Claude API Pricing: All Models and Costs (Oct 2026)

Legacy models and pricing details

Sonnet 5 and Haiku 4.5 are still available, but they're listed as legacy models. Sonnet 5 is $2 input and $10 output per MTok. Haiku 4.5 is $1 input and $5 output per MTok, with a 200,000-token context window. Anthropic lists Haiku 4.5 for retirement no sooner than October 15, 2026, and requires advance notice before a model is retired. If you're building something that depends on a legacy model, check the current deprecation schedule first.

Sonnet 5's $2/$10 price is now its standard rate. It launched at that rate as an introductory offer, but the planned increase to $3/$15 didn't happen. Other legacy options include Fable 5 at $10/$50, Opus 5 and Opus 4.6-4.8 at $5/$25, and Sonnet 4.6 at $3/$15 per MTok. Availability and retirement timing vary by model and platform.

Two details matter when you interpret the table:

• Thinking tokens count as output. Reasoning tokens are billed at the output rate, including tokens that aren't shown in the final response. Thinking is enabled by default on the current models listed here; it can't be disabled on Fable 5.1 and Opus 5.5. The effort setting affects how many tokens a response may use.
• Token counts differ by model. Anthropic says Claude 4.7 and later models use a tokenizer that can produce around 30% more tokens for the same text. The real difference depends on what your content looks like. Compare cost per task, not just cost per token.

The tokenizer can change the apparent savings

A lower per-token price doesn't automatically translate into proportional savings. If the newer model produces more tokens for the same input, that extra token usage can erase some of the rate advantage.

For example: Sonnet 4.6 costs $3 per MTok for input, while Sonnet 5.5 costs $2. If the newer tokenizer increases the token count by 30%, the comparable input cost is roughly $2 × 1.3 = $2.60 instead of $3. That's about 13% lower, not the 33% you'd conclude from price alone. This is an estimate - your prompts may behave differently.

The same idea applies to Haiku: the move from Haiku 4.5 to Haiku 5.5 is estimated to be about 87% cheaper per task at the lower rate. Going from Opus 4.6 or 4.5 to Opus 5.5 could even come out slightly more expensive for the same text if token counts rise by 30%. If you're planning a migration, measure a few representative inputs using Anthropic's token-counting endpoint before you do the math.

Customer-support chat

Assume each conversation uses 500 input tokens and 200 output tokens, with 10,000 calls per day (so 5 million input tokens and 2 million output tokens daily).

• Haiku 5.5: (5 × $0.10) + (2 × $0.50) = $1.50 a day, or about $45 a month.
• Sonnet 5.5: (5 × $2) + (2 × $10) = $30 a day, or about $900 a month.
• Haiku 4.5: (5 × $1) + (2 × $5) = $15 a day, or about $450 a month with the same token counts.

Yes, the price gap is big - but the bill won't tell you whether Haiku is accurate enough for your customers. Run tests on real support conversations, including edge cases and escalation scenarios. A "cheaper" model that triggers more follow-up work can end up costing you more.

Code review assistant

Suppose the assistant processes 1,000 requests per day, with 3,000 input tokens and 800 output tokens per request. That's 3 million input tokens and 800,000 output tokens daily.

• Haiku 5.5: (3 × $0.10) + (0.8 × $0.50) = $0.70 a day, or $21 a month.
• Sonnet 5.5: (3 × $2) + (0.8 × $10) = $14 a day, or $420 a month.
• Opus 5.5: (3 × $4) + (0.8 × $20) = $28 a day, or $840 a month.

If you're moving from Sonnet 4.6 to Sonnet 5.5, headline rates suggest roughly a one-third drop. With an estimated 30% increase in token count, the comparable daily cost comes out to about $18.20 - around 13% lower. That's why the only meaningful comparison is based on measured prompts and outputs.

