Claude Sonnet 5 Pricing and API Guide

Claude Sonnet 5 Pricing

Claude Sonnet 5 launched on June 30, 2026 with an introductory pricing window. At $2.00 per million input tokens and $10.00 per million output tokens, the intro rate is competitive for high-volume workloads. The introductory window closes on August 31, 2026 -- after which input rises to $3.00 and output to $15.00 per MTok.

Subscription and API billing are separate systems. Free plan users get Sonnet 5 as the default model as of June 30, 2026. Pro subscribers also get Sonnet 5 by default, with Claude Max providing extended context limits. API access is always billed per token -- no subscription plan is required to start making API calls. If you are weighing Sonnet 5 against OpenAI on cost and capability, see Claude Sonnet 5 vs GPT-5.6.

How tokens are counted

Input tokens include everything sent in the request: system prompt, conversation history, tool definitions, and the user message. Output tokens are what the model generates. In multi-turn conversations or agentic loops, the full conversation history is resent each turn, so input costs accumulate quickly without prompt caching. Getting your caching strategy in place before you scale matters more at the standard rate than at the intro rate.

Choosing the Right Model ID

Anthropic provides three ways to reference Sonnet 5 in your API calls. They differ in stability and appropriate use case. Picking the wrong one for production will cause silent breaking changes when Anthropic releases future models.

The three model IDs

• claude-sonnet-5 -- Canonical short name for the model family. Points to the current Sonnet 5 generation. Appropriate for initial integration and development work.
• claude-sonnet-5-20260630 -- Versioned ID pinned to the June 30, 2026 release. The only ID guaranteed not to change when Anthropic updates the Sonnet family. Use this in all production deployments.
• claude-sonnet-latest -- Alias that always resolves to the newest Sonnet model Anthropic has released. As of June 30, 2026, this points to Sonnet 5. When a newer Sonnet ships, this alias switches automatically. Never use this in production systems.

The API returns the resolved model ID in every response object. If you inherit a codebase that uses claude-sonnet-latest, check your response logs to confirm which model is actually being called. Swap to the versioned ID before moving to production.

Context window

All three IDs access the same model with a 1,000,000-token (1M) input context and a 128,000-token (128K) maximum output. The full context window is available immediately with no configuration or special API tier. For most workloads the 1M context is more than sufficient, but enabling prompt caching on repeated large contexts becomes important at scale because large re-sent histories accumulate cost fast. For a deeper look at what Sonnet 5 can actually do with that context, see What Is Claude Sonnet 5?

The Tokenizer Change: What It Costs You

Sonnet 5 ships with an updated tokenizer that produces more tokens from the same input text compared to prior Claude Sonnet and Haiku versions. The multiplier ranges from 1.0x to 1.35x depending on content type. This is not a model quality regression -- the updated tokenizer improves handling of multilingual text and dense technical content. But it has direct cost implications for high-volume workloads.

Token counts by content type

• Structured data and configuration (JSON, CSV, XML, code): tends toward the 1.0x end. Minimal change expected for code-heavy prompts.
• Natural language prose (instructions, conversation, documents): tends toward the 1.35x end. Expect noticeably more tokens per call on prose-heavy workloads.
• Mixed content (instruction blocks wrapping technical content): lands between 1.1x and 1.25x depending on the ratio.

Cost impact at standard rates

A prompt previously counted as 1,000 input tokens that the new tokenizer now counts as 1,350 tokens costs 35% more at the same rate. At standard rates ($3.00/MTok), that same call goes from $0.003 to $0.00405. At 1 million calls per month, that difference is $405 vs $300 on input alone -- before output tokens.

The intro pricing window through August 31, 2026 was partly designed to cushion this transition. At $2.00/MTok input, the same 35% token increase is a smaller dollar impact. The time to benchmark your actual workloads against the new tokenizer is now, while costs are lower and before the standard rate kicks in.

Prompt Caching: When It Pays Off

Prompt caching stores large content blocks on Anthropic's servers so subsequent calls retrieve them at a fraction of the standard input cost. Writing to the cache is more expensive than a normal input token, so the economics only work when the same content is reused many times within the cache TTL period.

Cache pricing

• Cache write: $3.75/MTok -- paid once when content is first written to cache.
• Cache read: $0.30/MTok -- paid each time cached content is retrieved instead of re-processed.

Cache write at $3.75/MTok is 1.875 times the intro input rate of $2.00/MTok. That cost front-loading is the core constraint: you need enough cache reads to recover the write cost before caching becomes net positive.

Break-even math

During the intro period: write cost = $3.75, savings per cache read vs regular input = $2.00 - $0.30 = $1.70. Break-even at approximately 13 reads. At standard rates (from September 1, 2026): write cost = $3.75, savings per read = $3.00 - $0.30 = $2.70. Break-even drops to approximately 5 reads. Caching becomes more cost-effective at standard rates than at the intro rate.

How to enable caching

Set cache_control breakpoints in your Messages API request to designate content blocks for caching. The cache applies to content up to each breakpoint. Multiple breakpoints are supported within a single request. See the Anthropic prompt caching documentation for minimum cacheable token lengths, TTL durations, and implementation examples.

Batch API: 50% Off for Non-Interactive Work

The Batch API processes requests asynchronously. In return for accepting processing delays -- typically minutes to a few hours -- you pay 50% of the standard per-token rate. For any workload where latency does not affect user experience, Batch API is the most direct way to cut API costs in half with no change to model quality or output.

Effective Batch API rates

Workloads that fit Batch API

• Nightly document classification or enrichment pipelines
• Bulk sentiment analysis across a dataset
• Overnight prompt evaluation runs for model testing
• Pre-generating structured summaries from large document sets
• Background content moderation at scale
• Generating metadata for content catalogs

Workloads that do not fit Batch API

• Real-time customer chat -- requires immediate responses, incompatible with batch delays
• Interactive coding assistants
• Live search augmentation where latency affects perceived quality
• Streaming responses displayed incrementally to users

Submitting a Batch API request returns a batch ID you poll for completion. Results are returned as individual response objects, each matched to its original request via a custom ID you supply. See the Anthropic Message Batches documentation for batch size limits, request format, and polling patterns.

Dev Team Mode and Self-Verification

Sonnet 5 introduces native multi-agent orchestration: the model can autonomously decompose a complex software task into parallel sub-agents, run them concurrently, and synthesize results into a single coherent output. This capability is built into the model itself -- there is no separate API product, endpoint, or billing tier.

What Dev Team mode does

• Task decomposition: Sonnet 5 analyzes a complex request and splits it into parallel sub-tasks -- planner, implementer, tester, reviewer -- without explicit prompting.
• Parallel sub-agent execution: Sub-agents run concurrently, reducing wall-clock time on multi-step tasks compared to sequential single-agent approaches.
• Synthesis: Results from all sub-agents combine into a coherent final output presented as a single response.
• Self-verification: For software repair tasks, Sonnet 5 can write a reproducing test, implement the fix, and re-run the test in a single autonomous pass with no human checkpoints between steps.

API access -- no special tier needed

Dev Team mode uses the standard Messages API with extended tool-use and multi-turn patterns. Any API key that can call Sonnet 5 with tool use already has access to the full multi-agent orchestration capabilities. There is no application process, enterprise agreement, or separate endpoint required.

The typical implementation passes tools including shell execution, file-write access, and a test-runner command that returns exit codes. The model reads the exit codes and iterates if the fix did not pass the test -- completing all three steps (write test, implement fix, re-run) in a single agentic pass.