Claude Sonnet 5 Pricing: API Rates and Plan Costs

Claude Sonnet 5 is priced at $2 per million input tokens and $10 per million output tokens via the API through August 31, 2026. Standard pricing kicks in the following day, rising to $3 for input and $15 for output.

We route AI model workloads for SMB clients in every implementation we build, and the detail most likely to catch a budget off guard isn't the headline price, it's the tokenizer change behind this release.

Anthropic launched Sonnet 5 on June 30, 2026, positioning it as a mid-tier, high-volume model between Haiku 4.5 and the Opus tier. Per-token pricing only tells part of the story, because the new tokenizer changes how many tokens your actual documents generate.

Claude Sonnet 5 API Pricing

Sonnet 5's introductory rate runs through August 31, 2026, at $2 per million input tokens and $10 per million output tokens on the Anthropic API. That changes the next day. Standard pricing takes over on September 1, 2026, at $3 input and $15 output per million tokens. Cached input stays at $0.20 per million tokens under both rate cards.

That cache rate matters for any workflow with a repeated prompt prefix, such as a support bot that reuses the same system instructions on every call. The savings compound fast. Caching those repeated tokens cuts their cost by 90 percent or more compared to a fresh input token, on either rate card.

Confirm the live figures on Anthropic's pricing page before you budget a large workload. Rates move without much notice. Sonnet 5 already has one scheduled change built in, and another could follow it.

• Introductory input: $2 per million tokens, through Aug 31, 2026
• Introductory output: $10 per million tokens, through Aug 31, 2026
• Standard input: $3 per million tokens, from Sep 1, 2026
• Standard output: $15 per million tokens, from Sep 1, 2026
• Cached input: $0.20 per million tokens, both rate cards

The Tokenizer Change That Moves Your Real Bill

Sonnet 5 runs on a newer tokenizer than Sonnet 4.6. Anthropic says it produces roughly 30 percent more tokens for the same text. A contract that billed at 10,000 tokens on the prior model now bills closer to 13,000 tokens on Sonnet 5, even though the per-token price dropped.

The lower headline rate can hide this shift. Test it yourself before you trust the sticker price. Run a real sample of your own documents through Sonnet 5's tokenizer, then compare its rate card against a competitor's, instead of estimating from a word or character count.

We now tokenize a real sample document when we scope Claude spend for a new client build. Word-count estimates fall short. The gap between the two methods has grown wide enough to blow a monthly budget on a long-document workflow.

• Sonnet 5's tokenizer produces about 30% more tokens per document than Sonnet 4.6
• A lower per-token price can still raise your total bill once the token count rises
• Test with a real document sample, not a character-count estimate

Claude Consumer and Team Plan Pricing

Sonnet 5 is the default model on Claude's Free plan. Free costs $0 a month. It includes a 200,000-token context window and roughly half of the Pro plan's usage allowance.

Pro costs more but does more. It runs $17 a month billed annually, or $20 a month billed monthly, and adds the Opus tier plus Claude Code.

Max is the heavy-usage tier. It starts at $100 a month for 5 times the Pro usage allowance, with a 20-times tier available at a higher price, and it sets Opus as the default model instead of Sonnet 5.

Team and Enterprise scale by seat. Team runs $20 to $25 a month per standard seat, or $100 to $125 for a premium seat. Enterprise costs $20 a seat plus usage billed at API rates.

• Free: $0/month, Sonnet 5 only, 200k context, ~50% of Pro's usage cap
• Pro: $17/month annual or $20/month monthly, adds Opus and Claude Code
• Max 5x / 20x: from $100/month, Opus is the default model
• Team: $20-25/month standard seat, $100-125 premium seat
• Enterprise: $20/seat plus usage billed at API rates

A Realistic Monthly Cost Example

Take a support-automation workload of 10,000 conversations a month, averaging 800 input tokens and 400 output tokens per conversation once you fold in the tokenizer's 30 percent bump.

Run the standard-rate math and that is 10.4 million input tokens for $31.20 and 5.2 million output tokens for $78.00. The monthly API bill lands near $109. Swap in the introductory rate and the same workload runs closer to $73 through August 31, 2026.

One lever moves that number the most. Cache the system prompt on every call instead of resending it fresh each time, since the same instructions repeat on every conversation.

• 10,000 conversations/month, 800 input + 400 output tokens each (post-tokenizer)
• Standard pricing: ~$109/month total (input $31.20 + output $78.00)
• Introductory pricing: ~$73/month total, through Aug 31, 2026
• Caching a repeated system prompt cuts the input side further

Who Sonnet 5's Pricing Fits, and When to Route to Opus

Sonnet 5's pricing fits high-volume, everyday work: drafting, summarizing, customer replies, and agentic tasks that plan and use tools without needing the hardest reasoning ceiling available.

A workload split we see hold up across client builds routes about 80 percent of tasks to Sonnet 5. The hardest 20 percent goes to Opus instead, the tasks where a wrong answer costs more than the token savings. That split keeps quality high on the work that matters and controls spend on the rest.

One workload is the exception. Skip Sonnet 5's pricing for a job that leans mostly on the hardest coding or research tasks. Opus's higher per-token rate buys a meaningfully stronger reasoning ceiling there, and routing that work to Sonnet 5 to save money usually costs more in rework than it saves in tokens.

Our answer would change if Anthropic ever prices Opus within a small premium of Sonnet 5. At that point the case for routing anything difficult to Sonnet 5 weakens. The quality gap would no longer be worth defending on cost alone.

• Good fit: high-volume drafting, support replies, planning and tool-use tasks
• Route to Opus instead: the hardest 15-20% of tasks, where errors are expensive
• Revisit this split if Opus and Sonnet 5 pricing ever converges