· CostMyTokens · Analysis · 7 min read
Claude Sonnet 5 vs 5.5 vs 4.5: API Pricing Compared
Three generations of Claude Sonnet compared on official Anthropic list prices — Sonnet 5 and 5.5 share $2/$10, Sonnet 4.5 costs 50% more and retires November 30, 2026, plus the tokenizer change that shrinks the headline saving.
Three generations of Anthropic’s mid-tier sit on the price list at once, and the shape is unusual: current Claude Sonnet 5.5 and legacy Claude Sonnet 5 (still available) cost exactly the same — $2 input / $10 output per 1M tokens — while the older Sonnet 4.5 costs 50% more per token at $3 / $15 and is already deprecated, retiring on November 30, 2026. So this comparison is not really “which Sonnet is cheapest per token.” It is two sharper questions: if 5 and 5.5 cost the same, what actually separates them — and if you are still pinned to 4.5, how much of the headline 33% saving survives the tokenizer change that arrived with the newer models?
Prices are Anthropic’s official API list prices from its pricing page, verified 2026-10-28, the same source behind our Anthropic provider page. Standard workload for monthly figures: 10M tokens, split 8M input / 2M output, no caching, standard tier. Claude subscriptions (Pro, Max, Team) are a separate product and are not mixed in.
Disclosure: this article contains no affiliate links. Where CostMyTokens does use affiliate links elsewhere, they never affect order — every table is sorted by price only. Read our disclosure.
The three price rows
| Model | Status | Context | Input / 1M | Cached input / 1M | Output / 1M | Cost for 10M tokens / month |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | Current (released Sep 28, 2026) | 1M | $2.00 | $0.10 | $10.00 | $36.00 |
| Claude Sonnet 5 | Legacy (still available) | 1M | $2.00 | $0.20 | $10.00 | $36.00 |
| Claude Sonnet 4.5 | Deprecated Sep 30, 2026 — retires Nov 30, 2026 | 200K | $3.00 | $0.30 | $15.00 | $54.00 |
One history note matters for anyone who budgeted earlier this year: Sonnet 5’s $2 / $10 launched as introductory pricing with a scheduled rise to $3 / $15. Anthropic cancelled that rise and made the introductory rate permanent in August 2026, and Sonnet 5.5 launched at the same $2 / $10 in September. Pages still quoting a September price step for Sonnet 5 are out of date — the current list price for both Sonnet 5.5 and Sonnet 5 is $2 / $10.
Sonnet 5 vs 5.5: same price, three real differences
With input and output rates identical, the 5-to-5.5 decision is decided off the main price columns:
- Cache reads are half the price. Sonnet 5.5 reads cached input at $0.10 per 1M against Sonnet 5’s $0.20 (Anthropic cut the 5.5 cache-read rate in October 2026). On cache-heavy traffic this is the whole ballgame: at the standard workload with half the input served from cache, Sonnet 5.5 costs $28.40 a month against Sonnet 5’s $28.80 — a small gap at that mix, but it widens with every point of cache share, and cached reads are where agent workloads actually live. Cache writes (5-minute) cost the same $2.50 per 1M on both.
- Speed. Anthropic rates Sonnet 5.5’s output generation at more than 30% faster than Sonnet 5. At identical prices, latency is a free upgrade.
- Fewer tokens per task. Anthropic’s launch claim for 5.5 is a cost per completed task up to 30% below Sonnet 5, because the model reaches a finished answer in fewer tokens and steps. That is the vendor’s figure, measured on their evaluations — treat it as a reason to A/B your own workload, not as a budgeting input.
There is no price reason to choose Sonnet 5 over 5.5. Sonnet 5 remains on the list as a pinning option for teams mid-migration; new work should start on 5.5.