Long prompts on Haiku 5.5

Haiku's 100,000-token threshold is worth testing directly. Assume each request has 150,000 input tokens and 2,000 output tokens, with 100 requests per day. Because the input exceeds 100,000 tokens, the higher Haiku 5.5 rate applies:

• Input: 15 million × $0.50 = $7.50 a day.
• Output: 0.2 million × $2.50 = $0.50 a day.
• Total: $8 a day, or $240 a month.

At the lower-rate tier, the same token volume would be $1.60 a day. For comparison, Sonnet 5.5 is $32 a day in this example. Haiku is still cheaper here - but the threshold makes its bill about five times higher than a lower-rate calculation would suggest.

The threshold is sharp. A 99,000-token prompt costs about $0.0099 in Haiku input. A 101,000-token prompt costs about $0.0505 - roughly five times as much for just over 2% more text. Cache reads and cache writes count toward the 100,000-token limit, so caching part of the prompt doesn't get you out of the higher-rate tier.

Prompt caching: cheaper repeated context

Prompt caching lets later requests reuse a prompt prefix so you don't pay the full input rate each time. Cache reads are charged as a fraction of the normal input price:

Writing to cache costs more than regular input tokens: a five-minute cache write is 1.25× the input rate, while a one-hour write is 2×. For example, Sonnet 5.5 cache writes cost $2.50 per MTok for five minutes and $4 per MTok for one hour.

Consider a 2,000-token system prompt sent 50,000 times daily to Sonnet 5.5. Without caching, those 100 million input tokens cost $200 a day. With cache reads at $0.10 per MTok, reads cost $10 a day, plus the initial cache write. If the prefix stays identical and requests arrive often enough to keep the cache entry active, you're saving close to $190 a day on that repeated context.

Caching isn't automatic savings. Before relying on it, check these conditions:

• Minimum prefix size: the models in the table require at least 512 tokens for a cached prefix. Some legacy models require more.
• Exact prefix matching: if you change a timestamp, user ID, or any other content inside the cached prefix, reuse may fail and you'll trigger a new write.
• Cache lifetime: the lifetime is measured from the start of the request that writes or reads the entry.
• Break-even: at a 0.1× read rate, Anthropic says a five-minute write pays for itself after one cache read and a one-hour write after two.

For implementation details, use Anthropic's prompt caching documentation. Keep stable instructions and shared context in the reusable prefix, and put request-specific content after it.

Batch API: lower prices for work that can wait

The Message Batches API charges 50% of the standard input and output rates. Anthropic says most batches finish within an hour, though results are available only when all requests complete or after 24 hours. If a request isn't processed, it expires without charge. A batch can include up to 100,000 requests or 256 MB.

Batch is a good fit for evaluations, document processing, backfills, and other "we can wait" jobs. It's a poor fit for interactive support or any in-product assistant that needs to respond right away. You can combine caching and batch discounts, but Anthropic recommends using a one-hour cache when shared context is used across a batch that may run longer than five minutes. Fast mode isn't available in Batch.

Choose by measured quality, not price tier

Anthropic positions Haiku 5.5 for high-volume tasks like classification, extraction, and routing. Sonnet 5.5 is presented as a balance of speed and capability. Opus 5.5 and Fable 5.1 are intended for more demanding reasoning and agentic work. Those descriptions can narrow your test plan, but they don't replace evaluation.

A practical evaluation process:

1. Pick 200-500 representative requests, including difficult cases and normal traffic.
2. Run them through two candidate models with the same tools and application instructions.
3. Score outputs against your criteria: correctness, output format, latency, escalation rate, and downstream rework.
4. Record actual input and output tokens (including thinking tokens), then estimate monthly cost at your expected volume.
5. Keep the cheaper option only if it meets your quality bar. Route harder requests to a stronger model when that split is reliable.

The right comparison is total operating cost, not just API spend. A low-cost model that produces unreliable structured output, misses important issues, or forces more human review can end up being the expensive option. On the other hand, using a top-tier model for every routine classification request is hard to justify if the smaller model passes your evaluation.