Search for Claude in the Model Hub and you get more than a dozen entries across four families — Fable, Opus, Sonnet and Haiku — each with several versions. Teams usually react in one of two ways: route everything to the most expensive model, or try the cheapest one and get frustrated with the results. Neither is a strategy. This guide explains what each family is for, how their prices relate, which one fits which job, and how to switch models in Claude Code and in your own code.
What Each Family Is For
Anthropic's naming is straightforward: the family sets the capability tier, the version number tells you how recent it is. From most to least capable:
| Family | Positioning | Recommended version | Best for |
|---|---|---|---|
| Fable | Sits above Opus; the most capable Claude available | claude-fable-5-2 | The hardest reasoning, long-horizon agents, large codebases |
| Opus | Complex reasoning and long-running agents | claude-opus-5-5 | Architecture work, stubborn bugs, multi-step tasks |
| Sonnet | The balance point between capability and cost for coding and agents | claude-sonnet-5-5 | Day-to-day coding and most production calls |
| Haiku | Fast and inexpensive | claude-haiku-5-5 | High-volume classification, extraction, lightweight tasks |
Each family also keeps a few older versions (Opus, for example, still offers 5, 4.8, 4.7 and 4.6). Those exist for projects that need to pin behavior or run regression comparisons. For new work, start with the recommended version.
What They Have in Common
A lot of the spec sheet is identical across the families, so you can stop comparing these points:
- Modalities: all accept text and image input and return text
- Capabilities: extended thinking, tool use, structured output, prompt caching and batch processing are supported across the board
- Context window: every 5.x model, plus the Opus and Sonnet 4.x versions, has a 1M-token context; Haiku 4.5 is the exception at 200K
- Protocols: the same model id works over Anthropic Messages, OpenAI Chat Completions and the OpenAI Responses API
That leaves two variables that actually matter: how hard the task is, and how much each call is worth to you.
Relative Cost: Each Tier Roughly Doubles
At Model Hub prices when this article was written, the input rates of the four recommended versions compare roughly like this:
| Model | Relative input price | Output vs input price |
|---|---|---|
| Haiku 5.5 | 1 | 5x |
| Sonnet 5.5 | 4 | 5x |
| Opus 5.5 | 8 | 5x |
| Fable 5.2 | 16 | 5x |
Two practical takeaways:
- Fable costs 4x Sonnet and 16x Haiku per token. Sending every request to the top model multiplies your bill without improving most answers
- Output tokens cost five times input tokens. Asking for concise answers and capping max_tokens often saves more than switching models
Prices change as upstream providers adjust them, so treat the Model Hub as the source of truth. For turning token counts into money, see the token cost guide.
Recommendations by Scenario
| Scenario | Pick | Why |
|---|---|---|
| Everyday coding and tests in Claude Code | Sonnet 5.5 | Strong enough for most code work at half the price of Opus |
| Architecture design, cross-module refactor planning | Opus 5.5 | Deeper reasoning; getting it right once beats repeated rework |
| Problems that both Sonnet and Opus fail on | Fable 5.2 | The strongest option, reserved for where it pays off |
| Long unattended agent runs | Opus 5.5 or Fable 5.2 | Mistakes deep into a long run are expensive |
| Support bots, Q&A, content generation | Sonnet 5.5 | Consistent quality at a predictable cost |
| Classification, extraction, tagging, routing decisions | Haiku 5.5 | High throughput, little need for deep reasoning |
| Deep reasoning without changing request parameters | claude-opus-4-6-thinking | Extended thinking is on by default |
A rule that works for most teams: default to Sonnet, escalate to Opus when it gets stuck, reach for Fable only after that, and hand rule-like work to Haiku. The full tiering approach is in the multi-model routing guide.
Switching Models in Claude Code
Switch at any point in a session with /model:
/model claude-opus-5-5To set a default, export the model before launching:
export ANTHROPIC_BASE_URL=https://zivv.pro
export ANTHROPIC_AUTH_TOKEN=sk-your-key-here
export ANTHROPIC_MODEL=claude-sonnet-5-5
claudeClaude Code's built-in opus, sonnet and haiku aliases can be pointed at specific ids with ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL and ANTHROPIC_DEFAULT_HAIKU_MODEL, so /model opus resolves to the version you chose. Installation and persistent configuration are covered in the macOS and Linux setup guide and the Windows setup guide.
Switching Models in Code
All three protocols use the same model id, so changing models means changing one field. With the Anthropic SDK:
import anthropic
client = anthropic.Anthropic(base_url="https://zivv.pro", api_key="sk-your-key-here")
def ask(prompt: str, hard: bool = False) -> str:
model = "claude-opus-5-5" if hard else "claude-sonnet-5-5"
msg = client.messages.create(
model=model,
max_tokens=2048,
messages=[{"role": "user", "content": prompt}],
)
return msg.content[0].textProjects on the OpenAI SDK set base_url to https://zivv.pro/v1 and pass the same id; the Python quickstart shows both styles side by side.
Choosing a Group
When you create a key in Zivv you pick a group. Claude models are currently offered in two groups, and the difference is where the calls come from:
| Group | What it is | Use it when |
|---|---|---|
| Claude MAX | Proprietary account pool, limited to Claude Code and Claude Desktop | You only use Claude through Claude Code |
| Claude Max Full | Proprietary account pool with no client restriction | SDKs, scripts, Cursor and other tools |
Neither group may be used for model distillation. Their price multipliers differ; check the group prices on each model page. If one key serves both Claude Code and your own scripts, Claude Max Full is the simpler choice. The reasoning behind the Claude MAX group is explained on the Claude Code page.
FAQ
Q: Is Fable worth twice the price of Opus? A: On routine tasks the difference is small. On the hardest reasoning and long agent runs, Fable is more likely to get it right in one pass. Start with Opus and switch to Fable for tasks that keep failing, then compare.
Q: Should I ever use the older versions? A: Only when you need to reproduce earlier behavior or are running model regression tests. New projects should use the recommended versions.
Q: What are model ids with a date, like claude-haiku-4-5-20251001? A: Fixed snapshots. Their behavior does not change with later updates, which suits production tasks that need consistent output.
Q: How do I see which models my key can use? A: Call GET /v1/models with the key to list them, or check the group prices on a model page. Calling a model outside your key's group returns an error; the 403 troubleshooting guide walks through it.
Wrapping Up
You do not need the newest, strongest Claude for everything. Tiering by task difficulty gives the best value: Sonnet as the default, Opus for hard problems, Fable as the last resort, and Haiku for volume. Review your usage monthly to confirm the expensive models are earning their keep. Sign up for Zivv and one key gives you every Claude model — switch with /model in Claude Code and find the mix that fits your workflow.