The tokenizer caveat: a 33% cut that lands at about 13%
Against Sonnet 4.5, the sticker math looks decisive: $3 / $15 down to $2 / $10 is a 33% cut on both columns, and the standard workload falls from $54.00 to $36.00. One change eats part of that. Sonnet 5 and 5.5 use the newer Claude tokenizer (introduced with the 4.7 generation), which produces roughly 30% more tokens for the same text than the tokenizer Sonnet 4.5 uses, depending on content.
Priced on identical text rather than identical token counts, the same workload that costs $54.00 on Sonnet 4.5 costs about $46.80 on Sonnet 5.5 — the token count inflates by ~30% in both directions (10.4M input-equivalents × $2, plus 2.6M output-equivalents × $10). The real saving is roughly 13%, not 33%. It is still a saving, and two lines escape the tokenizer effect entirely:
- Cache reads fall from $0.30 to $0.10 — even after re-counting cached text at the new tokenizer’s rate (~$0.13 effective), that is still well over 50% cheaper per repeated prompt.
- Context grows from 200K to the full 1M window at standard rates, which removes a class of chunking workarounds 4.5 users pay for in engineering time instead of tokens.
Before migrating, re-run your real prompts through Anthropic’s token-counting endpoint: your inflation factor depends on your content mix, and budgets copied from 4.5 dashboards will understate token counts on either newer model.
Discounts: Batch halves all three
Anthropic’s two standing discounts apply across all three generations:
- Batch API: 50% off everything. Sonnet 5.5 and Sonnet 5 batch at $1 / $5 per 1M ($18.00 at the standard workload); Sonnet 4.5 batches at $1.50 / $7.50 ($27.00). For work that can wait, Batch erases the generational gap proportionally — 4.5 stays 50% more expensive than 5.5 at every tier.
- Prompt caching: reads at the cached rates in the table above; 5-minute cache writes cost 1.25× the input rate ($2.50 on Sonnet 5 and 5.5, $3.75 on Sonnet 4.5). On Sonnet 5.5 a 1-hour cache write is listed at $4.00 per 1M.
A worked example at app scale: a chat assistant handling 10,000 conversations a month at 2,000 input and 500 output tokens each costs $90.00 a month on Sonnet 5.5 or Sonnet 5, and $135.00 on Sonnet 4.5 — before caching, which all three support and which widens the gap further in 5.5’s favour.
The November 30 deadline
Sonnet 4.5 was deprecated on September 30, 2026 and retires from Anthropic’s first-party API on November 30, 2026 (per Anthropic’s model deprecations page, checked 2026-10-29). After that date, claude-sonnet-4-5 calls stop working regardless of price, so for anyone still pinned to it this comparison has a forced answer: migrate to Sonnet 5.5 (not 5 — same price, cheaper cache reads, current generation), re-count your tokens before trusting the saving, and do it in November, not on the deadline weekend. If you reach Sonnet 4.5 through a cloud platform rather than the Anthropic API directly, check that platform’s own lifecycle page — retirement dates there can differ.
Verdict
- Default to Sonnet 5.5. It matches Sonnet 5’s $2 / $10 sticker, halves its cache-read price, generates faster, and is the current generation — at the standard workload it is $36.00 a month, $28.40 with half the input cached.
- Choose Sonnet 5 only if you are pinned to it. Same rates, double the cache-read cost, no offsetting advantage.
- Treat Sonnet 4.5 as a migration task, not an option. It costs 50% more per token ($54.00 at the standard workload), its real saving after the tokenizer change is ~13% rather than 33% — and it stops being callable on November 30, 2026 anyway.
Run your own mix on the CostMyTokens calculator, price the current model on its Sonnet 5.5 page, or see the whole Claude lineup, Haiku and Opus included, in our Claude API pricing guide. For how the $2 / $10 tier sits against other makers’ flagships, see Grok API pricing — or ChatGPT API pricing, where GPT-6.1 Sol carries the same $2 / $10 sticker.
Prices: Anthropic official API pricing, re-checked 2026-10-28; retirement date from the deprecations page. Prices and retirement dates change; check the official page before committing budget or a migration date.