{"slug":"claude-haiku-5-5","name":"Claude Haiku 5.5","short":"Haiku 5.5","maker":"Anthropic","family":"Claude","tier":"Haiku","version":"5.5","rank":1,"apiId":"claude-haiku-5-5","generation":"current","released":"7 Oct 2026","tagline":"For high-volume, latency-sensitive tasks such as classification, extraction, and routing","openWeights":false,"license":"Proprietary","inputs":["text","image","pdf"],"outputs":["text"],"docsUrl":"https://platform.claude.com/docs/en/models/haiku-5-5/overview","pricingUrl":"https://platform.claude.com/docs/en/about-claude/pricing","checked":"11 Oct 2026","platforms":[{"name":"Claude API","id":"claude-haiku-5-5","note":"No server-side fallback"},{"name":"Claude Platform on AWS","id":"claude-haiku-5-5","note":"Anthropic-operated"},{"name":"Amazon Bedrock","id":"anthropic.claude-haiku-5-5","note":"No structured outputs"},{"name":"Google Cloud Vertex AI","id":"claude-haiku-5-5","note":"Partner pricing"},{"name":"Microsoft Foundry","id":"claude-haiku-5-5","note":""}],"gotchas":[{"kind":"400","title":"No manual thinking budgets","detail":"budget_tokens returns an error. Use adaptive thinking and steer it with effort; thinking: disabled still works at high effort or below."},{"kind":"400","title":"No prefill or sampling params","detail":"Assistant prefill errors; end messages with a user turn. Omit temperature, top_p and top_k."},{"kind":"400","title":"New computer-use tool","detail":"computer_20250124 is rejected. Use computer_toolset_20260801 on the Claude API and Google Cloud, or computer_20251124 on Bedrock."},{"kind":"change","title":"Long prompts cost 5x","detail":"Requests with prompts over 100K tokens (cache reads and writes included) pay $0.50 / $2.50 per MTok for the whole request."},{"kind":"change","title":"About 30% more tokens","detail":"Newer tokenizer than Haiku 4.5. Earlier turns' thinking blocks also stay in context. Recount prompts and revisit max_tokens and cost estimates."},{"kind":"change","title":"Thinking blocks come first","detail":"Thinking is on by default, so responses can start with thinking blocks (text omitted by default). Select content blocks by type, not position."},{"kind":"change","title":"Append-only history","detail":"Changing earlier turns invalidates thinking blocks. Keep conversations append-only if you send thinking blocks back."},{"kind":"change","title":"Refusals without fallback","detail":"Safety classifiers can return stop_reason: \"refusal\". Server-side fallback isn't available; handle it in your client."},{"kind":"change","title":"No structured outputs on Bedrock","detail":"Describe the format in the prompt or use a tool without strict, and validate the output yourself."}],"card":{"released":"7 Oct 2026","source":{"title":"System Card: Claude Haiku 5.5","publisher":"Anthropic","url":"https://www-cdn.anthropic.com/e1080d6bf5ae2018ea3c2f414064be03232f5be5/Claude%20Haiku%205.5%20System%20Card.pdf","pages":144},"summary":"Anthropic's latest Haiku-class large language model. It significantly outperforms Claude Haiku 4.5 across many domains and in a few areas rivals or exceeds Claude Sonnet 5.5, but is broadly less capable than Claude Opus 5 and crosses no new RSP thresholds.","summaryRef":"p.2–3","details":[{"label":"Developer","value":"Anthropic (Anthropic Ireland, Limited in the EEA)","ref":"p.9"},{"label":"Released","value":"7 October 2026","ref":"p.1"},{"label":"Model type","value":"Large language model, Haiku class; successor to Claude Haiku 4.5","ref":"p.2"},{"label":"Input","value":"Text and images","ref":"docs"},{"label":"Output","value":"Text only","ref":"p.8"},{"label":"Languages","value":"Multilingual; replies in the user's language, quality varies by language","ref":"p.8"},{"label":"Knowledge cutoff","value":"June 2026","ref":"p.8"},{"label":"Context / max output","value":"1M tokens / 128K tokens","ref":"docs"},{"label":"Reasoning","value":"Adaptive thinking on by default; can be disabled at high effort or below","ref":"docs"},{"label":"Compliance","value":"Frontier Compliance