[{"data":1,"prerenderedAt":1026},["ShallowReactive",2],{"blog-claude-haiku-5-5-en":3},{"slug":4,"published":5,"minutes":6,"category":7,"tags":8,"tagPages":14,"keywords":15,"about":25,"sources":32,"cover":51,"og":52,"expertise":53,"locales":54,"lang":55,"title":58,"description":59,"coverAlt":60,"metaTitle":61,"takeaways":62,"faq":68,"toc":84,"blocks":112,"others":795},"claude-haiku-5-5","2026-10-19",15,"llmops",[9,10,11,12,13],"Claude Haiku 5.5","LLM pricing","Reasoning effort","Model routing","Subagents",[],[16,17,4,18,19,20,21,22,23,24],"claude haiku 5.5","haiku 5.5 pricing","haiku 5.5 vs haiku 4.5","haiku 5.5 vs sonnet 5.5","haiku 5.5 benchmarks","haiku 5.5 effort","claude subagent model","cheapest claude model","haiku 5.5 api",[26,29],{"name":27,"url":28},"Claude (AI)","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FClaude_(AI)",{"name":30,"url":31},"Anthropic","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FAnthropic",[33,36,39,42,45,48],{"title":34,"url":35},"Anthropic: Introducing Claude Haiku 5.5 (October 7, 2026)","https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-haiku-5-5",{"title":37,"url":38},"Claude Platform docs: Claude Haiku 5.5 overview","https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fmodels\u002Fhaiku-5-5\u002Foverview",{"title":40,"url":41},"Claude Platform docs: Pricing","https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fabout-claude\u002Fpricing",{"title":43,"url":44},"Claude Platform docs: Effort","https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fbuild-with-claude\u002Feffort",{"title":46,"url":47},"Anthropic: Introducing Claude Sonnet 5.5 (September 28, 2026)","https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-sonnet-5-5",{"title":49,"url":50},"Claude Platform docs: Models overview","https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fabout-claude\u002Fmodels\u002Foverview","\u002Fimages\u002Fblog\u002Fclaude-haiku-5-5\u002Fcover.webp","\u002Fimages\u002Fblog\u002Fclaude-haiku-5-5\u002Fog.jpg","ai-engineer",[55,56,57],"en","de","hu","Claude Haiku 5.5: pricing, effort levels and the jobs it takes over from Sonnet","Claude Haiku 5.5 pricing by tier, benchmarks against Haiku 4.5 and Sonnet 5.5, effort levels, and which Sonnet 5.5 jobs can move to it.","Claude Haiku 5.5: a light paper airplane carrying a small coin flies quickly past a large, heavy jar.","Claude Haiku 5.5 pricing and effort · Balázs Csorba",[63,64,65,66,67],"Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, as of October 2026. Prompts over that size pay $0.50 and $2.50.","Anthropic’s table puts Haiku 5.5 far ahead of Haiku 4.5 and behind Sonnet 5.5. The widest gap to Sonnet is on Terminal-Bench 4.0, which measures agentic coding.","Effort defaults to medium. Use low for simple, high-volume calls, and raise the level only where your evals show a gain.","Move classification, extraction, routing, summaries and subagent steps first. Keep complex agentic coding and judgment-heavy work on Sonnet 5.5 or Opus 5.5.","Re-test a move from Haiku 4.5 as a whole request: token counts rise by around 30% for the same text, sampling parameters are removed, and thinking is on by default.",[69,72,75,78,81],{"q":70,"a":71},"How much does Claude Haiku 5.5 cost?","As of October 2026, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 above that. The Batch API halves those prices, and a cache read costs $0.01 per million tokens in the lower tier.",{"q":73,"a":74},"Is Claude Haiku 5.5 better than Haiku 4.5?","Anthropic’s launch table shows large gains over Haiku 4.5 on the benchmarks it lists, and says the model costs around 75% less to run on average. Those are vendor figures, so test your own workload before you switch.",{"q":76,"a":77},"What do the Haiku 5.5 effort levels do?","Effort sets how much work the model puts into each response, including text, tool calls and thinking. Haiku 5.5 supports low, medium, high, xhigh and max, with medium as the default. Use low for simple, high-volume calls, and raise the level only where your evals show a quality gain.",{"q":79,"a":80},"Can Claude Haiku 5.5 replace Sonnet 5.5?","Only for narrower jobs. Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding, and that Haiku 5.5 suits narrowly scoped work such as compaction, summarization and subagent tasks. Test each workload on your own evals first.",{"q":82,"a":83},"What changes when I migrate from Haiku 4.5?","The newer tokenizer makes the same text count as roughly 30% more tokens, so recount with the 5.5 model ID. Remove temperature, top_p and top_k, and choose an effort level, because thinking is on by default and effort is a new control.",[85,88,91,94,97,100,103,106,109],{"id":86,"title":87},"what-haiku-55-is","What is Claude Haiku 5.5, and where does it run?",{"id":89,"title":90},"pricing-tiers-and-batch","What does Haiku 5.5 cost per million tokens?",{"id":92,"title":93},"benchmarks-against-haiku-45-and-sonnet-55","How does Haiku 5.5 benchmark against Haiku 4.5 and Sonnet 5.5?",{"id":95,"title":96},"effort-levels-on-haiku-55","How do effort levels work on Haiku 5.5?",{"id":98,"title":99},"jobs-that-move-from-sonnet","Which Sonnet 5.5 jobs can move to Haiku 