Claude Sonnet 5.5 vs Haiku 5.5: 20x Price Gap, Benchmarks and Coding Fit

Sonnet 5.5 leads all eight rows of Anthropic's Haiku 5.5 table

Anthropic's Haiku 5.5 launch post (October 7, 2026) is the one page that runs both models on the same benchmark variants; it lists Sonnet 5.5 "for reference" beside Haiku 5.5, the legacy Haiku 4.5 and GPT-6 Luna. The figures below are Anthropic's, from that table only.

Read the gaps, not the leader: Haiku 5.5 is within about 6 to 7 points on FrontierCode and Humanity's Last Exam with tools, about 11 points behind on computer use and tool-free reasoning, and 31 points behind on Terminal-Bench 4.0. Anthropic draws the same line in words. It says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks" of the Terminal-Bench kind, and that Haiku 5.5 fits narrowly scoped jobs such as compaction, summarization and subagent work that "might otherwise have been cost-prohibitive". Its model guide gives Haiku 5.5 "the lowest latency and price" for real-time apps, high-volume processing and subagents, and Sonnet 5.5 everyday coding, agent and enterprise workloads. For Claude Code, Anthropic's help center suggests Sonnet for most coding and Haiku for quick mechanical work such as renames, log lines, regex explanations and boilerplate.

Sonnet 5.5's own launch post (September 28, 2026) reports different Sonnet 5.5 figures on some of these rows: 80.1% on an OSWorld 2.1 "partial" variant rather than 83.9% on the offline subset, 1844 on GDPval-AA and 1811 on AA-Briefcase. Those numbers belong to a different table, so they are not set against Haiku 5.5 here.

Haiku 5.5 costs one-twentieth of Sonnet 5.5 until a prompt passes 100,000 tokens

Anthropic prices Haiku 5.5 by prompt length: once a request's prompt is longer than 100,000 tokens, the whole request pays the higher tier. Sonnet 5.5 has no such step and includes its 1M-token window at standard pricing.

The bottom row is the decision row. Two details move it. First, cache reads: since October 7, 2026 a Sonnet 5.5 cache read costs 5% of its input price ($0.10, cut from $0.20), while Haiku 5.5 keeps the usual 10%, so on cached tokens Sonnet 5.5 costs 10x more on short prompts and only 2x more above 100,000 tokens. Second, the 100,000-token count includes cache reads and writes, so a cached 150,000-token context puts every request that reuses it in Haiku's higher tier. In ToolColumn arithmetic, an illustrative single 300,000-token prompt with 5,000 output tokens costs about $0.65 on Sonnet 5.5 and $0.16 on Haiku 5.5. The Batch API halves input and output rates on both models, so it does not change the ratio.

Against the legacy Haiku 4.5 ($1 / $5), Anthropic says Haiku 5.5 is priced 90% lower on prompts up to 100,000 tokens and 50% lower above. The same text counts as about 30% more tokens on Haiku 5.5 than on the legacy Haiku 4.5, so compare bills on real traffic rather than on per-token rates alone.

A million support tickets a month: about $230 on Haiku 5.5, $4,300 on Sonnet 5.5

An illustrative workload, in ToolColumn arithmetic at Claude API list prices: a support desk classifies 1,000,000 tickets a month. Each request reuses 3,000 tokens of cached instructions, adds 1,000 tokens of ticket text and returns 200 output tokens. Cache writes are left out because at that volume the cache stays warm.

The totals show caching narrowing the gap only slightly, from 20x to about 19x, because the fresh input and output still dominate. Without caching the same month costs $10,000 on Sonnet 5.5 and $500 on Haiku 5.5. The $4,070 difference is the monthly budget Haiku 5.5's misclassifications would have to eat before Sonnet 5.5 pays for itself on this job; measure both on a labeled sample of your own tickets before deciding.

Same 1M context and June 2026 cutoff; Sonnet 5.5 defaults to high effort on the API

Read the effort rows before benchmarking your own prompts: on the API the two models start at different effort levels, so a default-versus-default test compares Sonnet 5.5 at high with Haiku 5.5 at medium. Haiku 5.5 is the first Haiku-class model with an adjustable effort setting. Sonnet 5.5 rejects requests that set thinking to disabled; its between_tools setting only turns off up-front thinking. In the Claude apps, neither model can turn thinking off. Anthropic gives Sonnet 5.5 a Medium default in the apps and does not state Haiku 5.5's app default. Both can return up to 300K output tokens through the Batches API with the output-300k-2026-03-24 beta header.

A prompt routed to both models needs to avoid what either rejects: Sonnet 5.5 returns a 400 error for temperature, top_p or top_k and errors on forced tool use, while Haiku 5.5 errors on assistant prefill and on manual budget_tokens, and its sampling parameters can return an error.

Both are on Claude Free; Claude Code's haiku alias needs v2.1.293

• Claude apps: Free, Pro, Max, Team and Enterprise users can select both Sonnet 5.5 and Haiku 5.5 on web, iOS and Android. Opus 5.5 is not on Free. Anthropic does not publish which model each plan opens on by default, so check the model picker.
• Claude Code: the default model is Opus 5.5 on Pro, Max, Team, Enterprise and Anthropic API accounts, so you pick either model yourself. On the Anthropic API the sonnet alias selects Sonnet 5.5 (Claude Code v2.1.284 or later) and haiku selects Haiku 5.5 (v2.1.293 or later). On Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and Claude Platform on AWS, haiku still maps to the legacy Haiku 4.5 and sonnet to an earlier Sonnet, so use the full model ID there.
• API and cloud: the IDs are claude-sonnet-5-5 and claude-haiku-5-5 on the Claude API, Google Cloud (Google managed and Anthropic managed), Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-sonnet-5-5 and anthropic.claude-haiku-5-5 on Amazon Bedrock. Anthropic's rates above are for the Claude API.