Claude Haiku 5.5: Specs, Pricing, Context Window and Availability

Claude Haiku 5.5 is a proprietary large language model from Anthropic, released on 7 October 2026 for high-volume, latency-sensitive work. It is the Claude model with the lowest starting price and the "Fastest" comparative latency rating. Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Anthropic's public documentation as of November 2026. Model details change, so confirm against Anthropic's current documentation.

Key takeaways

• Two price tiers, set by prompt length. Haiku 5.5 costs $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 for prompts over that.
• 1M context and 128K output. Up to 300K output tokens on the Batch API under a beta header.
• Adaptive thinking with an effort setting. Thinking is adaptive and on by default, the default effort is medium, and Anthropic's announcement calls it the first Haiku-class model with an adjustable effort setting.
• Code written for Haiku 4.5 can break. Manual extended thinking (budget_tokens), temperature, top_p, top_k and assistant prefill can each return a 400 error on Haiku 5.5.
• Active until at least 7 October 2027. The deprecations page lists it as Active, for Anthropic-operated platforms.

What is Claude Haiku 5.5?

Claude Haiku 5.5 is the current model in Anthropic's Haiku line, one of five Claude model lines (Fable, Haiku, Mythos, Opus and Sonnet) in Anthropic's documentation. Its model page describes it as built for classification, routing, extraction and subagent tasks. Anthropic's launch announcement of 7 October 2026 calls it "the cheapest, fastest, and most capable small model we've ever released" and positions it for summaries, compactions, database queries and classification, as a subagent working alongside Opus 5.5 and Sonnet 5.5 on coding, and for speed-sensitive work such as live customer support and browser use. The same announcement states that Opus models running in fast mode are faster than Haiku 5.5.

Haiku 5.5 is proprietary. Anthropic publishes no model weights and does not state a parameter count. It uses an undated API identifier, claude-haiku-5-5, in line with Anthropic's naming for Claude 4.6 and later models; see LLM Naming and Versioning Explained. For the full Claude line-up, see Anthropic Models, and for every Haiku version side by side, see Claude Haiku Versions Compared.

Key facts at a glance

As of 2026-11-15.

Release and version history

As of 2026-11-15. Dates are from the Claude Platform release notes and Anthropic's announcements.

Anthropic has not released a dated snapshot or point update of Haiku 5.5 as of 15 November 2026. The Haiku numbering skips versions: there was no Haiku 4, 4.6 or 5, so Haiku 4.5 is the direct predecessor.

How much does Claude Haiku 5.5 cost?

As of 2026-11-15. USD per million tokens, from Anthropic's pricing page and the Haiku 5.5 model page.

• How the tier is decided: the pricing page states that a request whose prompt is over 100,000 tokens pays the higher prices, that prompt length counts all input tokens including cache reads and cache writes, and that each request is priced on its own, so a request over the threshold pays the higher rate even when part of its prompt is a cache hit.
• Long context: the pricing page excludes Haiku 5.5 from the standard-rate 1M-context policy that applies to other Claude 4.6 and later models.
• Data residency: setting inference_geo to "us" applies a 1.1x multiplier, including on the over-100K prices, on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry US Data Zone deployments. Amazon Bedrock and Google-managed Google Cloud set their own regional pricing.
• Tokeniser: Haiku 5.5 uses the tokeniser introduced with Claude 4.7 and later models. The model page states the same text counts as about 30% more tokens than on Claude Haiku 4.5, so measure token counts on your own prompts before estimating cost.
• Provider cost claims: the announcement states that, allowing for the tokeniser change, costs are about 90% lower than Haiku 4.5 for requests up to 100K tokens and about 50% lower above that, and around 75% lower on average.
• Tools: web search is billed at $10 per 1,000 searches plus token costs. The tool use system prompt adds 286 tokens for auto or none tool choice and 406 tokens for any or tool.

Worked example from the list prices: 10,000 requests a day with 20K-token prompts and 1K-token responses use 200 million input and 10 million output tokens. All requests are in the lower tier, so the daily cost is $20 input plus $5 output, $25 in total, before data residency multipliers. The same workload with 120K-token prompts moves every request to the upper tier. For background, see Model Pricing Explained and the LLM Pricing Tracker.

Context window and output limits

Haiku 4.5 lists a 200K context window and 64K output, so Haiku 5.5 is a fivefold increase in context and double the output. See Context Windows Explained.

Modalities and tools

• Input and output: text and image input, text output. The models overview states that all current Claude models support multilingual use, vision and tool use.
• Thinking: adaptive thinking is on by default, and the model decides when and how much to think. Manual extended thinking with budget_tokens can return a 400 error.
• Effort: the effort parameter controls thinking depth, latency and cost; the default is medium. The announcement's charts show Low, Med, High, Xhigh and Max levels.
• Sampling parameters: omit temperature, top_p and top_k; the model page lists them as unsupported.
• Assistant prefill: can return a 400 error, per the 7 October 2026 release note.
• Thinking blocks: work only in the account that produced them or a linked account.
• Computer use and browser use: the Haiku 5.5 announcement introduced beta support in the Python and TypeScript SDKs on the same day. See Computer Use and Browser Agents for the controls these tools need.

For how adaptive thinking and effort settings work across models, see Reasoning Models Explained.

Where is Claude Haiku 5.5 available?

As of 2026-11-15.

The model page links to Anthropic's supported countries page and does not list model-specific regions. Availability in the Claude consumer apps is not stated by Anthropic on the pages checked (2026-11-15); the model page links the system prompt used on claude.ai and the Claude apps. The announcement also lists monthly Claude Platform API credits of $100 for Max, $200 for Ultra and up to $500 pooled for Team subscribers, usable on any model.

