Sonnet 5 costs twice what Haiku 4.5 does per token. Per page, closer to 2.6 times. What that buys.
Haiku 4.5 at $1 / $5 scores 18. Sonnet 5 at $2 / $10 scores 38, with five times the context and a year newer knowledge. The real price gap, and when Haiku wins.
GTM Architect & Growth Operator · Now · 16 September 2026 · 6 min read
On this page
Claude Sonnet 5.5 has replaced Sonnet 5
Anthropic shipped Claude Sonnet 5.5 on 28 September 2026 at the same $2 / $10 per million tokens Sonnet 5 listed at, and it scores 56 on Intelligence Index v4.3.2 against Sonnet 5’s 38. Sonnet 5 has moved to Anthropic’s legacy list. Everything below still describes Sonnet 5. For the current model, read Sonnet 5.5 against Sonnet 5.
Inside Anthropic’s lineup, the small-model decision used to be simple: Haiku for volume, Sonnet for anything that needs thinking.
It is still roughly that. The price gap is bigger than the rate card shows, though, and the speed gap is smaller.
| Info | Claude Haiku 4.5 | Claude Sonnet 5 |
|---|---|---|
| Input / output per 1M tokens | $1 / $5 | $2 / $10 |
| Tokenizer | Previous | Newer, ~30% more tokens |
| Effective per page vs Haiku | 1x | ~2.6x |
| Intelligence Index v4.3 | 18 (reasoning) | 38 (max) |
| Output speed, tokens/s | 84.7 | 79.5 |
| Anthropic's latency label | Fastest | Fast |
| Context / max output | 200K / 64K | 1M / 128K |
| Reliable knowledge cutoff | Feb 2025 | Jan 2026 |
| Thinking | Extended | Adaptive |
| Retirement | Not sooner than 15 Oct 2026 | Not sooner than 30 Jun 2027 |
The price gap is wider than 2x
Sonnet 5 lists at exactly twice Haiku 4.5’s price, and that is the number most comparisons stop at.
Anthropic’s pricing page adds that Claude 4.7 and later models use a newer tokenizer, one that produces approximately 30% more tokens for the same text. Sonnet 5 is on it. Haiku 4.5 predates it.
So the same prompt costs more tokens on Sonnet 5 before any rate is applied, and twice the rate times about 1.3 times the tokens is roughly 2.6 times the cost per page. Still a reasonable trade. Budget for 2.6, not 2.
Anthropic notes the exact increase depends on content and workload shape. I would count real prompts on both before signing a forecast, the same advice I gave for Sonnet 4.6 against Sonnet 5, where the same tokenizer shift eats most of a price cut.
What 2.6 times buys
Twenty index points, 38 against 18. That is wider than the whole step from Sonnet 5 up to Opus 5, and it comes at the cheap end of the lineup where nobody expects a gap that size.
Five times the context: 1M tokens against 200K, and twice the output at 128K against 64K.
Knowledge about a year newer: a reliable cutoff of January 2026 against February 2025. That alone can decide it for anything touching recent tools, APIs or events.
Key takeaway
At the small end, the jump from Haiku 4.5 to Sonnet 5 is the largest capability step per dollar in Anthropic’s lineup. It is not a premium tier. It is the tier where most work belongs.
The speed gap is smaller than the label
Anthropic labels Haiku 4.5 its fastest model and Sonnet 5 fast. On Artificial Analysis’s measurement the two are close: 84.7 output tokens per second for Haiku 4.5 with reasoning, 81.5 without, and 79.5 for Sonnet 5 at max effort.
Labels describe the lineup relative to itself, and time to a finished answer also depends on how much each model writes. But if speed was the main reason a pipeline sits on Haiku 4.5, it is worth measuring again.
What changes in the code
Haiku 4.5 uses extended thinking, the mode set with a thinking type of enabled and a budget_tokens value. Anthropic’s models page says that mode is deprecated on the 4.6 generation and not accepted on later models, and Sonnet 5 uses adaptive thinking steered by an effort parameter instead.
Any request that sends budget_tokens to Haiku 4.5 needs changing before the model ID moves. Context and output limits change in your favour, so nothing breaks there.
The retirement window
Anthropic lists Haiku 4.5’s retirement as not sooner than 15 October 2026, and Sonnet 5’s as not sooner than 30 June 2027. Both are floors, not scheduled shutdowns. The first one is a month away, which is reason enough to put a Haiku 4.5 migration plan on the list now.
Which one I would use
| The work | Model | Why |
|---|---|---|
| Most new work inside Anthropic's lineup | Sonnet 5 | Twenty more index points, five times the context, newer knowledge. |
| Anything needing knowledge from 2025 or 2026 | Sonnet 5 | Haiku 4.5's reliable cutoff is February 2025. |
| Short, simple classification at huge volume | Haiku 4.5 | Half the rate, and the index gap may not show on trivial tasks. Plan for its retirement window. |
| Prompts over 200K tokens | Sonnet 5 | Haiku 4.5 cannot take them. |
| Volume work, vendor is flexible | Neither | GPT-5.6 Luna scores 38 at $0.20 / $1.20, cheaper than Haiku 4.5. |
If your vendor is not fixed, the cheaper comparison is Haiku 4.5 against GPT-5.6 Luna. If Sonnet 5 is not enough, the next step up is Sonnet 5 against Opus 5.
What would change my mind
A new small Claude model on the current tokenizer and a newer knowledge cutoff. That would reset every row of this table.
A measured tokenizer difference on your own prompts well away from 30%. The 2.6 figure rests on Anthropic’s approximation.
Questions people asked
Is Claude Sonnet 5 better than Haiku 4.5?
Yes: 38 against 18 on the Intelligence Index v4.3, with five times the context window, twice the output limit and a knowledge cutoff about a year newer. Haiku 4.5 is half the price per token.
How much more expensive is Sonnet 5 than Haiku 4.5?
Twice as much per token, $2 / $10 against $1 / $5. With Sonnet 5’s newer tokenizer producing about 30% more tokens for the same text, closer to 2.6 times per page.
Is Claude Haiku 4.5 faster than Sonnet 5?
Anthropic labels it faster, but Artificial Analysis measured them close: 84.7 tokens per second for Haiku 4.5 with reasoning against 79.5 for Sonnet 5.
Should I use Claude Haiku 4.5 or Sonnet 5?
Sonnet 5 for almost all new work inside Anthropic’s lineup. Haiku 4.5 only for short, simple, very high-volume tasks, with a plan for its retirement window.
If you are sizing a model change across a large pipeline, the contact page is the fastest route to me.
