Anthropic API Pricing in 2026: Opus, Sonnet, Haiku Token Costs

Claude Opus 4.6 on the API: when frontier capability is worth the premium

Claude Opus 4.6 at about 15 dollars per 1M input tokens and 75 dollars per 1M output tokens is the most capable model in the Anthropic lineup as of May 2026. It is built for the hardest reasoning, the longest agent traces, and the tasks where Sonnet misses. The instinct for many developers is to use Opus as the default because it is the flagship. In practice, that approach burns money. The right pattern is to default to Sonnet for routine work and reserve Opus for the calls that genuinely need frontier capability. In our own stack only about 10 to 15 percent of calls touch Opus, and most of those are escalations from a Sonnet call that failed a quality check.

Claude Sonnet 4.6 at 3 dollars per 1M input: the production workhorse

Claude Sonnet 4.6 at about 3 dollars per 1M input tokens and 15 dollars per 1M output tokens is the model most chatbots, copilots, coding assistants, and content tools should run on in 2026. It is roughly 5x cheaper than Opus on input and 5x cheaper on output. Quality is close enough on routine work that most users cannot tell the difference, and on coding tasks specifically Sonnet has been the developer favorite across multiple recent generations. Pair it with prompt caching for a stable system prompt and tool definitions, and you are paying a fraction of the headline rate on most calls.

Claude Haiku 4.5 at 1 dollar per 1M input: when you only need a quick decision

Claude Haiku 4.5 at about 1 dollar per 1M input tokens and 5 dollars per 1M output tokens is built for high-volume narrow tasks. Classification, intent routing, tag extraction, content moderation, simple summaries, and the inner loops of multi-model pipelines. At this price you can embed a Claude call in places where the unit economics would not have worked at flagship pricing. A million classification calls on Haiku cost about 100 dollars total assuming short prompts. The same workload on Opus is 15 times more expensive on input alone.

Prompt caching: up to 90 percent off cached input tokens

Prompt caching is the single biggest cost lever on the Anthropic API. You mark a stable chunk of your prompt (system prompt, tool definitions, retrieved documents, few-shot examples) and Anthropic stores it on the server side. On subsequent calls within the cache window, that cached portion bills at up to 90 percent off the normal input rate. That is a far deeper discount than the typical 50 percent caching discount you see on other providers. Anthropic offers two cache TTL options: 5 minutes by default, and an extended 1-hour option for longer-lived caches. The price of writing to the cache is slightly higher than a normal input token, so the discount only pays back once the cached content is reused at least a couple of times.

Prompt caching example math, worked out

Take a coding agent that ships a 20,000 token system prompt and tool definitions block on every call, plus 500 tokens of user input, running on Sonnet 4.6. Input cost without caching: 20,500 tokens times 3 dollars per 1M = 0.0615 dollars per call. With caching applied to the 20,000 token stable block at a 90 percent discount, the cached portion bills at 0.30 dollars per 1M while the new 500 tokens still bill at the full 3 dollars per 1M. New per-call input cost: roughly 0.006 dollars cached portion plus 0.0015 dollars new tokens = about 0.0075 dollars. That is an 88 percent reduction in input cost per call, achieved by structuring the prompt with stable content first and variable content last. At a million calls per month, that is roughly 54,000 dollars in input cost saved per month on Sonnet alone.

Batches API: 50 percent off both input and output with a 24-hour SLA

The Batches API on Anthropic gives a flat 50 percent discount on both input and output tokens compared to realtime endpoints. The trade is asynchronous processing within a 24-hour SLA. You submit a batch of message requests, the system processes them when capacity allows, and you collect results once the batch completes. Use cases that fit perfectly include nightly classification of yesterdays content, bulk summarization of a document backlog, weekly report generation, monthly evaluation runs, and large-scale embedding-free retrieval reranking jobs. The Batches API is a cost lever most teams underuse. Walking through your workload once and tagging anything that does not require a realtime response is usually a fast path to cutting the bill by 30 to 50 percent on the eligible portion of traffic.

Batches API example math, worked out

Take a nightly classification pipeline that runs 5 million calls on Haiku 4.5 with 800 input tokens and 100 output tokens per call. Realtime cost: 5M calls times 800 tokens = 4B input tokens at 1 dollar per 1M = 4,000 dollars, plus 5M calls times 100 tokens = 500M output tokens at 5 dollars per 1M = 2,500 dollars. Total realtime: 6,500 dollars per month. Same workload on the Batches API at a 50 percent discount: 2,000 dollars input plus 1,250 dollars output = 3,250 dollars per month. That is a 3,250 dollar saving achieved by accepting a 24-hour processing window. Stack prompt caching on the stable system prompt portion and the effective saving climbs even higher.

