Claude Sonnet 5 Guide: Legacy Status, Pricing and Reasons to Stay

Claude Sonnet 5 had a short reign. Anthropic released it on 30 June 2026 as its most capable Sonnet yet, and on 28 September Sonnet 5.5 replaced it as the current model. The documentation now labels Sonnet 5 "legacy": still fully available, no longer updated, and with a recommendation to migrate.

Legacy is not deprecated. Anthropic commits to keeping Sonnet 5 running on its own platforms until at least 30 June 2027. That gives you a real choice, and for a few specific workloads staying put is the better engineering decision.

The specification in one place

From Anthropic's model page:

• API ID — claude-sonnet-5, a pinned snapshot despite having no date suffix.
• Context — 1M tokens as standard, no beta header.
• Output — 128K tokens on the Messages API, and up to 300K on the Batches API with the output-300k-2026-03-24 beta header.
• Thinking — adaptive, steered by an effort parameter that defaults to high.
• Knowledge cutoff — January 2026, against June 2026 for Sonnet 5.5.
• Input — text and images, text out.

One detail catches people moving up from Sonnet 4.6 or earlier: setting temperature, top_p or top_k to a non-default value returns a 400 error. Sonnet 5 also introduced an updated tokenizer, which Anthropic says produces roughly 1.0 to 1.35 times as many tokens for the same input depending on content type.

The pricing history matters for your forecasts

Sonnet 5 launched with introductory pricing of $2 per million input tokens and $10 per million output, with a standard rate of $3 and $15 due from 1 September. On 10 August Anthropic made the introductory rate permanent and cancelled the increase. Charts in the launch post still show $3 and $15 because they predate the change.

If you built a cost forecast in July, it probably assumed a price rise that never happened. Re-run it. Current rates are $2 and $10, cache reads at $0.20 per million, five-minute cache writes at $2.50, and the Batch API at half price, which is what the :batch variant in OpenRouter's catalogue reflects at $1 input and $5 output. Batch API savings math covers when that asynchronous discount is worth the latency.

Reading its benchmarks against the successor

At launch, the figures attributed to Anthropic's materials put Sonnet 5 at 63.2% on SWE-bench Pro, against 58.1% for Sonnet 4.6 and 69.2% for Opus 4.8, and at 80.4% on Terminal-Bench 2.1. Anthropic's own pitch was that Sonnet 5 performs close to Opus 4.8 at a lower price, and can match it on some tasks at higher effort.

Three months later the comparison that matters is with Sonnet 5.5, and Anthropic's 5.5 launch table includes Sonnet 5 as the baseline:

• Terminal-Bench 4.0 — 10.3% for Sonnet 5, against 70.6% for Sonnet 5.5.
• CursorBench 4.0 — 34.1% against 55.5%.
• GDPval-AA v2.1 — 1,449 against 1,844.
• OSWorld 2.1 (partial) — 57.0% against 80.1%.
• Humanity's Last Exam with tools — 54.9% against 64.5%.

The Terminal-Bench 4.0 gap looks catastrophic, and it is worth reading carefully. It is a newer and harder version of the benchmark than the 2.1 run where Sonnet 5 scored above 80, so the two numbers are not comparable. What it does tell you is that on the hardest long-horizon terminal tasks Anthropic now measures, Sonnet 5 mostly fails and its successor mostly does not. Vendor-claimed vs independent benchmarks is a useful reminder that both columns come from the same vendor harness.

What Sonnet 5 does that Sonnet 5.5 will not

The case for staying is mostly about API behaviour, not quality.

• Forced tool use. Sonnet 5 accepts a tool_choice of type any or tool. Sonnet 5.5 rejects both with a 400 error. If a pipeline depends on forcing a specific extraction tool, Sonnet 5 runs it unchanged.
• The fallback target. Sonnet 5.5's safeguards can decline in five named categories, and Anthropic's server-side fallback retries its cyber and frontier_llm declines on Sonnet 5. In other words, Sonnet 5 is where Anthropic itself sends some requests its successor will not handle.
• Editable history. Sonnet 5.5 signs each thinking block over the conversation before it and, on newer accounts, rejects a replay after earlier history was edited. Pipelines that rewrite conversation history between calls have less to change on Sonnet 5.

Against that, Sonnet 5.5 costs the same per token, halves the cache read price to $0.10, lowers the minimum cacheable prompt from 1,024 to 512 tokens, and is claimed to need fewer tokens per task. For new work, it is the default.

When to stay and when to move

A short decision rule:

• Stay on Sonnet 5 if you depend on forced tool choice, if your logs show Sonnet 5.5 declining legitimate requests that Sonnet 5 answers (security tooling is the usual suspect), or if the pipeline is stable and validated and nothing about it is quality-limited.
• Move to Sonnet 5.5 for agentic coding, long tool loops and anything where the January 2026 knowledge cutoff hurts. The benchmark gap on long-horizon work is too large to ignore.
• Plan the move either way. A legacy model has a floor on its retirement date, not a guarantee beyond it.

The Sonnet 5.5 guide covers the successor's effort levels and migration changes, and if you are arriving from much further back, the Sonnet 4 guide covers the jump from the retired generation.