Short answer: Anthropic’s model overview recommends starting with Claude Opus 5.5 if you’re unsure which model to use. Consider Fable 5.1 for demanding reasoning and long-horizon agent work; Sonnet 5.5 for a speed-and-intelligence balance; and Haiku 5.5 for high-volume, latency-sensitive tasks. These are Anthropic’s positioning cues, not a guarantee of the best result for your workload. Run representative evaluations before choosing.
As of November 8, 2026, Anthropic’s model overview lists these four models with Claude API IDs. The table gives their published descriptions, context windows, maximum output sizes, and token rates where available. Haiku 5.5’s total-cost comparison is partial: its lower-tier output rate and the billing treatment when a prompt exceeds the threshold are not established here. Model listings don’t guarantee that a model is enabled for every account or available through every platform; check the relevant model page and your account before deployment. (Anthropic model overview; Anthropic pricing)
Which Anthropic models should you shortlist?
| Model | Claude API ID | Anthropic’s stated fit | Context window | Maximum output | Standard price, USD per million tokens (input / output) |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | claude-haiku-5-5 |
High-volume, latency-sensitive work such as classification, extraction, and routing | 1M tokens | 128K tokens | Prompts up to 100,000 tokens: input $0.10; output rate not established. Prompts over 100,000 tokens: input $0.50; output $2.50; threshold billing basis not established. |
| Claude Sonnet 5.5 | claude-sonnet-5-5 |
A balance of speed and intelligence | 1M tokens | 128K tokens | $2 / $10 |
| Claude Opus 5.5 | claude-opus-5-5 |
Long-running agentic coding and knowledge work; Anthropic suggests starting here for most workloads | 1M tokens | 128K tokens | $4 / $20 |
| Claude Fable 5.1 | claude-fable-5-1 |
Demanding reasoning and long-horizon agentic work, or when higher-effort Opus evaluations fall short | 1M tokens | 128K tokens | $10 / $50 |
The identifiers above are the Claude API IDs shown in Anthropic’s overview, not IDs for other providers. Anthropic says each model page lists the platforms where that model is available. The overview also gives “not sooner than” retirement dates: September 1, 2027 for Fable 5.1; September 22, 2027 for Opus 5.5; September 28, 2027 for Sonnet 5.5; and October 7, 2027 for Haiku 5.5. These dates are not a promise of account-specific access. Verify the current model page and your account’s access before relying on a model. (Anthropic model overview)
A practical first choice by workload
- Classification, extraction, routing, or other high-volume work: Evaluate Haiku 5.5. Its published fit is high-volume, latency-sensitive tasks. Confirm its complete rates and applicable threshold billing before estimating total cost.
- General application tasks where speed and capability both matter: Compare Sonnet 5.5 with Opus 5.5 using the same representative examples. Anthropic describes Sonnet as a speed-and-intelligence balance and recommends Opus as a starting point when you’re unsure.
- Long-running coding agents or knowledge work: Evaluate Opus 5.5. For demanding reasoning or long-horizon agents, include Fable 5.1 if your higher-effort Opus evaluations still fall short.
- Workflows that depend on a particular input, tool, or feature: Confirm that the model and platform you intend to use support it. A model’s high-level description alone doesn’t establish a specific API feature or request format.
Anthropic’s overview says the four models support text and image input, text output, multilingual capabilities, vision, and tool use. It lists a 1-million-token context window and 128,000-token maximum output for each. If those limits or a particular modality are important to your design, check the current documentation for the specific model and platform before deployment. (Anthropic model overview)
What do Claude API models cost?
