Comparison · verified August 21, 2026
Claude Haiku vs Sonnet — the gap closed in 2026
The standard advice — Haiku for volume, Sonnet for quality — was built on a price gap that no longer exists. When Sonnet cost $3 / $15 and Haiku cost $0.80 / $4, moving a classification pipeline down a tier cut the bill by nearly three quarters. That arithmetic drove a lot of architecture.
Today Sonnet 5 is $2 / $10 against Haiku 4.5 at $1 / $5. Factor in the newer tokenizer and the real gap is roughly 2.6×, not 3.75×. Meanwhile Sonnet 5 has five times the context, an eleven-month newer knowledge cutoff, and adaptive thinking. If you last made this decision in 2025, it is worth making again.
The verified differences
What each one costs
Last verified August 21, 2026
USD per million tokens, standard rates, no discounts applied. Read the three caveats underneath before you put these numbers in a spreadsheet.
| Model | Tier | Input | Output | Notes |
|---|---|---|---|---|
| Claude Fable 5 | frontier | $10 | $50 | The frontier model. Twice the price of Opus 5 for the hardest work. |
| Claude Opus 5 | frontier | $5 | $25 | The default flagship. Thinking on by default. Fast mode available at $10/$50. |
| Claude Sonnet 5 | balanced | $2 | $10 | What most production agent fleets run on. The September 2026 increase to $3/$15 was cancelled. |
| Claude Haiku 4.5 | fast | $1 | $5 | High-volume and latency-sensitive work. |
Anthropic changed tokenizers at Claude 4.7
Claude 4.7 and later — including Opus 5, Sonnet 5, and Fable 5 — use a newer tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. A $2/MTok model on the new tokenizer is not directly comparable to a $2/MTok model on an older one, or to another vendor. Compare cost per task, not cost per token.
OpenAI meters long context separately
GPT-5.6 publishes a second, higher rate for long-context requests: Sol goes from $5/$30 to $10/$45, Terra from $2/$12 to $4/$18, Luna from $0.20/$1.20 to $0.40/$1.80. Anthropic includes the full 1M-token window at standard pricing on Claude 4.6 and later — a 900k-token request costs the same per token as a 9k one. If your workload is context-heavy, that difference is larger than the headline gap.
Two of Google’s current rates are introductory
Gemini 3.7 Flash and 3.6 Flash are priced at $0.75/$3.75 only through December 31, 2026. Both double to $1.50/$7.50 in 2027. If you are building a twelve-month cost model on Gemini Flash, model the 2027 number, not the one on the page today.
Caching and batching move the number more than model choice
A Claude cache hit costs 10% of the standard input price, and the Batch API takes 50% off both directions; the two stack. A workload that reuses a large system prompt can land below a nominally cheaper model that you are calling uncached. Model selection is usually the third-largest lever, after caching and after not sending the context at all.
What the evidence says, and who is saying it
Almost every performance number in public circulation traces back to a vendor running its own harness. That does not make the numbers useless, but the attribution belongs next to the claim.
Verifiable from Anthropic’s published model and pricing pages
- Haiku 4.5: $1 / $5, 200k context, 64k max output, reliable knowledge cutoff February 2025.
- Sonnet 5: $2 / $10, 1M context, 128k max output, reliable knowledge cutoff January 2026.
- Sonnet 5’s introductory pricing became permanent on August 10, 2026 — the scheduled increase to $3 / $15 was cancelled.
- Claude 4.7-generation models produce roughly 30% more tokens for the same text than earlier ones.
All of the above is documented product fact rather than benchmark interpretation. The quality difference between the two tiers is real but not something either of us can put a number on for your workload.
How to actually pick, in about an hour
The tier question is unusually cheap to answer empirically, because you can run both against the same inputs for a few dollars. Do that instead of reasoning about it.
- Take 100 real inputs from the workload in question. Not synthetic examples. The distribution of weird cases is the whole point, and you cannot invent it.
- Run both tiers and diff the outputs. For classification and extraction, count exact disagreements. If Haiku matches Sonnet on 98 of 100, the tier question is settled and you should stop reading.
- Look only at the disagreements. On each one, decide which answer was right. Sometimes Haiku is. What matters is the rate at which Sonnet is right and Haiku is wrong, and what that costs you per occurrence.
- Check whether your prompt fits Haiku’s window. 200k versus 1M is the constraint that ends the discussion outright for document work and long agent sessions. Measure your real prompt, including retrieved context, not the version in your head.
- Check the knowledge cutoff against your domain. If your work touches libraries, products, or events after February 2025, Haiku 4.5 does not know about them. This is the failure mode people misdiagnose as "Haiku is dumber" — it is not reasoning, it is recency.
- Price both with caching on. A cache hit costs 10% of input on either tier. If your prompt is mostly a stable system prompt, caching compresses the gap between tiers far more than the tier choice itself does.
Choosing the right Claude model covers the wider family, and cost optimization covers the levers that outrank tier selection.
The honest bottom line
Use Haiku 4.5 when latency is the product — real-time chat, autocomplete, inline suggestions — or at genuinely high volume where a 2.6× effective difference still amounts to real money. It remains the right call for well-defined classification and extraction with a clear rubric, short prompts, and content that does not depend on anything after February 2025.
Use Sonnet 5 for most other things, and more often than the old advice suggests. At 2× the price with 5× the context, an eleven-month newer cutoff, adaptive thinking, and double the output ceiling, the case for reaching down a tier is much weaker than it was. A lot of pipelines are on Haiku because someone did this arithmetic in 2025 and never revisited it.
The mistake to avoid treating tier selection as your cost strategy. Caching cuts input by 90% and batching cuts everything by 50%, and the two stack — both of which move the bill more than dropping a tier does, without costing you any quality. See prompt caching before you downgrade a model.