Claude Sonnet 5 vs Sonnet 4.6: context, output ceiling, price

Same price, five times the room

On the pool lane the two ids are priced identically: $3 per million input tokens, $15 per million output, 0.1× multiplier, which bills at $0.30 and $1.50 per million. Nothing about the older generation is a budget option.

What separates them is capacity. claude-sonnet-5 takes a 1,000,000-token context and emits up to 128,000 tokens; claude-sonnet-4-6 takes 200,000 and emits up to 64,000. That is 5× the input room and 2× the output ceiling for the same per-token price.

Everything else lines up. Both list text, tools and thinking, both route through /v1/messages, and the request bodies are interchangeable. Releases are 2026-01-22 for 4.6 and 2026-05-14 for 5, just under four months apart.

curl -sS https://token-share.app/v1/messages \
  -H "x-api-key: $TOKEN_SHARE_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 2048,
    "messages": [{"role": "user", "content": "Summarise the failure modes in this log."}]
  }'

Whether 200K is a ceiling depends on what you send

A 200,000-token window holds a great deal of code. Most single-file edits, review threads and tool-calling loops never approach it, and in those sessions the extra 800,000 tokens on Sonnet 5 sit unused — you are not paying for them, but you are not getting anything from them either.

The ceiling becomes real in two places: a working set that genuinely does not fit, and a long agent run whose transcript keeps growing until the session has to be truncated. If either is a regular occurrence for you, that is the difference the comparison is actually about.

The 64,000-token output limit is the quieter constraint. Generating a long file, a full migration, or a large structured document in one response is where 4.6 stops and 5 keeps going.

On the stable lanes, the older id costs more

Both generations have a dedicated lane, and here the pricing does not match. stable-claude-sonnet-5 is billed off $2/$10 at 0.4×, which is $0.80 and $4.00 per million. stable-claude-sonnet-4-6 is billed off $3/$15 at the same 0.4×, which is $1.20 and $6.00 per million.

So the older generation is 1.5× the price of the newer one on the stable lane, while being exactly equal to it on the pool lane. If your reason for choosing 4.6 was cost, that reason disappears the moment you move to a dedicated lane.

A stable-* id is a dedicated lane served on our own key instead of the shared account pool, and it does not take part in pool failover. Same weights either way — the multiplier pays for the path, not for the model.

How to settle it on your own workload

Start from your token records rather than from the spec sheet. If your input tokens per request never come near 200,000 and your responses never approach 64,000, the capacity difference is not buying you anything and the two ids are, for your purposes, the same price and the same shape.

If they do approach those numbers, the question stops being a comparison and becomes a constraint — one id fits your requests and the other truncates them.

The remaining case is quality, which the catalog cannot answer. Same endpoint, same body, one field changed: run your evaluation set on both and compare the failures.

Common questions