Claude API Pricing 2026: Opus 5.5 vs Fable 5.1 vs Sonnet 5.5 vs Haiku 5.5

Claude API pricing is quoted per million tokens, which sounds cheap until you realize how fast tokens add up in real workflows. The headline rates are easy to find; what is hard is translating them into "what will this actually cost me per month?" This guide does that translation with real use-case math, then shows the two levers that cut the bill the most. Every rate below was read on Anthropic's pricing and model pages on October 23, 2026.

Current Claude API rates (October 2026)

• Claude Opus 5.5: $4 input / $20 output per million tokens, with cache reads at $0.20. Its fast mode ($8 / $40) left research preview on October 15, 2026.
• Claude Fable 5.1: $10 input / $50 output per million tokens, built for demanding reasoning and long-horizon agentic work.
• Claude Sonnet 5.5: $2 input / $10 output, the same list price as Sonnet 5. Anthropic cut its cache-read price from $0.20 to $0.10 on October 7, 2026.
• Claude Haiku 5.5: $0.10 / $0.50 for prompts up to 100K tokens and $0.50 / $2.50 above that, so long prompts cost five times as much per token.
• Legacy models are still available: Sonnet 5 at $2 / $10, Haiku 4.5 at $1 / $5, and Opus 5 at $5 / $25, which costs more than Opus 5.5.
• You pay separately for input tokens (your prompts plus context) and output tokens (Claude's responses).

On Haiku 4.5: the October 15, 2026 date in Anthropic's deprecation table was a "not sooner than" date, not a shutdown. As of October 23, 2026, Haiku 4.5 is still listed as active with no deprecation notice, and Anthropic's policy is to give at least 60 days' notice before retiring a publicly released model. Haiku 5.5 costs a tenth as much on prompts up to 100K tokens, so most Haiku 4.5 workloads have a reason to move anyway.

These are standard API prices, not a universal monthly bill. Prompt caching, batch processing, model routing, context size, and output length all change the effective cost. Anthropic also notes that models from Claude 4.7 onward use a tokenizer that produces about 30% more tokens for the same text, so re-measure token counts when you move off an older model.

Real monthly cost by use case

Use these as workload patterns, not bill forecasts. Your actual numbers depend on prompt size, output length, how much context you re-send, caching, and the model mix you route to.

• Solo developer using Claude for code review and debugging: start with Sonnet 5.5, measure a week of prompt and output volume, then test whether Opus 5.5 earns its double price on your hardest tasks.
• Power user running an agentic coding loop most days: use Opus 5.5 or Sonnet 5.5 by default, then measure how often difficult tasks justify Fable 5.1.
• Small team running a customer-facing feature on the API: route routine requests to Haiku 5.5, then selectively escalate to Sonnet 5.5 or Opus 5.5.
• High-volume production app (Haiku-first with caching): cost-per-request matters far more than headline rate, this is where caching, batch, and staying under the 100K prompt line pay off.

Worked example: 10M input and 2M output tokens a month

The same hypothetical workload on each model at standard rates, with no caching or batch discount and Haiku prompts under 100K tokens. Swap in your own token counts: monthly cost is input millions times the input rate plus output millions times the output rate.

• Haiku 5.5: $1 input + $1 output = $2 a month.
• Sonnet 5.5: $20 input + $20 output = $40 a month.
• Opus 5.5: $40 input + $40 output = $80 a month.
• Fable 5.1: $100 input + $100 output = $200 a month.
• Through the Batch API (50% off), the same workload costs $1, $20, $40, and $100.
• For comparison, legacy Haiku 4.5 costs $10 + $10 = $20 for the same tokens, ten times Haiku 5.5.

The two levers that cut your bill most

1. Prompt caching (cached reads cost a fraction of fresh input)

If you re-send the same system prompt, instructions, or document context across many requests, prompt caching bills that repeated context as cache reads instead of fresh input. On Opus 5.5, cache reads cost $0.20 per million tokens against $4 for uncached input, a 95% discount on the reused part. For RAG apps, coding agents, and anything with a large fixed prompt, this is the single biggest lever, often larger than switching models.

2. Batch processing (50% off)

For workloads that do not need a real-time response, overnight processing, bulk classification, data enrichment, the Batch API is 50% cheaper across all models. If latency does not matter, batching halves the cost with zero quality loss.

Model choice is a cost decision

The most common waste is running Fable or Opus on tasks Haiku or Sonnet would handle fine. Per million output tokens, Fable 5.1 costs $50, Opus 5.5 $20, Sonnet 5.5 $10, and Haiku 5.5 $0.50, so Fable output costs 100 times Haiku 5.5 output and Opus 5.5 costs 40 times. Pick a default deliberately (Sonnet 5.5 costs half as much as Opus 5.5), reserve Fable 5.1 for genuinely hard reasoning or long-horizon agents, then route routine work down.

Which Claude model should you pay for?

• Choose Haiku 5.5 if the job is classification, extraction, routing, or short summaries at volume, and your prompts stay under 100K tokens.
• Choose Sonnet 5.5 if you want one capable default at $2 / $10; it replaces Sonnet 5 at the same list price.
• Choose Opus 5.5 if your own evaluations show it fixes failures that Sonnet 5.5 misses; it costs twice as much per token.
• Choose Fable 5.1 only for the hardest reasoning and long-horizon agent runs, at 2.5 times the Opus 5.5 rate.

FAQ

How much does the Claude API cost per million tokens?

Prices verified October 23, 2026: Opus 5.5 is $4 input / $20 output, Fable 5.1 is $10 / $50, Sonnet 5.5 is $2 / $10, and Haiku 5.5 is $0.10 / $0.50 per million tokens for prompts up to 100K tokens ($0.50 / $2.50 above that). Input and output are billed separately.

Is Claude Haiku 4.5 being retired?

Not yet. As of October 23, 2026, Anthropic still lists Haiku 4.5 as active, with a tentative retirement of "not sooner than October 15, 2026" and no deprecation notice. Anthropic says it gives at least 60 days' notice before retiring a publicly released model. Haiku 5.5 is the cheaper replacement on prompts up to 100K tokens.

Is the Claude API cheaper than a Claude subscription?

It depends on volume. Light, occasional use is often cheaper on the API because you only pay for what you use. Heavy daily use is usually cheaper on a flat-rate Pro or Max subscription. The break-even depends on your token volume, tracking your real usage is the only way to know which side you are on.

How do I reduce my Claude API bill?

In order of impact: enable prompt caching for repeated context, use the Batch API for non-urgent work, keep Haiku 5.5 prompts under 100K tokens, and route simple tasks to Haiku or Sonnet instead of Fable or Opus. Check the official pricing page for the current cache and batch rates before modelling spend.

Sources

• Claude API Docs, Pricing
• Claude model overview
• Claude model deprecations
• Claude API release notes
• Anthropic: Introducing Claude Opus 5.5