a codebase - as covered on Claude for developers.

What is not defensible: staying for price. Opus 4.6 and Opus 5 cost the same, and Opus 5 is cheaper in practice through lower tool overhead and a lower cache floor. If cost is the constraint, the honest answer is a smaller model - Sonnet 5 at $2 / $10 or Haiku 4.5 at $1 / $5. The full lineup is on the Claude models page, and the Opus family history on the Opus version archive.

#Frequently asked questions