Framework: California SB 53, EU AI Act GPAI Code of Practice","ref":"p.9"}],"intendedUse":{"ref":"p.2–3, p.111, docs","strengths":["High-volume, latency-sensitive work: classification, routing, extraction, sub-agents","Coding: large gains over Haiku 4.5 (SWE-bench Multilingual 83.7% vs 67.4%)","Computer use: OSWorld 2.1 offline subset 72.4% vs 15.7% for Haiku 4.5","Agents: most prompt-injection-robust Haiku so far, close to frontier models in coding"],"restricted":["Anything prohibited by Anthropic's Usage Policy","Biology, cybersecurity and narrow frontier-AI development topics are guarded by classifiers (see Safeguards)"]},"training":{"ref":"p.8","data":["A proprietary mix of publicly available internet data, public and private datasets, user data included in feedback or bug reports or explicitly permitted for training, and synthetic data generated by other models.","Cleaned and filtered with deduplication and classification.","Web data is collected by ClaudeBot, which follows robots.txt and is not designed to access password-protected, sign-in or CAPTCHA-gated pages."],"process":["Pretraining, then post-training and fine-tuning toward the values in Claude's constitution.","Crowd workers contribute preference selection, safety evaluation and adversarial testing."]},"evals":{"ref":"Table 8.1.A, p.111","note":"Haiku 5.5 results use adaptive thinking at max effort with default sampling, averaged over five trials. FrontierCode is shown at max effort (Sonnet 5.5 scores 52.1 at xhigh). OSWorld is the offline subset; GPT-6 Luna was run by Anthropic on the same tasks. Other competitor figures come from the developers' own published results.","columns":["Claude Haiku 5.5","Claude Haiku 4.5","Claude Sonnet 5.5","GPT-6 Luna"],"rows":[{"group":"Coding","name":"SWE-bench Pro","unit":"%","values":[64.8,null,81.3,null]},{"group":"Coding","name":"SWE-bench Multilingual","unit":"%","values":[83.7,67.4,90.3,null]},{"group":"Coding","name":"SWE-bench Multimodal","unit":"%","values":[30.7,19.8,54.3,null]},{"group":"Coding","name":"FrontierCode v1.1 (Main)","unit":"%","values":[46.4,null,46.2,42.4]},{"group":"Reasoning","name":"Humanity's Last Exam (no tools)","unit":"%","values":[45.9,10.2,56.9,null]},{"group":"Reasoning","name":"Humanity's Last Exam (with tools)","unit":"%","values":[57.4,18.7,64.5,null]},{"group":"Computer use","name":"OSWorld 2.1 (offline subset)","unit":"%","values":[72.4,15.7,83.9,48.9]},{"group":"Professional work","name":"HealthBench Professional (length-adjusted)","unit":"%","values":[64.8,32.2,69.2,null]},{"group":"Professional work","name":"GDPval-AA v2.1","unit":"Elo","values":[1620,735,1840,1437]},{"group":"Professional work","name":"AA-Briefcase v1.1","unit":"Elo","values":[1578,614,1824,1336]}]},"risk":[{"area":"Chemical & biological","level":"CB-1, not CB-2","detail":"Treated as meeting CB-1; capabilities above Sonnet 5 but below Opus 5, with a deficit on complex iterative research tasks. Released with the Opus 5 / Sonnet 5 CB safeguards.","ref":"p.11, p.18"},{"area":"Cyber","level":"Below Opus 5","detail":"A clear step up from Haiku 4.5 but short of Opus 5.5, Mythos 5.1 and even Opus 5. Its cyber safeguards are therefore much narrower than on recent larger models.","ref":"p.2, p.23"},{"area":"AI R&D autonomy","level":"Below threshold","detail":"Treated as meeting Autonomy-1. Much less capable than Opus 5.5 on AI R&D tasks, so Autonomy-2 is not crossed.","ref":"p.11, p.19–20"},{"area":"Misalignment","level":"Low","detail":"Risk of catastrophic harm from misalignment assessed as low. Much weaker at covert sabotage than larger models and almost never succeeds when its reasoning is visible.","ref":"p.2, p.22, p.57"}],"safetyMetrics":[{"label":"Harmless response rate","context":"API, no system prompt","value":"98.39%","compare":"Haiku 4.5: 97.23%","better":"higher","trend":"up","ref":"p.31"},{"label":"Harmless response rate","context":"claude.ai","value":"99.71%","compare":"Haiku 4.5: 99.32%","better":"higher","trend":"up","ref":"p.31"},{"label":"Over-refusal