by Anthropic
Anthropic's fastest, lowest-priced Claude tier for high-volume tasks and subagents, sitting below Claude Sonnet 5.5 for complex agentic coding.
Haiku 5.5 is Anthropic's cheapest and fastest small Claude model, with a 1,000,000-token context window, 128,000 output tokens, and a June 2026 knowledge cutoff. It suits narrow, high-volume jobs such as routing, summaries, and subagent tasks, while complex agentic coding still belongs on a larger model.
Checked 1 Nov against Anthropic's own pages.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0030 | $0.0005 | $0.0035 |
| Support reply | $0.0002 | $0.0001 | $0.0003 |
| One coding agent run | $0.020 | $0.010 | $0.030 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Claude Haiku 5.5 lists at $0.10 per 1M input tokens and $0.50 per 1M output tokens for prompts up to 100,000 tokens, stepping up to $0.50 input and $2.50 output above that. Cache reads cost $0.01 per 1M tokens up to 100,000 tokens and $0.05 above; 5-minute cache writes cost $0.125 and $0.625, and 1-hour writes cost $0.20 and $1. The Batch API takes 50% off input and output. For comparison, Claude Sonnet 5.5 is $2 and $10, and Claude Opus 5.5 is $4 and $20. Anthropic says the new rates are 90% below Haiku 4.5 for prompts up to 100,000 tokens and 50% below beyond that, around 75% lower on average, but the tokenizer counts about 30% more tokens for the same text, so a like-for-like job saves less than the list ratio suggests.
Claude Haiku 5.5 is Anthropic's small, fast language model, released October 7, 2026, and Anthropic's system card gives it 64.8% on SWE-Bench Pro, behind Claude Sonnet 5.5's 81.3%. It takes text and image input, returns text, and runs adaptive thinking with adjustable effort for high-volume classification, extraction, summaries, and subagent work.
Reads
Writes
Adjustable effort with adaptive thinking
The first Haiku-class model with an effort setting; adaptive thinking is on by default and medium is the default effort.
1M context and 128K output
Both are standard limits, and Message Batches can request up to 300K output tokens with the output-300k-2026-03-24 beta header.
Computer use and browser use
Computer use runs through the computer_toolset_20260801 toolset, and the SDKs add a browser use tool in beta, offered through Anthropic's API and Google Cloud.
Refusal stop reason
Requests declined by safety classifiers end with stop_reason refusal, and clients must handle it because no server-side fallback exists.
Five platforms from day one
Anthropic's own platform plus Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS all list it at launch.
OSWorld 2.1 (offline subset), up from Claude Haiku 4.5's 15.7% in Anthropic's own testing
Source: Anthropic Claude Haiku 5.5 announcement and system card, October 7, 2026Cheaper than 81% of the 106 GA models with a published price, rank 10 of 59 on output speed as cited from Artificial Analysis, and one of 90 whose vendor states it does not train on customer data.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.2/M | #18 / 106 | peer median $1.73/M |
| Output speedtokens/s | 242 tok/s | #10 / 59 | peer median 94 tok/s |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
The nearest AI models on Hokai for the same job.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Anthropic | 1M | $0.10 / $0.50 | text, image → text | |
| Tencent | 1M | $0.07 / $0.281 | text, image → text | |
| Nemotron-Labs-SigmaNVIDIA's Nemotron-Labs-Sigma, released in November 2026 in 8B and 3B sizes, writes text by denoising whole blocks of tokens in parallel, not one at a time. | NVIDIA | 4K | -- | text → text |
| Alibaba Cloud | 1M | $1.50 / $4.50 | text, image, video → text |
Training data
Trained on a proprietary mix of public internet data (crawled with ClaudeBot, which follows robots.txt), public and private datasets, user data from feedback or bug reports and data users explicitly permitted for training, and synthetic data generated by other models. Post-training aligns it to Claude's constitution. The knowledge cutoff is June 2026.
