If you are unsure where to start, use Claude Opus 5.5 to establish a quality baseline. Evaluate Fable 5.1 for demanding reasoning and long-running agents, compare Sonnet 5 for everyday development and writing, and try Haiku 4.5 first for low-latency, high-volume tasks. Choose based on acceptance rate, latency, and total cost on your own work, not on model names as if they formed a universal leaderboard.
This guide covers Anthropic’s general-purpose Claude API models. Claude subscriptions, Claude Code plans, and billing through AWS, Google Cloud, or other platforms may differ. Model availability and prices change; the information here was checked against Anthropic’s official documentation on September 25, 2026.
Current main models: capabilities, context, and standard prices
| Model | API ID | Official positioning | Context / max output | Input / output price (USD per million tokens) |
|---|---|---|---|---|
| Claude Fable 5.1 | claude-fable-5-1 |
Demanding reasoning and long-horizon agents | 1M / 128K | $10 / $50 |
| Claude Opus 5.5 | claude-opus-5-5 |
Long-running agentic coding and knowledge work; Anthropic’s suggested starting point for most workloads | 1M / 128K | $4 / $20 |
| Claude Sonnet 5 | claude-sonnet-5 |
A balance of speed and intelligence | 1M / 128K | $2 / $10 |
| Claude Haiku 4.5 | claude-haiku-4-5 |
Fastest, with near-frontier intelligence | 200K / 64K | $1 / $5 |
These are base token prices for the first-party Anthropic API. Prompt caching, the Batch API, fast mode, data residency, and cloud platforms can change the bill. Check the current rates for your access channel before making a budget. Models overview · API pricing
Fable 5.1, Opus 5.5, Sonnet 5, and Haiku 4.5 all support text and image input, text output, multilingual capabilities, vision, and tool use. They differ mainly in task ceiling, speed, cost, and context capacity, rather than in a simple yes-or-no list of features.
How to divide work among Opus, Sonnet, and Haiku
Opus 5.5: establish a useful quality baseline
Anthropic currently recommends starting with Opus 5.5 when you are unsure. It is positioned for long-running agentic coding and knowledge work. Its standard API price is $4 per million input tokens and $20 per million output tokens. A 1M-token context window and 128K output limit suit work that must continue across files and multiple steps.
Try it first on complex repository changes, cross-tool investigations, synthesis across long documents, and multi-step business workflows. It does not have to be your cheapest default: when a task is clearly scoped and easy to validate, compare Sonnet or Haiku on the same inputs.
Fable 5.1: move up only when a difficult task needs it
Fable 5.1 is for demanding reasoning and long-horizon agentic work. Its input and output rates are 2.5 times those of Opus 5.5. Add it to an evaluation when Opus 5.5 at higher effort still falls short of the acceptance criteria, or when the task value justifies the additional cost.
Do not infer that it is automatically better for every coding, research, or long-context task from its name alone. Record which failures, rework, or manual checks it actually reduces, and compare those gains with the extra cost.
Sonnet 5: compare it with Opus on everyday work
Sonnet 5 costs $2 per million input tokens and $10 per million output tokens. Anthropic positions it as a combination of speed and intelligence. It is a good candidate for routine code changes, document work, analysis, ordinary tool calls, and tasks sensitive to latency.
Its unit price is half of Opus 5.5’s, but that does not mean every task will cost exactly half as much. Models can generate different amounts of thinking tokens, make different numbers of tool calls and retries, and require different amounts of manual rework. Compare them on real workloads.
Haiku 4.5: use it for well-scoped, high-throughput requests
Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens. Anthropic positions it as the fastest current model. Its context window is 200K tokens and its maximum output is 64K, both smaller than those of the other three main general-purpose models.
Classification, field extraction, short summaries, format conversion, simple question answering, and batch routing are useful starting points. Provide a clear output schema and validation rules. For long-horizon planning, difficult judgment, or multi-turn tool coordination, define an escalation path to Sonnet or Opus.
Costs beyond the price table
Prompt caching
Anthropic’s prompt caching reduces the cost of reading repeated context. Standard rules include a 5-minute cache write at 1.25 times the base input price, a 1-hour write at 2 times the base input price, and cache reads generally at 0.1 times the base input price. Reads are cheaper for Fable 5.1 and Opus 5.5: 0.025 and 0.05 times their base input rates, respectively. Whether caching saves money depends on the write cost and subsequent cache hits, so calculate it for each model. Caching and pricing details
Batch API
The Batch API is for requests that can be processed asynchronously in large volumes. It offers a 50% discount on both input and output tokens. It is not suitable when an interactive request must return immediately, but it can lower costs for offline classification, archiving, extraction, and evaluation.
Opus fast mode
As of the date checked, fast mode is available for Opus 5.5, Opus 5, and Opus 4.8 at higher rates. Opus 5.5 fast mode costs $8 per million input tokens and $40 per million output tokens. Opus 5 and Opus 4.8 cost $10 / $50. Fast mode is available only on the first-party Claude API; do not assume it is available on Claude Platform on AWS or other cloud platforms. Fast mode details
What to do with older and restricted models
Older models such as Opus 4.8, Opus 4.7, and Sonnet 4.6 may still appear in existing configurations and transition workflows. Before migrating, check the exact API ID, retirement date, context limits, prompt caching, and tokenizer behavior. Do not replace a model name and immediately switch all production traffic. Evaluate the old and new versions on the same set of sanitized examples.
Mythos and other restricted-preview models are not included in the general availability comparison. Evaluate them only if your account and platform have approved access. A model string in a news story or third-party interface does not mean your API project can call it.
Choose a default model with your own tasks
Build a sanitized sample set that represents daily work, including successful requests, edge cases, past failures, tool errors, and long-context tasks. Keep prompts, tool definitions, output limits, and acceptance rules fixed. Compare Opus 5.5 and Sonnet 5 first; add Fable 5.1 only if the hardest examples still fail. Test Haiku 4.5 separately on simple, high-volume requests.
Track these measures:
| Measure | What to record |
|---|---|
| Task quality | Whether it passes the same acceptance criteria, and what kinds of errors occur |
| Actual cost | Input, output, cache, retries, and extra tool charges |
| Completion time | Time to first token and end-to-end time |
| Manual rework | Time spent reviewing, editing, and rerunning |
| Tool reliability | Whether calls are correct and failures recover cleanly |
Set a minimum quality bar first, then compare total cost for tasks that pass. A practical routing policy might use Haiku for classification, Sonnet for routine work, Opus for difficult jobs, and Fable only for a small number of high-value, long-running tasks. A product does not have to use one model for every request.
Earlier Claude model comparisons on this site
The earlier guide to choosing Claude Fable, Opus, Sonnet, and Haiku, along with the older draft comparing Opus 4.7, Sonnet 4.6, and Haiku 4.5, has been consolidated here. This guide updates the main models, prices, and context information. The old URLs now redirect directly to this page, which is the current model-selection reference.
Official resources
- Claude Models Overview: current model recommendations, API IDs, context windows, and output limits.
- Claude API Pricing: base prices, prompt caching, Batch API, data residency, and fast mode.