Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 are the three tiers of Anthropic's Claude 5.5 family, released within 15 days of each other in September and October 2026, and they share a context window but differ sharply in price, latency, default effort and API behaviour. This page compares the three side by side: API IDs, release dates, lifecycle status, context and output limits, knowledge cutoff, prices including Haiku 5.5's two-tier rule, breaking changes, safeguards, provider-reported benchmarks and earliest retirement dates. It is written for developers choosing a tier or splitting work between tiers, buyers estimating cost and architects routing agent workloads. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Anthropic's own documentation and announcements, cited in the sources list.
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
- One window, three price points. All three list a 1M-token context window, 128K maximum output and a June 2026 reliable knowledge cutoff; list prices are $4/$20 (Opus 5.5), $2/$10 (Sonnet 5.5) and from $0.10/$0.50 (Haiku 5.5) per million tokens.
- Haiku 5.5 is priced by prompt length. Prompts over 100,000 tokens pay $0.50/$2.50 for the whole request, including cache reads; Opus 5.5 and Sonnet 5.5 bill the full 1M window at standard rates.
- Default effort differs. Anthropic lists default effort as medium for Haiku 5.5 and Opus 5.5 and high for Sonnet 5.5; thinking is always on for Opus 5.5 and adaptive for the other two.
- Each has its own breaking changes. Opus 5.5 and Sonnet 5.5 reject forced tool use; Sonnet 5.5 replaces disabled thinking with between_tools; Haiku 5.5 rejects manual thinking budgets, sampling parameters and assistant prefill.
- Safeguards fall back differently. Flagged cyber requests on Opus 5.5 are re-routed to Opus 4.8 and on Sonnet 5.5 to Sonnet 5; Anthropic places Haiku 5.5's cyber safeguards between Haiku 4.5 and Sonnet 5.5.
What is the Claude 5.5 family?
The Claude 5.5 family is Anthropic's set of Opus, Sonnet and Haiku models released from September 2026. Anthropic's models overview describes them in its own words: Opus 5.5 "for long-running agentic coding and knowledge work", Sonnet 5.5 as "the best combination of speed and intelligence", and Haiku 5.5 "for high-volume, latency-sensitive tasks such as classification, extraction, and routing". Comparative latency is listed as Moderate, Fast and Fastest respectively. All three are proprietary, with no published weights or parameter counts, and use undated API identifiers.
The family sits alongside Anthropic's Fable and Mythos lines, which are covered in Claude Fable vs Mythos vs Opus. For each model in depth, see the Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 profiles, and for the whole line-up the Anthropic Models hub.
Every model at a glance
As of 2026-11-15. Source: Anthropic models overview, model pages, model deprecations page and release notes.
| Model | API ID | Released | Status |
|---|---|---|---|
| Claude Haiku 5.5 | claude-haiku-5-5 | 2026-10-07 | Active |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | 2026-09-28 | Active |
| Claude Opus 5.5 | claude-opus-5-5 | 2026-09-22 | Active |
All three take text and image input and return text. Open weights: none. Licence: proprietary. Parameters: not stated by Anthropic.
When was each model released?
As of 2026-11-15. Dates are as shown in Anthropic's Claude Platform release notes and announcements.
| Date | Event | API ID |
|---|---|---|
| 2026-09-22 | Claude Opus 5.5 released, first model of the 5.5 family; fast mode research preview on the Claude API | claude-opus-5-5 |
| 2026-09-28 | Claude Sonnet 5.5 released at the same $2/$10 price as Sonnet 5 | claude-sonnet-5-5 |
| 2026-10-01 | Dreams supports Opus 5.5 and Sonnet 5.5 | claude-opus-5-5, claude-sonnet-5-5 |
| 2026-10-07 | Claude Haiku 5.5 released; Sonnet 5.5 cache read lowered from $0.20 to $0.10 | claude-haiku-5-5, claude-sonnet-5-5 |
| 2026-10-15 | Fast mode out of research preview for Opus 5.5; also on Microsoft Foundry (Azure-hosted deployments) | claude-opus-5-5 |
| 2026-10-22 | Fast mode for Opus 5.5 on Claude Models on Google Cloud (Anthropic managed) | claude-opus-5-5 |
How much does each model cost?