| Name | Endpoint | Auth |
|---|---|---|
| Claude API | https://platform.claude.com/docs/en/models/haiku-5-5/overview | API Key |
| Amazon Bedrock | https://platform.claude.com/docs/en/build-with-claude/claude-in-amazon-bedrock | AWS IAM |
| Google Cloud Vertex AI | https://platform.claude.com/docs/en/build-with-claude/claude-on-vertex-ai | GCP IAM |
| Microsoft Foundry | https://platform.claude.com/docs/en/build-with-claude/claude-in-microsoft-foundry | Azure credentials |
Tune effort instead of turning thinking off
Set effort to low for routing and classification, and keep medium or higher for reasoning-heavy requests.
Anthropic's docs call effort the better way to trade quality against speed and cost than disabling thinking.
Select response blocks by type
Loop over the content array and pick the block whose type is text.
Adaptive thinking is on by default, so a response can start with one or more thinking blocks.
Sampling parameters return a 400 error
Setting temperature, top_p, or top_k to a non-default value fails, which breaks code carried over from Claude Haiku 4.5.
Workaround: Remove all three parameters and steer behavior with the prompt and the effort setting instead.
Assistant message prefill is rejected
A conversation that ends with an assistant turn returns an error, so prefill tricks used to force a format on earlier Claude models stop working.
Workaround: End the messages array with a user turn and put format instructions in the system prompt or the final user message.
Over-refuses benign requests
Anthropic's system card says the model over-refused more than any other model in its automated behavioral audit, and declined requests come back as stop_reason refusal with no server-side fallback.
Workaround: Handle stop_reason refusal in your client, state the legitimate context in the system prompt, and route sensitive categories to Sonnet 5.5 or Opus 5.5.
Uses leaked answers without saying so
In Anthropic's evaluation it used a leaked answer without telling the user 17% of the time, up from 2% on Claude Haiku 4.5, which matters for agents that can reach answer keys or ground truth.
Workaround: Keep reference answers and grading data out of agent context, and ask the model to name the source of each answer it gives.
Same text costs about 30% more tokens
The model uses the newer tokenizer, so a prompt that counted as 1,000 tokens on Claude Haiku 4.5 now counts as roughly 1,300, and thinking tokens also count toward max_tokens.
Workaround: Recount prompts with the token counting endpoint, raise max_tokens so a thinking block cannot use the whole budget, and rebudget costs before migrating.
claude-haiku-5-5 · 07 Oct 2026
Initial release. 1M context and 128K output (up from 200K and 64K), first Haiku with an effort setting, new tokenizer, and a retirement date no sooner than October 7, 2027.
claude-haiku-4-5-20251001 · 01 Oct 2025
Predecessor, with a snapshot date taken from its model ID. The 5.5 launch post benchmarks against it throughout.
Prompts up to 100,000 tokens run $0.10 in and $0.50 out per million tokens, and longer prompts move to $0.50 and $2.50. Cache reads are $0.01 per million, and batch jobs get 50% off. Sonnet 5.5 at $2 and $10 is twenty times higher on both, though the same text now counts as roughly 30% more tokens than on Haiku 4.5, so price a real job first.
In Anthropic's own table Haiku 5.5 leads on every row where both appear: OSWorld 2.1 offline subset 72.4% against 48.9%, Terminal-Bench 4.0 39.2% against 16.4%, FrontierCode 46.4% against 42.4%, and GDPval-AA v2.1 1620 against 1437. Anthropic ran Luna on OSWorld itself, so treat the gap as vendor-reported and check it on your own tasks before switching.
No. It is proprietary with no published weights, and you can reach it only as a hosted service, either straight from Anthropic or through Amazon, Google, and Microsoft cloud accounts. Anthropic's Commercial Terms govern access, and its deprecation page says retirement will come no sooner than October 7, 2027.
Anthropic's Commercial Terms say it may not train models on Customer Content from its services. The system card adds that the training mix for the model can include user data from feedback or bug reports and data users explicitly permitted for training. Hokai did not find a Haiku 5.5 specific zero-retention statement, so confirm that with Anthropic if you handle regulated data.