5.5?",{"id":101,"title":102},"router-and-subagent-pattern","How should Haiku 5.5 sit next to Sonnet 5.5 and Opus 5.5?",{"id":104,"title":105},"migrating-from-haiku-45","What should you re-test when you move from Haiku 4.5?",{"id":107,"title":108},"practical-checklist","What is a practical checklist before you switch a workload?",{"id":110,"title":111},"sources","Sources",[113,121,124,127,142,176,179,182,185,186,189,271,282,285,288,332,340,347,348,351,447,450,459,462,463,473,480,482,497,503,527,541,548,549,552,555,640,643,644,647,651,654,666,667,670,673,683,693,696,697,700,757,774,775],{"type":114,"content":115},"paragraph",[116,117,120],"Claude Haiku 5.5 is Anthropic’s small model for high-volume work. It was released on October 7, 2026 under the API model ID ",{"tag":118,"children":119},"code",[4],", and Anthropic presents it as “the cheapest, fastest, and most capable small model we’ve ever released.” It has a 1M-token context window, up to 128K output tokens, adaptive thinking on by default, and an effort setting, which makes it the first Haiku-class model with one. The short answer for most teams is narrow: use it for classification, extraction, routing, summaries and subagent steps, and keep complex agentic coding on Sonnet 5.5 or Opus 5.5.",{"type":114,"content":122},[123],"This article covers the price tiers and what they mean for each request, the benchmark table against Haiku 4.5 and Sonnet 5.5, how effort works on this model, which Sonnet 5.5 jobs can move to it, and what to re-test when you migrate from Haiku 4.5. I have spent more than ten years on Vue, Nuxt, TypeScript, Node and PHP work, and I run LLM features on my own Nuxt site, so the question I bring to a launch like this is practical: which of my workloads could move to a cheaper model without a measurable quality loss? Every price and benchmark figure comes from Anthropic’s launch posts or its platform documentation, and prices are as of October 2026.",{"type":125,"level":126,"id":86,"text":87},"heading",2,{"type":114,"content":128},[129,130,132,133,136,137,141],"Haiku 5.5 takes text and images as input and returns text. It is available now on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. The API ID is ",{"tag":118,"children":131},[4],"; on Amazon Bedrock it is ",{"tag":118,"children":134},[135],"anthropic.claude-haiku-5-5",". The ",{"tag":138,"href":50,"children":139},"a",[140],"platform’s model table"," lists it for high-volume, latency-sensitive tasks such as classification, extraction and routing.",{"type":143,"ordered":144,"items":145},"list",false,[146,152,161,166,171],[147,151],{"tag":148,"children":149},"strong",[150],"Context and output:"," 1M tokens of context, up to 128K output tokens, and up to 300K output tokens on the Message Batches API under a beta header.",[153,156,157,160],{"tag":148,"children":154},[155],"Thinking and effort:"," adaptive thinking on by default, with effort as the control for how much it thinks. The default effort is ",{"tag":118,"children":158},[159],"medium",".",[162,165],{"tag":148,"children":163},[164],"Speed:"," Anthropic calls it its fastest model to date at each model’s standard speed. In Fast Mode it runs less quickly than the Opus models.",[167,170],{"tag":148,"children":168},[169],"Knowledge:"," the reliable knowledge cutoff and the training data cutoff are both June 2026.",[172,175],{"tag":148,"children":173},[174],"Lifecycle:"," released October 7, 2026, with retirement not sooner than October 7, 2027.",{"type":114,"content":177},[178],"Anthropic positions the model for work where volume and speed count for more than depth. The launch post names summaries, compactions, database queries and classification requests as quick, repetitive work it handles reliably, and says it pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. It also names live customer support and browser use as speed-sensitive tasks where it fits well.",{"type":114,"content":180},[181],"Two details need attention before production. The 1M window is priced differently from the rest of the 4.6-and-later family: those models include the full window at standard pricing, while prompts over 100,000 tokens pay higher rates on Haiku 5.5. And its cybersecurity safeguards are stricter than on Haiku 4.5 but looser than on other recent models; they still block penetration testing and similar techniques. Anthropic says organizations doing wider cyber work can apply to its Cyber Verification Program. Plan for that if you build security tooling.",{"type":114,"content":183},[184],"The timing was signposted. In its Sonnet 5.5 post on September 28, 2026, Anthropic said Haiku 5.5 would join the Claude 5.5 family “in the coming weeks.” It shipped nine days later, on October 7.",{"type":125,"level":126,"id":89,"text":90},{"type":114,"content":187},[188],"As of October 2026, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Prompts over 100,000 tokens pay $0.50 for input and $2.50 for output. The tier is set per request, by the prompt length of that request.",{"type":190,"head":191,"rows":202},"table",[192,194,196,198,200],[193],"Per 1M tokens",[195],"Haiku 5.5, up to 100K \u002F over 100K",[197],"Haiku 4.5",[199],"Sonnet 