Anthropic lists rates in USD per million tokens (MTok), with separate input and output rates. For Haiku 5.5, the listed input rate is $0.10/MTok for prompts up to 100,000 tokens; for prompts over 100,000 tokens, the listed rates are $0.50/MTok input and $2.50/MTok output. The lower-tier output rate and whether the higher rates apply to the whole request or only excess tokens are not established here, so a complete Haiku request-cost comparison cannot be calculated from these figures alone. (Anthropic pricing)
Worked example: For an uncached request with 100,000 input tokens and 10,000 output tokens, the examples below use the published standard rates for Sonnet, Opus, and Fable. No Haiku total is shown because its lower-tier output rate and threshold billing treatment are not established here.
| Model | Example calculation | Approximate cost per request |
|---|---|---|
| Sonnet 5.5 | (0.1 × $2) + (0.01 × $10) | $0.30 |
| Opus 5.5 | (0.1 × $4) + (0.01 × $20) | $0.60 |
| Fable 5.1 | (0.1 × $10) + (0.01 × $50) | $1.50 |
These are illustrations, not usage forecasts. They assume the stated token counts, standard uncached rates, and no other charges. Estimate your own cost from measured input and output tokens and the pricing conditions that apply. (Anthropic pricing)
Anthropic’s announcement archive also makes qualified comparisons: it says Sonnet 5.5 runs “30%+ faster” and costs “up to 30% less” than Sonnet 5 “for most work,” and that Opus 5.5 costs 40% less than Opus 5 on “typical workloads.” Those are Anthropic’s announcement claims, not exact API rates or guaranteed savings for your application. Use published rates and your own usage to estimate costs; don’t convert those percentages into a universal per-request saving. (Anthropic models archive)
Anthropic also says its tokenizer for Claude 4.7 and later produces approximately 30% more tokens for the same text, with the increase depending on the content and workload. When comparing estimated costs for the same prompts, measure token counts on the model you plan to use rather than assuming identical tokenization. (Anthropic pricing)
How to choose a model for production
Use a small evaluation set based on real or representative inputs before committing to a model:
- Define the acceptance bar. Specify what counts as a correct answer, an acceptable failure, and an unacceptable failure. Include any latency or output-format requirements.
- Run the same cases through a shortlist. Start with a candidate whose published fit matches the task, then compare it with a plausible alternative—for example, Haiku and Sonnet for high-volume extraction, or Sonnet and Opus for a multi-step coding workflow.
- Measure what affects the decision. Record task quality against your rubric, latency under comparable conditions, and billed input and output tokens. Keep prompts, settings, and test cases consistent.
- Check features and access. Confirm the exact API ID, model and platform availability, input support, tool requirements, limits, and retirement information in current official documentation for your intended deployment.
- Estimate cost from observed usage. Apply the relevant input and output rates to the token counts you measured. Include caching or other pricing features only if you’ve confirmed they apply to your implementation.
- Retest before changing model IDs. Anthropic documents model IDs as pinned versions: an existing ID does not silently change to new model weights. Treat a model change as something to evaluate and roll out deliberately. (Anthropic model ID and versioning guide)
Keep API availability distinct from announcements and subscriptions
Anthropic’s announcements and system-card pages can help explain model positioning, launch history, and documented evaluations. They are not substitutes for the current API model overview, model-specific availability information, or pricing documentation. A model appearing in an announcement or system-card index does not, by itself, confirm that your production API account can call it. (Anthropic models archive; Anthropic model overview)
Claude app subscriptions and API usage are different purchasing routes. The prices in this guide are API token rates, not subscription fees. If you deploy through a cloud provider rather than the Anthropic-operated Claude API, check that provider’s model availability and billing terms; identifiers and pricing may differ. (Anthropic pricing)
Bottom line: shortlist by workload—Haiku for high-volume, latency-sensitive tasks; Sonnet for a speed-and-intelligence balance; Opus as Anthropic’s suggested starting point for most workloads and for agentic coding; and Fable for especially demanding reasoning or long-horizon agents. Then verify the model ID, access, feature support, quality, latency, and total token cost against your own production requirements.
Sources
Models overview — Claude Platform Docs
Pricing — Claude Platform Docs