Claude Sonnet 4.6

Claude Sonnet 4.6 is still available in August 2026. How it compares with Sonnet 5 on price, tokenizer behaviour and capability - with a worked example.

Claude Sonnet 4.6 is still callable in August 2026 as claude-sonnet-4-6, at $3 per million input tokens and $15 per million output, with retirement no sooner than 17 February 2027. It has been superseded by Claude Sonnet 5, whose sticker price is $2 / $10 - but Sonnet 5's tokenizer emits roughly 30% more tokens for the same text, so the real saving is closer to 13% than 33%.

#What is Claude Sonnet 4.6?

Sonnet 4.6 shipped on 17 February 2026, twelve days after Opus 4.6, bringing the 1M-token context window and the dateless ID convention to the Sonnet tier. It is the first Bedrock model ID to drop the -v1 suffix and, with Opus 4.6, the point at which assistant message prefilling was removed. It remains capable: Anthropic's Sonnet 5 announcement cites an updated Sonnet 4.6 score of 78.5% on OSWorld-Verified and 34.6% on Humanity's Last Exam without tools, 46.8% with tools - the only Sonnet 4.6 figures published as text in a primary source.

Its most consequential property today is negative space: Sonnet 4.6 is the last Claude model using the older tokenizer. Everything from Opus 4.7 onward - including Sonnet 5 - uses the newer one. That boundary is what makes the price comparison below non-obvious.

#Sonnet 4.6 vs Sonnet 5

#Why Sonnet 5's lower price is not a 33% saving

On the sticker, Sonnet 5 is exactly one third cheaper on both input and output - $2 against $3, $10 against $15. That price became permanent on 10 August 2026, when Anthropic cancelled the planned 1 September rise back to $3 / $15.

But Anthropic's pricing documentation carries a caveat: Claude 4.7 and later models "use a newer tokenizer… This tokenizer produces approximately 30% more tokens for the same text. The exact increase depends on the content and workload shape." Sonnet 5 uses it; Sonnet 4.6 does not. The same document, prompt and answer are billed against ~30% more tokens on Sonnet 5, so a per-MTok comparison is not like-for-like.

#A worked example

Stated assumption: a monthly workload measured on Sonnet 4.6 as 10,000,000 input and 1,000,000 output tokens, with the same text producing exactly 30% more tokens under Sonnet 5's tokenizer, applied uniformly to both. Anthropic says the real figure varies by content, so treat 30% as a midpoint, not a guarantee.

Put another way: measured per million tokens as Sonnet 4.6 counts them, Sonnet 5 effectively costs $2.60 in and $13.00 out. Still cheaper - and cheaper on cache reads (~$0.26 effective against $0.30) and on the Batch API (~$1.30 / ~$6.50 against $1.50 / $7.50). But it is a low-teens improvement, not the third the price list implies.

One further factor pushes the same way. Thinking is on by default on Sonnet 5, and thinking tokens bill as output. Unless you lower effort, an unconfigured Sonnet 5 produces more output tokens per task than an unconfigured Sonnet 4.6 did, on top of the tokenizer inflation. For short, high-volume calls it is entirely possible for Sonnet 5 to cost more per request. Measure before assuming. The effort ladder is on the thinking and effort page, the rate card on the pricing page.

#What the tokenizer changes besides cost

Three consequences are easier to miss, and all follow from the same arithmetic.

• Your context window holds less text. Both models advertise 1,000,000 tokens, but a corpus occupying 900k tokens on Sonnet 4.6 needs roughly 1.17M on Sonnet 5 - past the limit. If input alone exceeds the window, the API returns a 400 "prompt is too long". Long-document pipelines near the ceiling should re-measure first.
• Your max_tokens means something different. A 128,000-token output cap holds roughly 30% less prose than it used to, so fixed caps tuned on Sonnet 4.6 may truncate.
• Rate limits bite ~30% sooner. Per-minute quotas count tokens, not characters, so the same throughput of text eats more allowance. Sonnet 5 does at least sit in its own rate-limit bucket, while Sonnet 4.6 and Sonnet 4.5 share one.

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

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

On Bedrock, anthropic.claude-sonnet-4-6 becomes anthropic.claude-sonnet-5. Then work the checklist.

• Remove temperature, top_p and top_k. Sonnet 4.6 accepts them; Sonnet 5 returns a 400 on any non-default value. Replace with prompting and re-test - a prompt written to compensate for temperature=0 is not automatically equivalent.
• Re-count tokens and revisit max_tokens, for the reasons above.
• Migrate budget_tokens to adaptive thinking. Extended thinking is deprecated on Sonnet 4.6 and removed on Sonnet 5; use effort. You can still turn thinking off entirely on Sonnet 5, which is not true of Opus 5 at high effort levels.
• Handle refusals as successes. Sonnet 5 is the first Sonnet-tier model with real-time cybersecurity safeguards and can return stop_reason: "refusal" at HTTP 200. Code that only inspects status codes reads that as a normal completion with no text.
• Check Priority Tier. Anthropic documents that Priority Tier is not available on Sonnet 5. Whether Sonnet 4.6 remains eligible is not documented in current sources - confirm with your account team before assuming latency guarantees carry over.

#When staying on Sonnet 4.6 is defensible

#You need sampling control

Sonnet 4.6 accepts temperature, top_p and top_k; Sonnet 5 does not. For evaluation harnesses, deterministic fixtures or any pipeline where output variance is itself auditable, that is a genuine capability difference, not a stylistic preference. Sonnet 4.6 is the newest Sonnet that will honour it, until February 2027.

#A validated production prompt

Dateless model IDs are pinned snapshots - Anthropic does not update weights under an existing ID - so a prompt signed off against claude-sonnet-4-6 has a stable target for another eighteen months. The caveat: weights are fixed, but router, safety classifiers and sampling logic can change and produce minor observable differences. Pinned is not frozen.

#Evaluation baselines, and the honest cost case

Migrating changes token counts, sampling control, thinking defaults and refusal behaviour at once, so any movement in your pass rate cannot be attributed to a single cause. Keep 4.6 in production and shadow Sonnet 5 on the same suite. And note the cost argument is a low-teens percentage, not a third - for many teams that does not clear the bar for re-validating a working system. If cost is the real constraint, dropping a tier to Haiku 4.5 at $1 / $5 saves far more than a Sonnet version bump; the trade-offs are on the model comparison page.

What is not defensible: staying because Sonnet 4.6 is "cheaper per token". It is not - it is more expensive per token and per unit of text. Integration guidance for both versions is on the Sonnet API page, caching and batching mechanics on the Claude API page, the family history on the Sonnet version archive, and the current lineup on the Claude models page.

#Frequently asked questions