5.5",[201],"Opus 5.5",[203,214,225,235,244,254,263],[204,206,208,210,212],[205],"Input",[207],"$0.10 \u002F $0.50",[209],"$1.00",[211],"$2.00",[213],"$4.00",[215,217,219,221,223],[216],"Output",[218],"$0.50 \u002F $2.50",[220],"$5.00",[222],"$10.00",[224],"$20.00",[226,228,230,232,234],[227],"5-minute cache write",[229],"$0.125 \u002F $0.625",[231],"$1.25",[233],"$2.50",[220],[236,238,240,241,242],[237],"1-hour cache write",[239],"$0.20 \u002F $1.00",[211],[213],[243],"$8.00",[245,247,249,251,252],[246],"Cache read",[248],"$0.01 \u002F $0.05",[250],"$0.10",[250],[253],"$0.20",[255,257,259,261,262],[256],"Batch input",[258],"$0.05 \u002F $0.25",[260],"$0.50",[209],[211],[264,266,268,269,270],[265],"Batch output",[267],"$0.25 \u002F $1.25",[233],[220],[222],{"type":114,"content":272},[273,274,277,278,281],"Three details change the bill more than the headline rate. First, the prompt length that sets the tier counts every input token, including cache reads and cache writes, so caching a long prompt does not move the request into the lower tier. Second, the Batch API halves input and output prices, which suits backlogs that do not need an answer in seconds. Third, setting ",{"tag":118,"children":275},[276],"inference_geo"," to ",{"tag":118,"children":279},[280],"\"us\""," adds a 1.1x multiplier to every token category, and on Haiku 5.5 that multiplier also applies to the over-100K prices.",{"type":114,"content":283},[284],"Cache reads cost 10% of the base input price on Haiku 5.5: $0.01 per million tokens in the lower tier, a tenth of the $0.10 that Sonnet 5.5 charges for a cache read. Anthropic also halved the Sonnet 5.5 cache-read price on the day of the launch, from $0.20 to $0.10 per million tokens, and says that cuts Sonnet 5.5 cost on most agentic tasks by around 20%. The same day it announced, with rollout that week, a monthly API credit for Max and Team subscribers: $100 a month for Max 5x, $200 for Max 20x, and up to $500 for a Team, pooled across its users. The credit is meant for building tools and agents that call the API, and it works on any model.",{"type":114,"content":286},[287],"A worked example makes the tiers concrete. Assume one million classification calls, each with a 1,500-token prompt and a 50-token answer. That is 1.5 billion input tokens and 50 million output tokens. The table prices that at the lower tier, with and without batch pricing and caching, and on Sonnet 5.5 for comparison.",{"type":190,"head":289,"rows":296},[290,292,293,294],[291],"1M calls, each 1,500 input + 50 output tokens",[205],[216],[295],"Total",[297,306,314,323],[298,300,302,304],[299],"Haiku 5.5, standard",[301],"$150.00",[303],"$25.00",[305],"$175.00",[307,309,311,312],[308],"Haiku 5.5, 1,200 input tokens cached",[310],"$42.00",[303],[313],"$67.00",[315,317,319,321],[316],"Haiku 5.5, batch",[318],"$75.00",[320],"$12.50",[322],"$87.50",[324,326,328,330],[325],"Sonnet 5.5, standard",[327],"$3,000.00",[329],"$500.00",[331],"$3,500.00",{"type":114,"content":333},[334,335,277,337,339],"The cached row is $12 for the 1.2 billion cached tokens at $0.01 per million, plus $30 for the remaining 300 tokens per call at $0.10 per million. It leaves out the one-time cache write and assumes the prefix stays warm; 1,200 tokens clears the 512-token minimum for caching. On the standard row, setting ",{"tag":118,"children":336},[276],{"tag":118,"children":338},[280]," raises the $175 total to $192.50.",{"type":341,"variant":342,"title":343,"body":344},"callout","warn","The 100,000-token line applies per request",[345],[346],"A request with a prompt over 100,000 tokens pays the higher Haiku 5.5 rates, even when part of that prompt is a cache hit. A long system prompt or retrieved context can push a workload over the line without anyone noticing. Measure the prompt-length distribution, trim where quality holds, and treat the tier as part of the design rather than a footnote.",{"type":125,"level":126,"id":92,"text":93},{"type":114,"content":349},[350],"On Anthropic’s table, Haiku 5.5 beats Haiku 4.5 by a wide margin wherever Haiku 4.5 has a score, trails Sonnet 5.5 on every row, and beats GPT-6 Luna on each row where Luna has a score. These are vendor-reported numbers. Read them as direction and rough size, not as a verdict on your workload.",{"type":190,"head":352,"rows":361},[353,355,357,358,360],[354],"Benchmark (Anthropic’s label)",[356],"Haiku 5.5",[197],[359],"GPT-6 Luna",[199],[362,373,384,395,406,416,427,437],[363,365,367,369,371],[364],"GDPval-AA v2.1 (knowledge work)",[366],"1,620",[368],"735",[370],"1,437",[372],"1,840",[374,376,378,380,382],[375],"AA-Briefcase v1.1 (knowledge work)",[377],"1,578",[379],"614",[381],"1,336",[383],"1,824",[385,387,389,391,393],[386],"OSWorld 2.1, offline subset (computer use)",[388],"72.4%",[390],"15.7%",[392],"48.9%",[394],"83.9%",[396,398,400,402,404],[397],"Humanity’s Last Exam, no tools",[399],"45.9%",[401],"10.2%",[403],"—",[405],"56.9%",[407,409,411,413,414],[408],"Humanity’s Last Exam, with tools",[410],"57.4%",[412],"18.7%",[403],[415],"64.5%",[417,419,421,423,425],[418],"Terminal-Bench 4.0 (agentic coding)",[420],"39.2%",[422],"0.0%",[424],"16.4%",[426],"70.6%",[428,430,432,433,435],[429],"FrontierCode 