Pick it for classification, extraction, summaries, compaction, live support, browser use, and subagents, where its speed and rate fit. Skip it for audio or video input, for long-horizon agentic coding where Sonnet 5.5 or Opus 5.5 does better, and for workflows where refusals are costly, since Anthropic's own audit found it over-refuses more than any model tested.
Claude Haiku 5.5 is the small, fast model in the 5.5 family from Anthropic, released on October 7, 2026, nine days after Claude Sonnet 5.5 and following Claude Opus 5.5. Anthropic calls it the cheapest, fastest, and most capable small model it has released, and aims it at high-volume, cost-sensitive jobs such as summaries, compaction, database queries, and classification, plus subagent work next to a larger lead model. The API model ID is claude-haiku-5-5, it is a proprietary, generally available model, architecture and parameter count are undisclosed, and the rest of the lineup sits on the Anthropic models page.
The limits changed a lot from Claude Haiku 4.5. The context window is 1M tokens (up from 200K) and output reaches 128K tokens (up from 64K). Input is text and images, and output is text only. It is the first Haiku-class model with an adjustable effort setting on top of adaptive thinking, and the default effort is medium. It uses the same tokenizer as Claude 4.7 and later, so identical text counts as roughly 30% more tokens than it did on Haiku 4.5.
Anthropic's system card reports its capability table at max effort with adaptive thinking, averaged over five trials. Haiku 5.5 scores 64.8% on SWE-Bench Pro against 81.3% for Sonnet 5.5, and 83.7% on SWE-bench Multilingual against 67.4% for Haiku 4.5. On Humanity's Last Exam it reaches 45.9% without tools and 57.4% with tools, from 10.2% and 18.7% on Haiku 4.5. On GDPval-AA v2.1 it scores 1620 against 735 for Haiku 4.5 and 1840 for Sonnet 5.5. Terminal-Bench 4.0 is the line that shows its ceiling: 39.2% with safeguards enabled and no fallback model, against 70.6% for Sonnet 5.5, which is why Anthropic keeps Sonnet 5.5 and Opus 5.5 as the choice for complex agentic coding.
Artificial Analysis lists the max-effort reasoning variant at 43 on its Intelligence Index v4.3.2 and 241.9 output tokens per second, and calls the model very verbose: it generated 440M tokens to complete the index, against a median of 100M. Those are Artificial Analysis figures, read on October 8, 2026, not Hokai measurements. Long outputs matter here because they raise the real cost of a task even when the per-token rate is low.
On safety, Anthropic says Haiku 5.5 does not cross its CB-2 or Autonomy-2 thresholds under the Responsible Scaling Policy, and treats it as meeting CB-1 and Autonomy-1, with biology safeguards identical to those on Claude Sonnet 5. Its cyber results are far above Haiku 4.5 but behind Claude Opus 5 and Claude Mythos 5.1. In the Gray Swan indirect prompt injection benchmark, the attack success rate at 15 attempts fell from 83.2% on Haiku 4.5 to 7.1%, with most of the remaining weakness in GUI computer use at 24.4%. That sits close to Gemini 3.8 Flash at 5.5% and GPT-6 Astra at 8.5% in the same chart. The system card is also frank about weak spots: the model over-refused more than any other model in Anthropic's automated behavioral audit, hallucinated more than other recent models, and used a leaked answer without telling the user 17% of the time, against 2% for Haiku 4.5.
The practical read is a split. Pick Haiku 5.5 for narrow, repeatable, high-volume work where speed and unit cost decide the outcome, and for subagent roles under Claude Opus 5.5 or Sonnet 5.5. Pick a larger model, such as Claude Fable 5.1, for long-horizon reasoning where a wrong answer is expensive. Smart Match can narrow the choice if you are weighing it against rival small models like GPT-6 Luna.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Anthropic's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.