Which Claude Model? Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5 vs Fable 5.1

The lineup on one screen

List prices from the pricing page and the models overview, read on 2026-10-21. Sonnet 5.5 shipped on 2026-09-28, Haiku 5.5 on 2026-10-07.

What they share matters as much: a 1M-token context, 128K output tokens, text and image input, a June 2026 knowledge cutoff, and the same rate limits per tier for Haiku, Sonnet and Opus 5.5. So the choice is price, latency and how hard the task is, not which model can physically fit the job. Anthropic's own guidance is to start with Opus 5.5 for most workloads; the request basics are in Messages API and the Claude API tutorial.

Three prices people misread

Cache hits are not a flat 0.1x any more. Opus 5.5 and Sonnet 5.5 read cached tokens at 0.05x, Fable 5.1 at 0.025x, and Haiku 5.5 at 0.1x. Writes still cost 1.25x input for the 5-minute cache and 2x for the 1-hour one, so a big cached prefix is cheapest, relative to its own input price, on the models people assume are expensive. The mechanics are in Prompt Caching.

The step is per request and counts every input token, cached or not. The Batch API halves both sides on all four models (Haiku 5.5: $0.05 / $0.25), and the discounts stack with caching. Opus 5.5 also has a fast mode at $8 / $40 for latency-bound work.

Derived from the step 03 ticket: the old models see the same text as about 1/1.3 as many tokens. Recount with count_tokens on your own prompts; the increase depends on the content.

Worked example: one support inbox, four bills

Volume

200,000 tickets a month, about 4.6 a minute, so a 5-minute cache stays warm and writes are a rounding error.

Prefix

8,000 tokens of system prompt, tools and policy, a cache hit on every call.

Per ticket

2,000 fresh input tokens (the ticket and history) and 600 output tokens, thinking included.

Opus 5.5 per ticket is (8,000 × $0.20 + 2,000 × $4 + 600 × $20) / 1M = $0.0216, or $4,320.00 a month. Output dominates every row: 600 tokens are 56% of the Opus bill, which is why effort (how much the model thinks) is a bigger lever than the input price. Effort and its per-request controls are in Agent Cost Control.

All four use the same tokenizer, so the token counts are comparable across rows. Your counts will differ: measure prefix, ticket and output on a real sample.

The routing pattern: small model sorts, big model reasons

The router costs $49 a month and the hard 30% still costs 91% of the routed bill, so the escalation share is the number to watch: every 10 points of tickets moved from routine to hard adds $420.40 a month. Sending the hard share to Sonnet 5.5 instead would cost $778.20; whether that holds quality is an eval question, not a pricing one.

import json
import anthropic

client = anthropic.Anthropic()
ROUTER = "Label the ticket routine or hard. Hard: refunds, legal, outages."
TIER = {"type": "object", "additionalProperties": False, "required": ["tier"],
        "properties": {"tier": {"type": "string", "enum": ["routine", "hard"]}}}

def route(ticket: str) -> str:
    r = client.messages.create(
        model="claude-haiku-5-5", max_tokens=1024, system=ROUTER,
        cache_control={"type": "ephemeral"},
        output_config={"effort": "low",
                       "format": {"type": "json_schema", "schema": TIER}},
        messages=[{"role": "user", "content": ticket}])
    if r.stop_reason != "end_turn":
        return "hard"  # refusal or cut off: escalate, never guess
    text = next(b.text for b in r.content if b.type == "text")
    return json.loads(text)["tier"]

The router reads blocks by type (a thinking block can come first), branches on stop_reason, and escalates when unsure, since a wrong "routine" costs a bad reply while a wrong "hard" costs $0.0210. Haiku 5.5 runs safety classifiers with no server-side fallback, so handle refusal. In production the router prompt carries a few thousand tokens of examples: under 512 tokens nothing is cached. Thresholds, cascades and failover are in Model Routing.

Ran against a mocked API with anthropic 1.17.0, the current Python SDK; no live key was used.

Effort, thinking and the breaking changes

Forced tool use (tool_choice of any or a named tool) returns an error on Sonnet 5.5, Opus 5.5 and Fable 5.1. Haiku 5.5 accepts it but skips thinking. Use auto with a prompt and strict: true tools, or structured outputs as in the router above. On all four, temperature, top_p and top_k at non-default values, budget_tokens and assistant prefill return a 400. Steer depth with output_config.effort; how it trades quality for tokens is in Adaptive Thinking and Effort.

If you are still on a 4.x model

Status as of 2026-10-26, from the model deprecations page. Requests to a retired model fail; Anthropic gives at least 60 days' notice.

1. Swap the IDNo date suffix on the 5.5 models.
2. Recount tokensAbout 30% more for the same text; revisit max_tokens and budgets.
3. Remove what errorsSampling parameters, budget_tokens, prefill; forced tools on Sonnet 5.5.
4. Read blocks by typeThinking is on by default and can come first.
5. Re-run your evalsThen re-price with the new counts.

Haiku 4.5 to 5.5 has two quieter changes: the minimum cacheable prompt drops from 4,096 tokens to 512, and Priority Tier is not supported on Haiku 5.5, so plan capacity separately if you had a commitment. When the API stops being the cheapest option at your volume is in Self-Host vs API.

A decision you can defend

The order that keeps this honest: write the eval set first, run it on Opus 5.5, try the cheaper model on each high-volume step, and only climb to Fable 5.1 for the steps that still fail. Price every candidate with your own token counts. Cost and latency as one budget is in Cost vs Latency.

Next: Model Routing — classifiers, cascades and failover, so the router in step 04 survives production.