1.1, main set (agentic coding)",[431],"46.4%",[403],[434],"42.4%",[436],"52.1% (xhigh)",[438,440,441,443,445],[439],"Chartography, no tools (visual reasoning)",[431],[442],"6.4%",[444],"29.1%",[446],"61.6%",{"type":114,"content":448},[449],"Anthropic’s table is the source for every Haiku figure here. The Sonnet 5.5 column needs a caveat. Anthropic’s Sonnet 5.5 post gives 1,844 on GDPval-AA, 1,811 on AA-Briefcase and 80.1% on OSWorld 2.1, labeled “partial”, where the Haiku post gives 1,840, 1,824 and 83.9% on the offline subset. The sources do not explain the difference, so this article uses the Haiku post throughout and does not average the two. GPT-6 Luna is also a different model from the GPT-6 Sol in the Sonnet post.",{"type":114,"content":451},[452,453,458],"The widest gap to Sonnet 5.5 is on Terminal-Bench 4.0, an agentic coding evaluation, at 39.2% against 70.6%. That fits Anthropic’s own statement that Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding. The narrowest is FrontierCode, at 46.4% against 52.1%. The biggest step from Haiku 4.5 is on OSWorld 2.1, from 15.7% to 72.4%. For third-party numbers on Opus 5.5, see my ",{"tag":454,"to":455,"children":456},"link","\u002Fblog\u002Fartificial-analysis-leaderboard-claude-opus-5-5",[457],"Artificial Analysis post on Claude Opus 5.5","; that is a separate question from the vendor table.",{"type":114,"content":460},[461],"Anthropic’s post also quotes customers, and those are vendor-collected results, not independent tests. AlphaSense reports a statistically significant improvement on 400 queries against its Ask in Document feature: 0.84 for Haiku 5.5 against 0.76 for Haiku 4.5. Box reports a score 11 points higher than Haiku 4.5 at about half the latency in early testing. HubSpot reports 92.8% averaged over three runs on a CRM reporting suite, which it calls the best score it has seen on that suite.",{"type":125,"level":126,"id":95,"text":96},{"type":114,"content":464},[465,466,468,469,472],"Effort sets how much work the model puts into the whole response: text, tool calls and thinking. Haiku 5.5 supports low, medium, high, xhigh and max, and medium is the default, so sending ",{"tag":118,"children":467},[159]," behaves the same as omitting the parameter. Effort is a behavioral signal rather than a strict token budget, and ",{"tag":118,"children":470},[471],"max_tokens"," remains the hard limit on total output.",{"type":114,"content":474},[475,476,479],"Set it under ",{"tag":118,"children":477},[478],"output_config",". The minimal call below uses the shape from the effort guide, with the Haiku model ID:",{"type":118,"code":481},"import anthropic\n\nclient = anthropic.Anthropic()\n\nresponse = client.messages.create(\n    model=\"claude-haiku-5-5\",\n    max_tokens=2048,\n    messages=[\n        {\n            \"role\": \"user\",\n            \"content\": \"Label this support ticket as billing, bug, or other. Ticket: ...\",\n        }\n    ],\n    output_config={\"effort\": \"low\"},\n)\n\nfor block in response.content:\n    if block.type == \"text\":\n        print(block.text)",{"type":114,"content":483},[484,485,488,489,492,493,496],"The shape is the same one the effort guide uses for other models. Leave ",{"tag":118,"children":486},[487],"temperature",", ",{"tag":118,"children":490},[491],"top_p"," and ",{"tag":118,"children":494},[495],"top_k"," out, as the overview advises.",{"type":498,"src":499,"alt":500,"wide":501,"caption":502},"image","\u002Fimages\u002Fblog\u002Fclaude-haiku-5-5\u002Fdiagram-28.webp","Five effort levels of Claude Haiku 5.5, from low to max, with medium as the default setting",true,"Haiku 5.5 effort levels: low is cheapest, medium is the default, and xhigh and max need a measured quality gain.",{"type":143,"ordered":144,"items":504},[505,510,514,519],[506,509],{"tag":148,"children":507},[508],"low"," is the cheapest and fastest level. Anthropic suggests it for chat, short tool tasks and simple, high-volume requests. It also warns that in long agent prompts, the model is more likely to skip a search, stop early or skip a check at this level.",[511,513],{"tag":148,"children":512},[159]," is the default and the recommended start for most work, including agentic coding.",[515,518],{"tag":148,"children":516},[517],"high"," suits knowledge work, longer agent tasks and strict instruction following.",[520,492,523,526],{"tag":148,"children":521},[522],"xhigh",{"tag":148,"children":524},[525],"max"," are for workloads where your evals show a quality gain. Anthropic says to compare them with Sonnet 5.5 on performance, cost and speed before you pay for them.",{"type":114,"content":528},[529,530,532,533,536,537,540],"Thinking is on by default and counts toward ",{"tag":118,"children":531},[471],", so leave headroom in that limit. To reduce thinking, lower the effort level. You can also send ",{"tag":118,"children":534},[535],"thinking: {\"type\": \"disabled\"}"," at high effort or below. At xhigh or max, that request returns a 400 error, so use adaptive thinking there instead. Do not pass ",{"tag":118,"children":538},[539],"adaptive"," as an effort value, because it is a thinking mode, not a