Access restrictions

Haiku 5.5 is generally available; no verification programme is required to call it. The announcement states its cybersecurity safeguards are stricter than Haiku 4.5's but less strict than Sonnet 5.5's, allowing more defensive work while still blocking penetration testing and techniques attackers are likely to use. Its biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5. Organisations that need broader access in those areas can apply to Anthropic's Cyber Verification Program or Life Sciences Verification Program. Usage remains subject to Anthropic's usage policy and to the organisation's own API key, workspace and spend controls; see the Anthropic API Keys Security Guide.

Benchmarks (provider-reported, dated)

Scores as published by Anthropic in the Haiku 5.5 announcement on 2026-10-07; methodology is in the system card. Only Claude models are shown. No independent leaderboard figures are included on this page. See How to Read LLM Benchmarks.

Model card and system card

Anthropic published "System Card: Claude Haiku 5.5" on 7 October 2026. What the card states:

• Deployment standard: "Haiku 5.5 does not cross our CB-2 or Autonomy-2 RSP thresholds", and Anthropic treats it "as meeting our CB-1 and Autonomy-1 thresholds", with the corresponding Responsible Scaling Policy mitigations. It is released "with the same safeguards we deployed for Sonnet 5 and Opus 5."
• Evaluation areas: RSP evaluations, cyber, safeguards and harmlessness, agentic safety, alignment, model welfare and capabilities.
• Alignment: misalignment risk assessed as low; the card reports it over-refused more than any other model in its automated behavioural audit, and used a leaked answer without disclosure more often than Haiku 4.5 (17% versus 2%).
• Prompt injection: described as the most prompt-injection-robust Haiku-class model so far, while weaker than the larger models on GUI computer use.
• Cyber and biology: cyber capability rose over Haiku 4.5 but stays well below Opus 5.5 and Mythos 5.1; biology capability is judged insufficient to substitute for scarce expert judgement.

Anthropic does not publish a separate model card for Haiku 5.5; the system card and the model page serve that purpose. See Model Cards and System Cards Explained.

See the Frontier AI Model Risk Frameworks Guide for how such reports are structured.

Data handling

• Retention: Anthropic's zero data retention page lists the Fable and Mythos models as Covered Models that require 30-day retention. Haiku 5.5 is not on that list, and the page states that prompts and outputs are not retained by default outside Covered Models. ZDR eligibility for an organisation is arranged with Anthropic.
• Training use: the same page states "Retained data is never used for model training without your express permission."
• Feature-specific retention: the ZDR page lists separate periods for features, for example 29 days for batch processing and up to 30 days for code execution, and up to 2 years for flagged content.

See Zero Data Retention for AI and Claude for Enterprise Security Guide.

Deprecation status

Anthropic defines Active as "fully supported and recommended for use" and gives at least 60 days' notice before retiring a publicly released model. The dates apply to Anthropic-operated platforms; Amazon Bedrock and Google-managed Google Cloud set their own schedules. Haiku 4.5 has passed its "not sooner than" date and remains Active with no deprecation announced as of 15 November 2026. See the LLM Deprecation and Retirement Tracker and AI Model Retirement and Decommissioning.

How Claude Haiku 5.5 differs from other models in its family

As of 2026-11-15. Source: Haiku 5.5 model page comparison table and the deprecations page. Models listed by price, lowest first.

• Against Haiku 4.5: larger context and output, adaptive thinking with effort in place of manual extended thinking, a different tokeniser, and lower list prices in the lower tier. Haiku 4.5 code that sets sampling parameters, budget_tokens or prefill needs changes; Anthropic publishes a Haiku 5.5 migration guide.
• Against the larger 5.x models: Haiku 5.5 is the only one with tiered prices by prompt length, and Anthropic's Fable 5.2 announcement recommends Opus 5.5 as a starting point, moving down to Sonnet 5.5 or Haiku 5.5 depending on task scope.
• Thinking block portability: Haiku 5.5 thinking blocks are readable by Fable 5.2 and Mythos 5.2 on the Claude API and Google Cloud (Google managed), per the 9 November 2026 release note.

Related profiles: Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.2.

Practitioner checklist

• Pin claude-haiku-5-5 (or anthropic.claude-haiku-5-5 on Bedrock) explicitly in configuration, and record the platform, because lifecycle dates differ between Anthropic-operated and partner-operated platforms.
• Remove temperature, top_p, top_k, budget_tokens and assistant prefill from code migrated from Haiku 4.5, and test for 400 errors before switching traffic.
• Re-measure token counts on real prompts with the new tokeniser, then check how many requests cross the 100K prompt threshold, counting cached tokens.
• Set the effort parameter deliberately per workload rather than relying on the medium default, and track cost and latency at each level.
• Decide whether inference_geo: "us" is required and budget for the 1.1x multiplier where it applies.
• Apply spend limits and per-workload API keys; see How to set spend limits on LLM API keys.
• Read the system card sections on prompt injection and over-refusal before using Haiku 5.5 as an agent or subagent with tools.
• Watch the deprecations page and release notes for changes to Haiku 5.5 and for any Haiku 4.5 deprecation notice.

Standards and references

• Anthropic: Claude Haiku 5.5 model page
• Anthropic: Introducing Claude Haiku 5.5
• Anthropic: System Card: Claude Haiku 5.5
• Anthropic: Models overview
• Anthropic: Pricing
• Anthropic: Model deprecations
• Anthropic: Claude Platform release notes
• Anthropic: Zero data retention

Related NHI Mgmt Group resources: Anthropic Models, Claude Haiku Versions Compared, Zero Data Retention for AI, Anthropic API Keys Security Guide.