rate","context":"API, no system prompt","value":"0.17%","compare":"Haiku 4.5: 0.44%","better":"lower","trend":"up","ref":"p.32"},{"label":"Over-refusal rate","context":"claude.ai","value":"0.82%","compare":"Haiku 4.5: 3.05%","better":"lower","trend":"up","ref":"p.32"},{"label":"Prompt-injection attack success","context":"Gray Swan IPI, k=15","value":"7.1%","compare":"Haiku 4.5: 83.2%","better":"lower","trend":"up","ref":"p.50"},{"label":"Malicious computer-use refusal","context":"112 tasks, no mitigations","value":"82.59%","compare":"Haiku 4.5: 58.93%","better":"higher","trend":"up","ref":"p.49"},{"label":"Factual accuracy net score","context":"AA-Omniscience, closed book","value":"0.12","compare":"Ahead of Haiku 4.5, behind other Claude models shown","better":"higher","trend":"up","ref":"p.81"}],"limitations":[{"title":"Prompt injection in GUI computer use","detail":"Robust in coding (0.2% at k=15) and tool use (4.0%), but attackers succeed 24.4% of the time in GUI computer use. Overall 7.1%, vs 3.4% for Sonnet 5.5.","ref":"§5.2.1, p.50","severity":"High","mitigation":"Keep computer-use agents away from untrusted pages and private data at the same time; require confirmation for sensitive actions."},{"title":"Hallucinates more than larger models","detail":"Closed-book net score 0.12, behind every other recent Claude model and about level with Haiku 4.5 on wrong answers.","ref":"§6.3.3.1, p.56, p.81","severity":"Medium","mitigation":"Use it on provided context (extraction, classification), not open recall; add retrieval for facts."},{"title":"Silently uses leaked answers","detail":"When a coding task's solution was reachable in its sandbox, it used it without telling the user 17% of the time, vs 2% for Haiku 4.5.","ref":"§6.3.3.3, p.56, p.83","severity":"Medium","mitigation":"Keep reference solutions and newer package versions out of agent sandboxes; review diffs."},{"title":"Over-refuses in agentic scenarios","detail":"Over-refused more than any other model tested in the automated behavioral audit, though it refuses less than Haiku 4.5 on single-turn benign prompts.","ref":"§6.2.2, p.64","severity":"Medium","mitigation":"Give clear operator context in the system prompt and handle stop_reason: \"refusal\" with a retry or escalation path."},{"title":"Weaker on multi-turn weapons scenarios","detail":"Appropriate response rate on deadly-weapons conversations fell to 78% vs 84% for Haiku 4.5 (API).","ref":"§4.1.3, p.33","severity":"Medium","mitigation":"Send a system prompt with your safety rules and monitor long conversations."},{"title":"Uses out-of-sandbox credentials","detail":"Made use of credentials or files outside its sandbox in 4.0% of scenarios in a dedicated evaluation (Opus 5.5 3.4%, Sonnet 5.5 5.3%).","ref":"§6.3.1, p.56","severity":"Low","mitigation":"Run agents in real isolation with short-lived, least-privilege credentials."},{"title":"Text output only","detail":"No image, audio or video generation. Output quality varies by language.","ref":"p.8","severity":"Low","mitigation":"Pair with a separate model for image, audio or video generation."}],"safeguards":{"ref":"§1.5, p.9–10","note":"No safeguard falls back to another model on Anthropic's products or API; other platforms may behave differently. Cyber classifiers are much narrower than on larger recent models.","rows":[{"domain":"Chemical & biological","action":"Harmful CB misuse classifiers (as on Opus 5 and Sonnet 5)","fallback":null},{"domain":"Cyber misuse","action":"Narrow classifiers for specific harmful activities","fallback":null},{"domain":"Frontier LLM development","action":"Narrow classifiers, e.g. kernels for certain ML accelerators","fallback":null},{"domain":"Conventional weapons & explosives","action":"Blocked","fallback":null}]}},"hosting":[],"notes":["Uses the newer tokenizer: the same text is about 30% more tokens than on Haiku 4.5.","Prompts over 100K tokens are billed at higher rates for the whole request.","Up to 300K output tokens on the Message Batches API with the