As of 2026-11-15. USD per million tokens, from Anthropic's pricing page.
| Model | Input | Cached input | Cache write | Output | Batch (input / output) | Notes stated by provider |
|---|---|---|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100K) | $0.10 | $0.01 | $0.125 (5m), $0.20 (1h) | $0.50 | $0.05 / $0.25 | Cache read 0.1x; no fast mode listed |
| Claude Haiku 5.5 (prompts over 100K) | $0.50 | $0.05 | $0.625 (5m), $1.00 (1h) | $2.50 | $0.25 / $1.25 | Whole request billed at the higher tier |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $2.50 (5m), $4.00 (1h) | $10.00 | $1.00 / $5.00 | Cache read 0.05x since 2026-10-07; no fast mode listed |
| Claude Opus 5.5 | $4.00 | $0.20 | $5.00 (5m), $8.00 (1h) | $20.00 | $2.00 / $10.00 | Cache read 0.05x; fast mode $8 / $40 |
Haiku 5.5 tier rule: the pricing page states that a request's prompt length "counts all of its input tokens, including cache reads and cache writes", and a request over the threshold "pays the higher prices even when part of its prompt is a cache hit". Agents that accumulate long histories can cross 100K tokens mid-session, so watch prompt length, not only cache hit rate.
Shared rules: all three use the newer tokeniser that Anthropic says produces about 30% more tokens than pre-4.7 models for the same text, so they compare directly with each other. Setting inference_geo to "us" applies a 1.1x multiplier to all token categories, including Haiku 5.5's higher tier. See Model Pricing Explained and How to set spend limits on LLM API keys.
What changed between models?
These are three tiers rather than successive versions, so this section records what each changed against its own predecessor and the API differences between them, as documented by Anthropic.
- Haiku 4.5 to Haiku 5.5: the release notes state that code written for Haiku 4.5 can break: manual extended thinking (budget_tokens), temperature, top_p and top_k, and assistant message prefill can each return a 400 error. Adaptive thinking is on by default, with thinking text omitted unless thinking.display is set to "summarized". Anthropic calls it the first Haiku-class model with an adjustable effort setting and states it costs around 75% less to run than Haiku 4.5 on average.
- Sonnet 5 to Sonnet 5.5: Anthropic lists five ways Sonnet 5 code can break, including forced tool use returning a 400 error. Up-front thinking is turned off with thinking type between_tools (at effort high or below) instead of disabled, and thinking blocks work only in the producing account or a linked one. The announcement states output is "30%+ faster" than Sonnet 5 and up to 30% cheaper per task.
- Opus 5 to Opus 5.5: prices fall from $5/$25 to $4/$20; thinking cannot be disabled, and both disabled and manual enabled thinking return a 400 error; tool_choice any and tool return a 400 error, with auto plus strict tool use advised. Anthropic states about 40% lower cost than Opus 5 on typical workloads. See Claude Opus 5.5 vs Opus 5 vs Opus 4.8.
Across the family, Sonnet 5.5 and Haiku 5.5 model pages state that temperature, top_p and top_k are not supported. For effort and thinking background, see Reasoning Models Explained.
What workloads does Anthropic position each model for?
This section records the provider's own positioning wording; it is not an assessment.
| Model | Positioning stated by provider | Related notes stated by provider |
|---|---|---|
| Claude Haiku 5.5 | High-volume, cost-sensitive work such as summaries, compactions, database queries and classification; a subagent alongside Sonnet 5.5 and Opus 5.5 on coding work | Anthropic states Sonnet 5.5 and Opus 5.5 remain the choices for complex agentic coding |
| Claude Sonnet 5.5 | A faster, lower-cost companion to Opus 5.5, for well-scoped everyday tasks, bug fixing and polished documents, slides and spreadsheets | Opus 5.5 described as for complex work that requires careful judgment |
| Claude Opus 5.5 | Long-running agentic coding and knowledge work; performs at the level of Fable 5.1 on most work | Fast mode up to 2.5 times the speed at $8/$40 |
Worked cost example
Using list prices, 10,000 requests a day, each with a 20K-token prompt and a 1K-token response, use 200 million input and 10 million output tokens. On Opus 5.5 that is $800 plus $200, $1,000 a day. On Sonnet 5.5 it is $400 plus $100, $500 a day. On Haiku 5.5, each 20K prompt stays under the 100K threshold, so it is $20 plus $5, $25 a day. If the same requests carried 150K-token prompts, Haiku 5.5 would bill $0.50 input and $2.50 output for every request, five times its lower tier. Default effort also differs, which changes thinking and output tokens per task, so measure on your own data.
What is the context window of each model?
