Claude Sonnet 5.5 and Haiku 5.5: Pricing, Differences and Which One to Use

Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. Claude Haiku 5.5 starts at $0.10 and $0.50 for prompts up to 100,000 tokens, which makes it about 20 times cheaper on the rate card. Pick Sonnet 5.5 when output quality drives the result, and Haiku 5.5 for high-volume, bounded jobs. Prices are as of November 2026.

If you run training operations or administer an LMS, the useful question is not which model is better. It is what each model costs for the jobs you actually run: quiz generation, feedback analysis, a learner helpdesk. This post answers that with token arithmetic you can rerun with your own numbers.

Everything here is based on Anthropic’s published documentation. We did not run these models ourselves, so every token count in the cost model is a labeled assumption, not a measurement.

What are Claude Sonnet 5.5 and Claude Haiku 5.5, and when did they launch?

Anthropic released Claude Sonnet 5.5 on September 28, 2026, and Claude Haiku 5.5 on October 7, 2026. Sonnet is described as the best combination of speed and intelligence. Haiku is built for high-volume, latency-sensitive work such as classification, extraction and routing. Both are available through the Claude API and the major cloud platforms.

The two share most of their specifications, which is why the choice comes down to price and quality. Anthropic’s Claude Sonnet 5.5 overview and Claude Haiku 5.5 overview both list a 1M-token context window, 128K tokens of maximum output and a June 2026 knowledge cutoff.

Anthropic’s Sonnet 5.5 announcement says the model runs more than 30% faster than Sonnet 5 and costs up to 30% less per task, while keeping the same per-token rates. Its Haiku 5.5 announcement calls Haiku “the cheapest, fastest, and most capable small model we’ve ever released.” Those are vendor claims, not independent results.

If you are new to the family, start with our explainer on what is Claude.

How much do Claude Sonnet 5.5 and Haiku 5.5 cost per million tokens?

Sonnet 5.5 is a flat $2 input and $10 output per million tokens. Haiku 5.5 is $0.10 and $0.50 when the prompt is 100,000 tokens or fewer, and $0.50 and $2.50 above that. Batch requests are 50% off for both. The table below comes from Anthropic’s pricing page.

These are API list prices as of November 2026. They do not describe Claude app subscriptions; for seat plans, see our guide to Claude pricing.

What happens when a Haiku 5.5 prompt passes 100,000 tokens?

The whole request is billed at the higher Haiku rates, not just the tokens above 100,000. Anthropic counts every input token toward the threshold, including cache reads and cache writes, and prices each request on its own. Past the line, Haiku is still four times cheaper than Sonnet on both input and output.

Using Anthropic’s ratio of roughly 555,000 words per million tokens, 100,000 tokens is about 55,500 words. That is a long document set, but a bulk dump of learner comments or a full course library can cross it. The cost model below shows the effect.

What are the real differences between Claude Sonnet 5.5 and Haiku 5.5?

Price and intended workload are the main differences. Sonnet 5.5 defaults to high effort and fast latency. Haiku 5.5 defaults to medium effort and is labeled the fastest model. Anthropic says Sonnet and Opus remain better for complex agentic coding. For bounded, repeatable tasks, the published specifications are otherwise the same.

The quality gap is real but task dependent. In their launch posts, Anthropic reports 70.6% for Sonnet 5.5 and 39.2% for Haiku 5.5 on Terminal-Bench 4.0, an agentic coding test. That gap matters if you want an agent that writes and debugs code. It tells you little about drafting multiple-choice questions or tagging comments by theme.

Two practical points apply when you wire either model in. Anthropic says the same text counts as about 30% more tokens on Haiku 5.5 than on Haiku 4.5, so budgets built on older estimates run low. Sonnet 5.5 returns an error if you set temperature, top_p or top_k, so remove those from existing integrations.

How much does Claude Sonnet 5.5 cost for common L&D workloads?

For typical training workloads, Sonnet 5.5 costs roughly 16 to 20 times more than Haiku 5.5 per request. In the scenarios below, that is $0.094 against $0.0047 per quiz module, and $303 against $18 per month for a helpdesk bot. The token counts are our assumptions, so replace them with yours.

What assumptions sit behind the token counts?

Anthropic’s overview pages say 1M tokens is roughly 555,000 words on the current tokenizer, so one word is about 1.8 tokens. Every count below is built from that ratio plus a stated guess. They are not measurements.