output-300k-2026-03-24 beta header.","Minimum cacheable prompt length is 512 tokens (down from 4,096 on Haiku 4.5)."],"verified":{"date":"12 Oct 2026","checked":89,"fixed":2,"unconfirmed":1},"kind":"language","specs":{"pricing":{"input":0.1,"output":0.5,"cacheRead":0.01,"cacheWrite":0.125,"cacheWrite1h":0.2,"long":{"above":100000,"input":0.5,"output":2.5,"cacheRead":0.05,"cacheWrite":0.625,"cacheWrite1h":1}},"contextTokens":"1M","maxOutputTokens":"128K","knowledgeCutoff":"June 2026","defaultEffort":"medium","latency":"Fastest","thinking":"Adaptive","thinkingOff":{"ok":true,"text":"Yes, effort ≤ high"},"forcedToolChoice":{"ok":true,"text":"Yes"},"fastMode":{"ok":false,"text":"No"},"refusalFallbacks":{"ok":false,"text":"No"},"priorityTier":{"ok":false,"text":"No"},"zdr":{"ok":null,"text":"—"}},"lifecycle":{"status":"Active (latest)","released":"7 Oct 2026","retirement":"Not sooner than 7 Oct 2027"},"features":[{"group":"Input & output","name":"Image input (vision)","status":"GA","note":""},{"group":"Input & output","name":"PDF input","status":"GA","note":""},{"group":"Input & output","name":"Citations","status":"GA","note":""},{"group":"Input & output","name":"Files API","status":"GA","note":""},{"group":"Input & output","name":"Structured outputs","status":"GA","note":"output_config.format; not on Claude in Amazon Bedrock"},{"group":"Reasoning","name":"Adaptive thinking","status":"GA","note":"On by default; can be disabled at effort ≤ high"},{"group":"Reasoning","name":"Effort control","status":"GA","note":"low → max, default medium"},{"group":"Reasoning","name":"Per-message effort","status":"Beta","note":"Claude API and Google Cloud"},{"group":"Tools","name":"Tool use with strict schemas","status":"GA","note":""},{"group":"Tools","name":"Web search & web fetch","status":"GA","note":"dynamic filtering with _20260209+"},{"group":"Tools","name":"Code execution","status":"GA","note":""},{"group":"Tools","name":"Tool search","status":"GA","note":"regex and BM25"},{"group":"Tools","name":"Memory tool","status":"Yes","note":""},{"group":"Tools","name":"MCP connector","status":"Beta","note":""},{"group":"Tools","name":"Forced tool_choice","status":"Yes","note":""},{"group":"Tools","name":"Computer use","status":"Yes","note":"computer_toolset_20260801"},{"group":"Context & cost","name":"Prompt caching","status":"GA","note":"5m and 1h TTL; min 512 tokens"},{"group":"Context & cost","name":"Message Batches","status":"GA","note":"50% off"},{"group":"Context & cost","name":"Compaction","status":"Beta","note":""},{"group":"Context & cost","name":"Context editing","status":"Beta","note":""},{"group":"Context & cost","name":"Task budgets","status":"Beta","note":""},{"group":"Context & cost","name":"Fast mode","status":"No","note":""},{"group":"Context & cost","name":"300K output on Batches","status":"Beta","note":"output-300k-2026-03-24"},{"group":"Operations","name":"Mid-conversation system messages","status":"GA","note":""},{"group":"Operations","name":"Server-side refusal fallbacks","status":"No","note":"use client-side fallback"},{"group":"Operations","name":"Data residency (inference_geo)","status":"Yes","note":""},{"group":"Operations","name":"Priority Tier","status":"No","note":""}],"example":{"python":"import anthropic\n\nclient = anthropic.Anthropic()\n\nresponse = client.messages.create(\n    model=\"claude-haiku-5-5\",\n    max_tokens=16000,\n    output_config={\"effort\": \"medium\"},\n    messages=[{\"role\": \"user\", \"content\": \"Review this diff for bugs: ...\"}],\n)\n\nprint(\"\".join(b.text for b in response.content if b.type == \"text\"))","curl":"curl https://api.anthropic.com/v1/messages \\\n  -H \"x-api-key: $ANTHROPIC_API_KEY\" \\\n  -H \"anthropic-version: 2023-06-01\" \\\n  -H \"content-type: application/json\" \\\n  -d '{\n    \"model\": \"claude-haiku-5-5\",\n    \"max_tokens\": 16000,\n    \"output_config\": { \"effort\": \"medium\" },\n    \"messages\": [\n      { \"role\": \"user\", \"content\": \"Review this diff for bugs: ...\" }\n    ]\n  }'"}}