As of 2026-11-15. From Anthropic's models overview and model pages; reasoning controls from the release notes.
| Model | Context window | Max output | Knowledge cutoff | Input | Output | Reasoning controls |
|---|---|---|---|---|---|---|
| Claude Haiku 5.5 | 1M tokens | 128K tokens (300K on Batch API, beta) | Jun 2026 (reliable and training data) | Text, image | Text | Adaptive, on by default; default effort medium |
| Claude Sonnet 5.5 | 1M tokens | 128K tokens (300K on Batch API, beta) | Jun 2026 (reliable and training data) | Text, image | Text | Adaptive, on by default; default effort high; between_tools at effort high or below |
| Claude Opus 5.5 | 1M tokens | 128K tokens | Jun 2026 (reliable and training data) | Text, image | Text | Adaptive, always on; default effort medium |
The 300K batch output option uses the output-300k-2026-03-24 beta header, as stated on the Haiku 5.5 and Sonnet 5.5 model pages; the Opus 5.5 model page checked does not state a batch output limit. Both the Haiku 5.5 and Sonnet 5.5 pages list a minimum cacheable prompt length of 512 tokens.
Capabilities and tools compared
As of 2026-11-15. Anthropic's models overview states that all current models support "text and image input, text output, multilingual capabilities, vision, and tool use". Cloud IDs are from each model's overview page.
| Model | Function calling and structured outputs | Computer use tool | Fine-tuning | Cloud IDs stated by provider |
|---|---|---|---|---|
| Claude Haiku 5.5 | Tool use supported; assistant prefill returns a 400 error | OSWorld 2.1 (offline subset) figure reported in its announcement | Not stated by Anthropic on the pages checked (2026-11-15) | Amazon Bedrock anthropic.claude-haiku-5-5; Google Cloud, Microsoft Foundry and Claude Platform on AWS claude-haiku-5-5 |
| Claude Sonnet 5.5 | Tool use supported; forced tool use returns a 400 error | OSWorld 2.1 (partial) figure reported in its announcement | Not stated by Anthropic on the pages checked (2026-11-15) | Amazon Bedrock anthropic.claude-sonnet-5-5; Google Cloud, Microsoft Foundry and Claude Platform on AWS claude-sonnet-5-5 |
| Claude Opus 5.5 | Tool use supported; tool_choice any and tool return a 400 error; auto with strict tool use advised | OSWorld 2.1 (partial) figure reported in its announcement | Not stated by Anthropic on the pages checked (2026-11-15) | Amazon Bedrock anthropic.claude-opus-5-5; Google Cloud, Microsoft Foundry and Claude Platform on AWS claude-opus-5-5 |
For agent architecture beyond model choice, see Computer-Use and Browser Agents.
Where is each model available?
| Model | Availability stated by provider |
|---|---|
| Claude Haiku 5.5 | "All platforms", including Amazon Web Services, Google Cloud and Microsoft Azure |
| Claude Sonnet 5.5 | "All platforms", including Amazon Web Services, Google Cloud and Microsoft Azure |
| Claude Opus 5.5 | "All platforms", including Amazon Web Services, Google Cloud and Microsoft Azure; fast mode on the Claude API, Microsoft Foundry and Google Cloud (Anthropic managed) |
Anthropic's deprecations page states that Amazon Bedrock and Claude Models on Google Cloud (Google managed) set their own schedules. Confirm in each cloud's catalogue.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Claude Haiku 5.5 | Announcement | System card | Not stated by Anthropic on the pages checked (2026-11-15) |
| Claude Sonnet 5.5 | Announcement | System card | Not stated by Anthropic on the pages checked (2026-11-15) |
| Claude Opus 5.5 | Announcement | System card | Not stated by Anthropic on the pages checked (2026-11-15) |
System card links are as given in each announcement; this page does not summarise their contents.
Benchmarks
As of 2026-11-15. Provider-reported figures from Anthropic's announcements. Only benchmarks reported for all three models are shown; the source column names the announcement each figure came from.
| Benchmark | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 | Source |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 39.2% | 70.6% | 66.4% (xhigh) | Haiku 5.5 and Sonnet 5.5 announcements |
| FrontierCode 1.1 (Main) | 46.4% | 52.1% (xhigh) | 54.4% | Haiku 5.5 and Sonnet 5.5 announcements |
| Humanity's Last Exam (with tools) | 57.4% | 64.5% | 67.7% | Haiku 5.5 and Sonnet 5.5 announcements |
| GDPval-AA v2.1 | 1620 | 1844 (1840 in the Haiku 5.5 announcement) | 1846 | Sonnet 5.5 and Haiku 5.5 announcements |
| Chartography (no tools) | 46.4% | 61.6% | 64.4% | Haiku 5.5 and Sonnet 5.5 announcements |
Independent leaderboard results were not checked for this page. Anthropic states that Artificial Analysis ran GDPval-AA on a pre-release deployment with a since-fixed structured-output bug, and that production safeguards may have lowered Opus 5.5 scores. OSWorld figures use different subsets across announcements (partial and offline subset), so they are not placed in one row.