level.",{"type":114,"content":542},[543,544,547],"On the Claude API and Google Cloud, Haiku 5.5 also accepts a per-message effort change in beta, with the header ",{"tag":118,"children":545},[546],"mid-conversation-output-config-2026-07-01",", and that preserves the prompt cache. It does not work while thinking is disabled. Anthropic plots three benchmarks against effort in the launch post. Use those charts to see the shape of the trade-off, then run your own sweep: the guidance for Opus 5.5 and Sonnet 5.5 is to sweep on your own evals rather than carry settings over from an earlier model.",{"type":125,"level":126,"id":98,"text":99},{"type":114,"content":550},[551],"Move the narrow, high-volume jobs first: classification, extraction, routing, summaries and compaction, lookups and subagent steps. Keep complex agentic coding, bug fixing and document or slide work on Sonnet 5.5, and reserve Opus 5.5 for the hardest reasoning.",{"type":114,"content":553},[554],"The reason to test the move is price. At the lower tier, Sonnet 5.5 costs 20 times as much per token as Haiku 5.5, at $2 against $0.10 for input and $10 against $0.50 for output. Above 100,000 prompt tokens, Sonnet 5.5 keeps its standard rate, so the ratio narrows to a quarter: $0.50 against $2.00 for input, and $2.50 against $10 for output.",{"type":190,"head":556,"rows":563},[557,559,560,561],[558],"Task",[356],[199],[562],"Basis in the sources",[564,573,581,589,597,605,614,623,631],[565,567,569,571],[566],"Classification and routing",[568],"Start here",[570],"Only where Haiku misses on your evals",[572],"Overview names classification and routing",[574,576,577,579],[575],"Extraction from one or a few documents",[568],[578],"Where accuracy gaps show up",[580],"Overview names extraction; AlphaSense tested Haiku 5.5 on Ask in Document",[582,584,585,587],[583],"Summaries and compaction",[568],[586],"Where key facts get lost",[588],"Launch post names summaries and compaction for Haiku 5.5",[590,592,593,595],[591],"Database queries and quick lookups",[568],[594],"Multi-step analysis",[596],"Launch post; Rogo’s “quick lookups”",[598,600,601,603],[599],"Subagent steps in a larger run",[568],[602],"Planning and final judgment",[604],"Launch post pairs it with Opus 5.5 and Sonnet 5.5 for subagent work",[606,608,610,612],[607],"Live support and browser use",[609],"Start here for speed",[611],"Long multi-step runs",[613],"Launch post; OSWorld 2.1: 72.4% vs 83.9%",[615,617,619,621],[616],"Agentic coding across a codebase",[618],"Subagent role only",[620],"Lead role",[622],"Terminal-Bench 4.0: 39.2% vs 70.6%",[624,626,628,629],[625],"Well-scoped bug fixes and polished documents",[627],"Test first",[568],[630],"Sonnet 5.5 post: strongest at these tasks",[632,634,636,638],[633],"Complex judgment and long-horizon work",[635],"Not as the lead",[637],"Start here, or Opus 5.5 for the hardest",[639],"Sonnet 5.5 post: Opus 5.5 is built for careful judgment",{"type":114,"content":641},[642],"The table is a starting point. None of the sources reports results for these exact workloads, so run each one on both models before you move it. Anthropic draws the line itself: Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding, while Haiku 5.5 suits narrowly scoped work such as compaction, summarization and subagent tasks. Box says it would put Haiku 5.5 to work on analytical jobs at scale, such as cost reports, financial summaries and weekly recurring reviews.",{"type":125,"level":126,"id":101,"text":102},{"type":114,"content":645},[646],"Put Haiku 5.5 in front as the first hop, either as a router that sorts requests or as a subagent that handles narrow steps, and escalate only the requests that need more capability. The launch post shows both roles in customer examples. Cognition says Haiku 5.5 joins Devin Fusion as the sidekick, and that you can try it in the Devin CLI with Opus 5.5 as the lead. Rogo describes a Haiku 5.5 subagent that pulls the segment revenue line out of a 10-K while a bigger model builds the deck.",{"type":498,"src":648,"alt":649,"wide":501,"caption":650},"\u002Fimages\u002Fblog\u002Fclaude-haiku-5-5\u002Fdiagram-39.webp","Routing diagram with Claude Haiku 5.5 as the first hop, escalating harder work to Sonnet 5.5 and Opus 5.5","Router pattern: Haiku 5.5 sorts each request, and only the requests that need more capability move up to Sonnet 5.5 or Opus 5.5.",{"type":114,"content":652},[653],"A router needs a signal to escalate on, and the sources do not define one. Choose it from your evals: prompt length, with the 100,000-token tier in mind, a validation check on Haiku’s output, or the task type. Whatever you choose, log which path each request takes. The bill depends on the escalation rate as much as on the cheapest route.",{"type":114,"content":655},[656,657,661,662,160],"Long prompts deserve a rule of their own. Above 100,000 tokens, Haiku 5.5 is still a quarter of Sonnet 5.5’s price, so a long prompt sent to Sonnet should be justified by quality data, not by the tier alone. For the wider cost picture, see ",{"tag":454,"to":658,"children":659},"\u002Fblog\u002Fllm-cost-latency-prompt-caching-routing",[660],"prompt caching and model routing",", and