Comparison to OpenAI API pricing

Anthropic and OpenAI cover roughly the same product surface (a frontier model, a workhorse, and a budget option), but the headline rates are not equivalent across tiers. Here is how the two lineups line up as of May 2026.

At the frontier tier, Claude Opus 4.6 and GPT-5 Pro land in a similar premium price band. At the workhorse tier, GPT-5 mini is meaningfully cheaper per token than Claude Sonnet 4.6, while Sonnet is widely considered the stronger coding model. At the budget tier, GPT-5 nano is dramatically cheaper than Claude Haiku 4.5 on input. The Anthropic prompt caching discount is deeper at up to 90 percent than OpenAIs 50 percent, which can flip the effective cost ranking on workloads with very large stable prompt prefixes. Verified May 2026.

Bedrock and Vertex AI delivery

Claude models are also available through Amazon Bedrock and Google Vertex AI in addition to the direct Anthropic API. The headline per-token rates on those platforms are typically the same or very close to the Anthropic rates listed on this page, but each cloud provider wraps the model with its own billing, region availability, rate limits, and committed-spend discount programs. Bedrock pricing flows through AWS billing under the Bedrock service line. Vertex AI pricing flows through Google Cloud billing under the Vertex AI service line. Both inherit cloud-side IAM, data residency, and compliance posture from the cross-region inference setup of their host platform. The right pick is usually the cloud where the rest of your infrastructure already runs. Verified May 2026.

Rate-limit tiers on the Anthropic API

Anthropic uses a usage tier system that scales rate limits with cumulative account spend and account age. New accounts start at Tier 1 after their first top-up and progress automatically as the account accumulates spend.

Hitting your tier cap returns a 429 response with a rate-limit error. Upgrades happen automatically once the spend threshold is met. Beyond Tier 4 you negotiate custom limits directly with the Anthropic sales team, typically as part of a committed-spend agreement. Verified May 2026.

Vision, tool use, citations, and extended thinking on the API

Vision input is billed per image based on resolution. Each image is converted into a token count tied to its dimensions and billed at the model standard input rate. Tool use is counted as output tokens for the structured tool-call JSON the model emits, while the tool result you send back on the next turn counts as input on that call. Citations released in the 1.0 generation add no per-call surcharge but they do add output tokens to the response. Extended thinking does not change the per-token rate but it does add output tokens, billed at the model standard output rate. For Opus that means thinking is billed at 75 dollars per 1M output tokens, which can dominate the bill on hard reasoning calls.

First-person: what we actually pay on the Anthropic API

Across our own production stack the routing mix is roughly 75 percent Claude Sonnet 4.6 for chat, coding, and most agent work, 15 percent Claude Haiku 4.5 for classification and routing, and 10 percent Claude Opus 4.6 reserved for the hardest reasoning calls and escalations from Sonnet. Prompt caching is applied to a 20,000 token stable system and tools prefix on every call, which pulls the effective input cost on Sonnet down to a fraction of the headline 3 dollars per 1M rate. The Batches API handles our nightly classification jobs and weekly report generation. The combined effect is that our average effective input cost on Sonnet sits well below 1 dollar per 1M tokens, against a headline rate of 3 dollars. That is the kind of number you should be targeting once you have the discount stacking dialed in. Verified May 2026.

Verdict: which Anthropic API model should you pick

My verdict after building on the Anthropic API since the early Claude generations: default new builds to Claude Sonnet 4.6 and only escalate to Claude Opus 4.6 on the calls that fail a quality check on Sonnet. Route classification and intent calls to Claude Haiku 4.5. Always turn on prompt caching by structuring your prompt with stable content first and variable content last, and route everything that does not need a realtime response through the Batches API for the 50 percent discount. Start on the 5 dollar free credit, hit Tier 1 after your first top-up, and let the tier ladder lift you naturally as production traffic accumulates. The single biggest mistake I see new developers make is calling Opus by default for everything. The single biggest cost win is structuring your prompts so the 90 percent caching discount actually applies. Verified May 2026.