Deprecations and retirements
As of 2026-11-15. Source: Anthropic model deprecations page. Anthropic states it gives at least 60 days' notice before retiring publicly released models.
| Model | API ID | Notice or deprecation date | Retirement or shutdown date | Stated replacement |
|---|---|---|---|---|
| Claude Haiku 5.5 | claude-haiku-5-5 | Active | Not sooner than 2027-10-07 | None stated |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | Active | Not sooner than 2027-09-28 | None stated |
| Claude Opus 5.5 | claude-opus-5-5 | Active | Not sooner than 2027-09-22 | None stated |
For lifecycle planning across providers, see the LLM Deprecation and Retirement Tracker and the AI Model Retirement Guide.
Data handling and security
Anthropic's zero data retention page designates only the Fable and Mythos 5.x models as Covered Models requiring 30-day retention; none of the three 5.5 tiers is on that list. The Opus 5.5 and Sonnet 5.5 announcements state zero data retention is available; the Haiku 5.5 announcement does not mention it, so confirm with Anthropic. Flagged content may be retained for up to 2 years under any arrangement. Sonnet 5.5 thinking blocks, and Opus 5.5 thinking for API accounts created on or after 31 August 2026, are tied to the originating account. All three carry cyber and biology safeguards; the Cyber and Life Sciences Verification Programs offer broader access.
The credential is the shared risk: one Claude API key reaches every tier the workspace allows, so a leaked key used on Opus 5.5 costs 40 times more per input token than on Haiku 5.5's lower tier. See the Anthropic API Keys Security Guide and the LLMjacking Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| List price per token | From $0.10/$0.50 (Haiku 5.5), $2/$10 (Sonnet 5.5), $4/$20 (Opus 5.5) |
| Prompts above 100K tokens | Haiku 5.5 moves to $0.50/$2.50; Sonnet 5.5 and Opus 5.5 keep standard rates |
| Latency rating | Fastest, Fast and Moderate respectively; fast mode listed for Opus 5.5 only |
| Default effort | Medium (Haiku 5.5, Opus 5.5); high (Sonnet 5.5) |
| Switching thinking off | Sonnet 5.5 via between_tools; not possible on Opus 5.5 |
| Assistant prefill | Returns a 400 error on Haiku 5.5 (per its Haiku 4.5 migration note) |
| Lifecycle runway | Earliest retirement 2027-10-07, 2027-09-28 and 2027-09-22 |
For cross-provider context, see Frontier AI Models Compared.
Practitioner checklist
- Pin each workload to an explicit 5.5 API ID and record the tier in the AI inventory.
- Track prompt length on Haiku 5.5 traffic and alert when requests cross 100,000 tokens.
- Remove temperature, top_p, top_k and assistant prefill from code moving to Haiku 5.5 or Sonnet 5.5.
- Replace forced tool_choice with auto plus strict tool use before routing to Opus 5.5 or Sonnet 5.5.
- Set effort explicitly rather than relying on differing tier defaults when comparing cost per task.
- Log which model answered security-related prompts, since safeguards can fall back to Opus 4.8 or Sonnet 5.
- Restrict which tiers each API key or workspace can call, and set spend limits.
Standards and references
- Anthropic: Models overview (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Model deprecations (accessed 2026-11-15)
- Anthropic: Claude Platform release notes (accessed 2026-11-15)
- Anthropic: Claude Haiku 5.5 model overview (accessed 2026-11-15)
- Anthropic: Claude Sonnet 5.5 model overview (accessed 2026-11-15)
- Anthropic: Claude Opus 5.5 model overview (accessed 2026-11-15)
- Anthropic: Zero data retention (accessed 2026-11-15)
- Anthropic: Introducing Claude Haiku 5.5 (accessed 2026-11-15)
- Anthropic: Introducing Claude Sonnet 5.5 (accessed 2026-11-15)
- Anthropic: Introducing Claude Opus 5.5 (accessed 2026-11-15)
Related NHI Mgmt Group resources: Anthropic Models, Claude Haiku Versions Compared, Claude Sonnet Versions Compared, LLM Pricing Tracker, Claude for Enterprise Security Guide.