for routing decisions with calibrated confidence and a human hand-off, see ",{"tag":454,"to":663,"children":664},"\u002Fblog\u002Fjev-typed-decisions-llm-routing",[665],"typed decisions for LLM routing",{"type":125,"level":126,"id":104,"text":105},{"type":114,"content":668},[669],"Re-test the whole request, not only the model name. Token counts rise, sampling parameters go, thinking is on by default, and effort is a new control you have to set. Start with a token count on real prompts, then run an effort sweep on your evals.",{"type":114,"content":671},[672],"Token counts go up. Haiku 5.5 uses the newer tokenizer that Claude 4.7 and later models share, and the overview says the same text counts as approximately 30% more tokens than on Haiku 4.5. The launch post’s reduction figures already account for that effect, according to its footnote. Your own budget does not, so count tokens with the 5.5 model ID on real prompts and budget from those counts.",{"type":114,"content":674},[675,676,488,678,492,680,682],"Sampling parameters go too. The overview says to omit ",{"tag":118,"children":677},[487],{"tag":118,"children":679},[491],{"tag":118,"children":681},[495],", and points to a section of the migration guide for the values that return a 400 error. Thinking changes as well. Adaptive thinking is on by default, and thinking blocks work only in the account that produced them or in an account linked to it, which matters if you move conversations between accounts.",{"type":114,"content":684},[685,686,492,689,692],"Read capabilities from the Models API rather than from notes. Each model’s capabilities object reports whether it accepts ",{"tag":118,"children":687},[688],"thinking: {type: \"disabled\"}",{"tag":118,"children":690},[691],"between_tools",", and which server tools it supports, so a migration script can check those flags before it sends a request.",{"type":114,"content":694},[695],"Then sweep effort. Start at medium for most work, measure low on the simple, high-volume paths, and measure high where long agent runs or strict instruction following matter. Anthropic warns that at low effort, long agent prompts make the model more likely to skip a search, stop early or skip a check, so add eval cases that test exactly those skips. Read the Prompting Claude Haiku 5.5 guide that the overview links to before you rewrite prompts, since it covers behavior specific to this model. The sources in this article list no other changes, so treat anything else as something your evals have to find.",{"type":125,"level":126,"id":107,"text":108},{"type":114,"content":698},[699],"Run these steps in order for each workload, and keep the Sonnet 5.5 path as the fallback until your evals say otherwise.",{"type":143,"ordered":501,"items":701},[702,704,723,738,743,753,755],[703],"Split traffic by prompt length, and count the requests above 100,000 tokens, because they pay the higher tier.",[705,706,708,709,712,713,716,717,712,720,160],"Count tokens with the ",{"tag":118,"children":707},[4]," ID on real prompts. Add the tool-use system prompt, which the pricing page lists at 286 tokens for ",{"tag":118,"children":710},[711],"auto"," or ",{"tag":118,"children":714},[715],"none"," tool choice and 406 tokens for ",{"tag":118,"children":718},[719],"any",{"tag":118,"children":721},[722],"tool",[724,725,488,727,492,729,731,732,735,736,160],"Remove ",{"tag":118,"children":726},[487],{"tag":118,"children":728},[491],{"tag":118,"children":730},[495]," from every request, and set ",{"tag":118,"children":733},[734],"output_config.effort"," explicitly, starting at ",{"tag":118,"children":737},[159],[739,740,742],"Leave room in ",{"tag":118,"children":741},[471]," for thinking, and keep thinking on unless the sweep shows it adds nothing for that workload.",[744,745,747,748,712,750,752],"Sweep effort on one eval set. Use ",{"tag":118,"children":746},[508]," only where the skip checks still pass, and move to ",{"tag":118,"children":749},[522],{"tag":118,"children":751},[525]," only where the score rises.",[754],"Send non-urgent work through the Batch API at half the price, and cache shared prefixes of at least 512 tokens.",[756],"Log the route, the escalation reason, the token counts and the eval score for each request, so you can see the escalation rate and the real cost per path.",{"type":114,"content":758},[759,760,764,765,769,770,160],"If you are building the routing and evaluation layer around models like this, two related posts on this site go further: ",{"tag":454,"to":761,"children":762},"\u002Fblog\u002Freasoning-models-effort-thinking-budget",[763],"reasoning effort and thinking budgets"," covers effort sweeps in depth, and ",{"tag":454,"to":766,"children":767},"\u002Fblog\u002Fsmall-language-models-when-enough",[768],"small language models"," covers when a small model is enough. For help putting LLM features like these into production, see my ",{"tag":454,"to":771,"children":772},"\u002Fexpertise\u002Fai-engineer",[773],"AI engineer