Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 4.5 compared

Anthropic's current lineup has four models: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5. All four take text and image input, produce text, and support tool use and vision¹. Beyond that they differ in almost every published dimension: list rate, cache-read rate, context window, output limit, thinking mode, knowledge cutoff and how long Anthropic commits to keep them.

This article sets out those published facts, read on 2026-09-29, and a method for deciding which model runs which task. It contains no quality ranking; Anthropic's own descriptions are quoted as descriptions, not as verdicts.

How Anthropic describes each model

Anthropic's overview gives each model a one-line purpose and a comparative latency¹:

• Claude Fable 5.1 — for demanding reasoning and long-horizon agentic work. Latency: slower.
• Claude Opus 5.5 — for long-running agentic coding and knowledge work. Latency: moderate.
• Claude Sonnet 5.5 — described as the best combination of speed and intelligence. Latency: fast.
• Claude Haiku 4.5 — the fastest model, with near-frontier intelligence by Anthropic's account. Latency: fastest.

Anthropic also gives a default: start with Opus 5.5 for most workloads, and use Fable 5.1 when evaluations on Opus 5.5 at higher effort still fall short¹. Treat that as the vendor's suggested starting point for your own evaluation.

List rates

Per million tokens, Fable 5.1 lists $10 input and $50 output, Opus 5.5 $4 and $20, Sonnet 5.5 $2 and $10, and Haiku 4.5 $1 and $5². The catalog entries INFRO estimates against are on the Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5 model pages.

The Batch API halves both rates for all four: $5/$25, $2/$10, $1/$5 and $0.50/$2.50². Opus 5.5 also has a fast mode in research preview at $8 input and $40 output, on the Claude API only².

Cache reads are not uniform

Most Claude models bill a cache hit at 0.1x the input rate. Two current models do not: Fable 5.1 bills cache hits at 0.025x input ($0.25 per million) and Opus 5.5 at 0.05x ($0.20 per million)². Sonnet 5.5 is at the standard 0.1x, also $0.20, and Haiku 4.5 at $0.10².

So for a workload dominated by cache reads — an agent resending a large, stable context — the input side of the bill is similar across Fable 5.1, Opus 5.5 and Sonnet 5.5. The difference then sits mostly in uncached input and output.

Cache writes follow Anthropic's multipliers: 1.25x input for a 5-minute write and 2x for a 1-hour write, so $5 and $8 per million on Opus 5.5, for example².

Limits, thinking and knowledge

• Context and output. Fable 5.1, Opus 5.5 and Sonnet 5.5 have a 1M-token window and 128K output on the synchronous Messages API; Haiku 4.5 has 200K and 64K¹. On the Message Batches API, Opus 5.5 and Sonnet 5.5 support up to 300K output tokens with a beta header¹.
• Long prompts. Claude 4.6 and later models include the full 1M window at standard pricing²; there is no long-prompt surcharge on the three 1M models.
• Thinking. Fable 5.1 and Opus 5.5 use adaptive thinking that is always on; Sonnet 5.5 uses adaptive thinking; Haiku 4.5 uses extended thinking. Default effort is high on Fable 5.1 and Sonnet 5.5, medium on Opus 5.5, and the effort parameter is not supported on Haiku 4.5¹.
• Knowledge. Reliable knowledge runs to June 2026 on the three newest models and to February 2025 on Haiku 4.5¹.
• Structured outputs. Generally available on all four³.
• Tokenizer. Claude 4.7 and later models use a tokenizer that produces approximately 30% more tokens for the same text². When comparing Haiku 4.5 against the newer models, compare measured tokens per task, not rates alone.

Overheads that differ by model

Two published overheads change the token count of the same request from one Claude model to the next, which matters most for small, frequent calls.

Tool definitions. When a request includes tools, Anthropic adds a tool-use system prompt whose size depends on the model: 286 tokens on Opus 5.5 and Sonnet 5.5, and 496 tokens on Haiku 4.5 with tool_choice set to auto or none, or 588 with any or tool². That is before your own tool names, descriptions and schemas, which are billed as input too. For a classification call of a few hundred tokens with one tool attached, the fixed overhead can be a large share of the input.

Tokenizer. The newer tokenizer on Claude 4.7 and later models produces approximately 30% more tokens for the same text, with the exact increase depending on content². Haiku 4.5 predates it. A rate comparison between Haiku 4.5 and Sonnet 5.5 therefore understates the difference in tokens counted for the same prompt; measure tokens per task on your own inputs rather than converting by rule of thumb.

The same models on other clouds

Anthropic publishes model ids for Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS for all four models — for example anthropic.claude-opus-5-5 on Bedrock and claude-haiku-4-5@20251001 on Google Cloud¹. The commercial terms differ by route:

• Partner-operated clouds (Bedrock, Google Cloud) invoice you themselves and set their own prices and lifecycle dates; their regional and multi-region endpoints carry a 10% premium over global ones².
• Anthropic-operated platforms billed through your cloud account rate usage at the standard per-model rates².
• US-only inference on the Claude API with inference_geo: "us" applies a 1.1x multiplier to every token category².
• Fast mode for Opus 5.5 is available on the Claude API only².

If your organization already buys Claude through a cloud commitment, the list rate on this page may not be what you pay. Compare against the rate your contract sets.

Retirement commitments

Anthropic commits not to retire Fable 5.1 sooner than September 1, 2027, Opus 5.5 sooner than September 22, 2027, Sonnet 5.5 sooner than September 28, 2027, and Haiku 4.5 sooner than October 15, 2026¹. The Haiku 4.5 date is a floor, not an announced retirement, but it is two and a half weeks from the date of this article. Any workload on Haiku 4.5 should have a tested successor ready.

A method for choosing

1. Group work by task type, not by application: classification, extraction, support drafting, research agents and code changes each get their own decision.
2. Evaluate upward. Start with the lowest-cost model that meets the hard constraints (window, output limit, latency), then move up only when it fails your check. Anthropic's own guidance on its pricing page follows the same shape: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning².
3. Test effort as well as model. Opus 5.5 at higher effort and Fable 5.1 at its default are different operating points; Anthropic suggests exactly that comparison¹.
4. Measure cost per successful task at the effort you will ship, with caching on.
5. Record the second choice. For every task tier, keep a tested alternative; it is your failover target and your migration path.

Routing Claude models with INFRO

Connect your Anthropic account as a provider connection; your provider bills you at your rates. INFRO's Bedrock connection speaks Chat Completions, which Bedrock does not offer for Claude, so Claude bought through AWS is not routable through INFRO yet. Anthropic notes that partner clouds price Claude independently²; if you have a negotiated rate, set it under Prices (billing).

Then put each task tier behind a route:

The first route fails over to the same model on another account, which changes infrastructure and not behaviour. The other two fail over to a different model, which changes behaviour, so list only a model you have evaluated on that task (failover). When Haiku 4.5 is retired, ticket-tags becomes a one-line edit: move its tested successor to first place.

Use a model allowlist to keep anthropic/claude-fable-5.1 to the projects approved for it; a route target that fails the allowlist is skipped. For Anthropic prompt caching, put cache_control on the text part that ends your stable prefix; INFRO passes it through, and usage.prompt_tokens_details.cached_tokens shows whether the cache was read (prompt caching). GET /v1/usage?group_by=route then gives estimated provider cost, error rate and latency per route, so the cost of each tier decision is visible month to month (usage analytics).