• Quiz items from a transcript: a 60-minute module transcript of 9,000 words (about 16,200 tokens) plus 800 tokens of instructions gives 17,000 input tokens. Twenty items at 200 tokens each is 4,000 output tokens, plus 2,000 for thinking, so 6,000. Volume: 500 modules.
• Learner feedback summary: 2,000 comments of 60 words each, about 110 tokens per comment. Sent in 8 requests of 250 comments, that is 28,000 input tokens and 1,500 output tokens each.
• Internal helpdesk bot: a 20,000-token policy and FAQ knowledge base, 750 tokens of question and history, and 600 output tokens including thinking. Volume: 1,000 questions a day, or 30,000 a month.

The thinking allowance is deliberate. Anthropic notes that thinking tokens count toward the output limit, and Sonnet defaults to high effort, so real output may run higher than these figures.

What does quiz generation and feedback summarizing cost?

Each row shows the arithmetic in tokens times the rate per million tokens. Generating quiz items for 500 modules costs $47.00 on Sonnet and $2.35 on Haiku, or half that through the batch API, which suits work that does not need an instant answer. Summarizing 2,000 comments costs cents on either model.

The last row is the threshold trap. Sending all 2,000 comments in one request pushes the prompt to 220,500 tokens, so Haiku bills at the higher tier and the job costs about four times more than the chunked version. The chunked run omits a final step that merges eight partial summaries, which would add a little.

The honest takeaway is that these dollar amounts are small. For quiz generation, the SME time spent reviewing items costs far more than any token bill, so choose the model that produces items needing the fewest edits.

What does an internal helpdesk bot cost per month?

A bot that loads a 20,000-token knowledge base into every question gets expensive without prompt caching: $1,425 a month on Sonnet and $71.25 on Haiku. With caching, those fall to about $303 and $18. This is the one scenario where volume makes the model choice a real budget line.

The cache-write figures assume the cache is rewritten 12 times a day because quiet periods let it expire, at the 5-minute write rates in the pricing table. Both models require at least 512 tokens in a cacheable prompt, which a knowledge base clears easily.

At $18 a month versus $303, a helpdesk bot is a Haiku job unless your tests show it gives wrong answers on policy questions. Wrong answers about leave, compliance deadlines or certification rules cost more than the savings, so test those first.

Which Claude model should you use for each training workflow?

Use Haiku 5.5 for classification, tagging, extraction and routing, and for helpdesk answers grounded in a document you supply. Use Sonnet 5.5 for drafting that learners or auditors will read, multi-step reasoning, and anything you cannot easily check automatically. The table maps common jobs to a starting model.

This table reflects our judgment from the published specifications, not a head-to-head test. Anthropic’s Haiku announcement lists summaries, compactions, database queries and classification as the work Haiku handles reliably, which supports the first three rows.

Quiz and scenario work is where the choice is least obvious. If you are building these with an AI assistant, our notes on using Claude for instructional design cover the prompting side.

Should you start with Haiku 5.5 and escalate to Sonnet 5.5?

Yes, when failures are detectable by an automated check. Run Haiku first, validate the output, and send only the failures to Sonnet. The cost is the Haiku price plus the failure rate times the Sonnet price. It works for structured output and fails for subjective quality, which a script cannot judge.

For the quiz scenario, the formula is $0.0047 + (1 minus the pass rate) x $0.094 per module. The pass rates below are hypothetical, not measured.

A useful pass check for quiz items is mechanical: valid JSON, one correct key, four distinct options, and a stem answerable from the transcript. Such checks catch format failures, not weak distractors, so a human still reviews a sample. The cascade also adds latency, since failures take two calls.

Is Claude Opus 5.5 or Fable 5.1 worth considering instead?

Rarely, for the workloads above. Anthropic’s pricing page lists Opus 5.5 at $4 and $20 per million tokens and Fable 5.1 at $10 and $50. Anthropic’s Sonnet announcement says Opus 5.5 “remains clearly stronger at complex, open-ended work requiring sustained judgment,” which describes research and strategy more than routine training operations.

Sonnet 5.5 at $2 and $10 is half the Opus 5.5 rate. Consider stepping up only if reviewers keep rejecting Sonnet output on a task that needs deep judgment, such as analyzing a skills gap across several data sources. Our Claude Opus 5.5 guide covers that model, and Claude Opus 5.5 vs Fable 5.1 covers the top tier.

What should you do next?

Pick one workflow, not all of them. List the real inputs: transcript length, comments per month, questions per day. Replace the assumed token counts in the tables above and recompute with the current rates on Anthropic’s pricing page, which can change after November 2026.

Then run the 30-transcript test described in the cost-per-accepted-item tip above. If Haiku’s output needs about the same edits as Sonnet’s, ship Haiku and keep Sonnet as the fallback. If it does not, the extra spend buys less SME editing time.