work",{"type":125,"level":126,"id":110,"text":111},{"type":143,"ordered":501,"items":776},[777,780,783,786,789,792],[778],{"tag":138,"href":35,"children":779},[34],[781],{"tag":138,"href":38,"children":782},[37],[784],{"tag":138,"href":41,"children":785},[40],[787],{"tag":138,"href":44,"children":788},[43],[790],{"tag":138,"href":47,"children":791},[46],[793],{"tag":138,"href":50,"children":794},[49],[796,870,917,956],{"slug":797,"published":798,"minutes":799,"category":7,"tags":800,"tagPages":805,"keywords":806,"about":817,"sources":827,"cover":864,"og":865,"expertise":53,"locales":866,"lang":55,"title":867,"description":868,"coverAlt":869},"small-language-models-when-enough","2026-10-17",11,[801,802,803,12,804],"Small language models","Distillation","Quantization","LLMOps",[],[768,807,808,809,810,811,812,813,814,815,816],"when to use a small language model","SLM vs LLM cost","model distillation","quantization aware training","Q4_K_M","LLM routing","small model evaluation","on-device LLM","Gemma 4 E4B","local LLM privacy",[818,821,824],{"name":819,"url":820},"Small language model","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FSmall_language_model",{"name":822,"url":823},"Knowledge distillation","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FKnowledge_distillation",{"name":825,"url":826},"Large language model","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FLarge_language_model",[828,831,834,837,840,843,846,849,852,855,858,861],{"title":829,"url":830},"Belcak et al.: Small Language Models are the Future of Agentic AI (arXiv 2506.02153)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.02153",{"title":832,"url":833},"Google: Gemma 4 model card","https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fdocs\u002Fcore\u002Fmodel_card_4",{"title":835,"url":836},"Google: Gemma 3 QAT models, state-of-the-art AI on consumer GPUs","https:\u002F\u002Fdevelopers.googleblog.com\u002Fen\u002Fgemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus\u002F",{"title":838,"url":839},"Google: Introducing Gemma 3 270M","https:\u002F\u002Fdevelopers.googleblog.com\u002Fen\u002Fintroducing-gemma-3-270m\u002F",{"title":841,"url":842},"Microsoft Research: Phi-4 Technical Report (arXiv 2412.08905)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2412.08905",{"title":844,"url":845},"Qwen: Qwen3 Technical Report (arXiv 2505.09388)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.09388",{"title":847,"url":848},"Apple: Apple Intelligence Foundation Language Models, Tech Report 2025 (arXiv 2507.13575)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.13575",{"title":850,"url":851},"Ong et al.: RouteLLM, Learning to Route LLMs with Preference Data (arXiv 2406.18665)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.18665",{"title":853,"url":854},"LMSYS: RouteLLM, an open-source framework for cost-effective LLM routing","https:\u002F\u002Flmsys.org\u002Fblog\u002F2024-07-01-routellm\u002F",{"title":856,"url":857},"Gomes et al.: On the Generalization vs Fidelity Paradox in Knowledge Distillation (arXiv 2505.15442)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.15442",{"title":859,"url":860},"llama.cpp: quantize tool documentation","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fblob\u002Fmaster\u002Ftools\u002Fquantize\u002FREADME.md",{"title":862,"url":863},"European Commission: AI Act","https:\u002F\u002Fdigital-strategy.ec.europa.eu\u002Fen\u002Fpolicies\u002Fregulatory-framework-ai","\u002Fimages\u002Fblog\u002Fsmall-language-models-when-enough\u002Fcover.webp","\u002Fimages\u002Fblog\u002Fsmall-language-models-when-enough\u002Fog.jpg",[55,56,57],"Small language models: when a 4B model is enough","When a small model beats a frontier API: cost per resolved case, latency, privacy, distillation, quantization, routing and an eval plan for your traffic.","Small language models: a hand picks a pocket knife over a large toolbox to cut a parcel string.",{"slug":871,"published":872,"updated":873,"minutes":874,"category":7,"tags":875,"tagPages":880,"keywords":881,"about":889,"sources":893,"cover":911,"og":912,"expertise":53,"locales":913,"lang":55,"title":914,"description":915,"coverAlt":916},"openrouter","2026-07-23","2026-08-01",10,[876,12,877,878,879],"LLM gateway","Fallbacks","OpenAI-compatible","Pay per token",[],[871,882,883,884,885,886,887,888],"openrouter vs litellm","llm gateway comparison","openrouter pricing","openai compatible api gateway","llm fallback routing","multi model api gateway","byok llm routing",[890],{"name":891,"url":892},"OpenRouter","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FOpenRouter",[894,897,900,903,906,909],{"title":895,"url":896},"OpenRouter documentation: quickstart","https:\u002F\u002Fopenrouter.ai\u002Fdocs\u002Fquickstart",{"title":898,"url":899},"OpenRouter pricing","https:\u002F\u002Fopenrouter.ai\u002Fpricing",{"title":901,"url":902},"OpenRouter documentation: model fallbacks","https:\u002F\u002Fopenrouter.ai\u002Fdocs\u002Fguides\u002Frouting\u002Fmodel-fallbacks",{"title":904,"url":905},"OpenRouter documentation: provider routing","https:\u002F\u002Fopenrouter.ai\u002Fdocs\u002Fguides\u002Frouting\u002Fprovider-selection",{"title":907,"url":908},"OpenRouter documentation index","https:\u002F\u002Fopenrouter.ai\u002Fdocs\u002Fllms.txt",{"title":910,"url":892},"Wikipedia: OpenRouter","\u002Fimages\u002Fblog\u002Fopenrouter\u002Fcover.webp","\u002Fimages\u002Fblog\u002Fopenrouter\u002Fog.jpg",[55,56,57],"OpenRouter: one API key in front of every model you might call","OpenRouter: 500+ models from 80+ providers behind one OpenAI-compatible endpoint, with fallbacks and pass-through pricing. Costs and where it breaks.","OpenRouter: one key on a ring with lines fanning out to a row of different doors.",{"slug":918,"published":919,"updated":920,"minutes":921,"category":7,"tags":922,"tagPages":926,"keywords":927,"about":936,"sources":941,"cover":950,"og":951,"expertise":53,"locales":952,"lang":55,"title":953,"description":954,"coverAlt":955},"llm-cost-latency-prompt-caching-routing","2026-06-01","2026-09-22",9,[923,12,924,925],"Prompt caching","LLM cost","Latency",[],[928,929,930,931,932,933,934,935],"prompt caching","LLM cost optimization","LLM latency","model routing","batch API LLM","LLM cost per request","cheaper LLM model","cache hit rate LLM",[937,938],{"name":825,"url":826},{"name":939,"url":940},"Latency (engineering)","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FLatency_(engineering)",[942,945,948],{"title":943,"url":944},"Claude API docs: Prompt caching","https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fdocs\u002Fbuild-with-claude\u002Fprompt-caching",{"title":946,"url":947},"OpenAI API docs: Prompt caching","https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fprompt-caching",{"title":949,"url":50},"Claude API docs: Models overview, context windows and prices (as of September 2026)","\u002Fimages\u002Fblog\u002Fllm-cost-latency-prompt-caching-routing\u002Fcover.webp","\u002Fimages\u002Fblog\u002Fllm-cost-latency-prompt-caching-routing\u002Fog.jpg",[55,56,57],"Prompt caching and model routing: cutting LLM cost and latency","Prompt caching, cheap-model-first routing and batch APIs are the levers that cut LLM cost and latency in production. Here is how to use each one.","LLM cost and routing: a railway switch splits a track toward a small and a large station, a watch and a coin beside it.",{"slug":957,"published":958,"minutes":959,"category":960,"tags":961,"tagPages":966,"keywords":969,"about":980,"sources":985,"cover":1020,"og":1021,"expertise":53,"locales":1022,"lang":55,"title":1023,"description":1024,"coverAlt":1025},"claude-in-opencode","2026-10-30",12,"agents",[962,963,964,965,10],"OpenCode","Claude Code","Claude Agent SDK","Coding agents",[967,968],"claude-code","coding-agents",[970,971,972,973,974,975,976,977,978,979],"claude in opencode","opencode claude","opencode claude agent","opencode claude subscription","opencode claude max","opencode anthropic api key","claude agent sdk opencode","opencode claude code login","opencode github copilot claude","opencode zen claude pricing",[981,982,983],{"name":27,"url":28},{"name":30,"url":31},{"name":962,"url":984},"https:\u002F\u002Fopencode.ai",[986,989,992,995,998,1001,1004,1007,1010,1013,1016,1017],{"title":987,"url":988},"OpenCode v2.2.3 release notes (October 30, 2026)","https:\u002F\u002Fgithub.com\u002Fanomalyco\u002Fopencode\u002Freleases\u002Ftag\u002Fv2.2.3",{"title":990,"url":991},"OpenCode pull request #58782: add Claude Agent provider","https:\u002F\u002Fgithub.com\u002Fanomalyco\u002Fopencode\u002Fpull\u002F58782",{"title":993,"url":994},"Claude Agent provider design spec","https:\u002F\u002Fgithub.com\u002FTarquinen\u002Fopencode\u002Fblob\u002Fclaude-agent-provider\u002Fspecs\u002Fv2\u002Fclaude-agent-provider.md",{"title":996,"url":997},"OpenCode docs: Providers","https:\u002F\u002Fopencode.ai\u002Fv2\u002Fdocs\u002Fproviders\u002F",{"title":999,"url":1000},"OpenCode docs: Models","https:\u002F\u002Fopencode.ai\u002Fv2\u002Fdocs\u002Fmodels\u002F",{"title":1002,"url":1003},"OpenCode docs (v1): Providers, note on Claude Pro\u002FMax plugins","https:\u002F\u002Fopencode.ai\u002Fdocs\u002Fproviders\u002F",{"title":1005,"url":1006},"OpenCode Console: models and pricing","https:\u002F\u002Fopencode.ai\u002Fv2\u002Fdocs\u002Fconsole\u002Fmodels\u002F",{"title":1008,"url":1009},"OpenCode Go: plans and usage limits","https:\u002F\u002Fopencode.ai\u002Fv2\u002Fdocs\u002Fconsole\u002Fgo\u002F",{"title":1011,"url":1012},"Claude Code docs: Legal and compliance","https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Flegal-and-compliance",{"title":1014,"url":1015},"Claude Help Center: Use the Claude Agent SDK with your Claude plan","https:\u002F\u002Fsupport.claude.com\u002Fen\u002Farticles\u002F15036540-use-the-claude-agent-sdk-with-your-claude-plan",{"title":40,"url":41},{"title":1018,"url":1019},"GitHub Docs: Supported AI models in GitHub Copilot","https:\u002F\u002Fdocs.github.com\u002Fen\u002Fcopilot\u002Freference\u002Fai-models\u002Fsupported-models","\u002Fimages\u002Fblog\u002Fclaude-in-opencode\u002Fcover.webp","\u002Fimages\u002Fblog\u002Fclaude-in-opencode\u002Fog.jpg",[55,56,57],"Claude in OpenCode: the Claude Agent provider, API keys and the other routes","How to use Claude in OpenCode: the new Claude Agent provider on your Claude plan, API keys, Console, Copilot and OpenRouter, with costs and limits.","Claude in OpenCode: a hand plugs one of two adapters into a laptop that shows a terminal window.",1793451245121]