Claude Sonnet 5 (2026): Specs, Pricing, API, Benchmarks & Comparisons

Claude Sonnet 5 is Anthropic's latest Sonnet model, released on June 30, 2026. It is built for coding, tool use, and long-running agent tasks, with a 1 million-token context window and adaptive thinking enabled by default.

Anthropic's Claude Sonnet 5 launch announcement calls it the most agentic Sonnet model yet. The company says the biggest gains over Sonnet 4.6 are in coding and agentic work. At higher effort settings, Sonnet 5 can reach Opus 4.8-level performance on some of Anthropic's tests.

For developers, the move from Sonnet 4.6 also changes how thinking, sampling controls, and token counts work. Manual thinking budgets are gone, non-default sampling values can return errors, and the new tokenizer changes how much text fits into the same token budget.

What is Claude Sonnet 5?

Claude Sonnet 5 is the newest model in Anthropic's Sonnet line. It is aimed at demanding work where teams need strong reasoning and tool use without moving to an Opus-class model.

Anthropic says the largest gains over Sonnet 4.6 are in coding and agentic tasks, especially work that uses tools across several steps. The earlier Claude 4 overview shows how the previous Sonnet and Opus models were positioned.

Claude Sonnet 5 release date and availability

Anthropic released Claude Sonnet 5 on June 30, 2026.

At launch, it became the default model for Claude Free and Pro users. Max, Team, and Enterprise users also received access, and Anthropic released the model in Claude Code and the Claude API.

Developers can also access Sonnet 5 through Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.

Claude Sonnet 5 specs at a glance

Anthropic's Sonnet 5 technical documentation confirms the 1 million-token context window and 128,000-token maximum output. It also documents adaptive thinking, the new tokenizer, and API behavior changes.

What's new in Claude Sonnet 5?

If you already run Sonnet 4.6 in production, check these three API changes before switching.

Adaptive thinking is on by default

On Sonnet 4.6, a request without a thinking field could run without extended reasoning. Sonnet 5 uses adaptive thinking by default.

Claude decides when a task needs more reasoning and how much to use. You can still turn thinking off when the task doesn't need it.

Manual thinking budgets are removed

Fixed thinking budgets are no longer supported.

A setting such as thinking: {type: "enabled", budget_tokens: N} returns a 400 error on Sonnet 5. Anthropic recommends adaptive thinking with the effort setting instead.

If you're setting up the API from scratch, the Claude API overview covers authentication, requests, and the Messages API.

Non-default sampling settings are not accepted

Sonnet 5 does not accept non-default values for temperature, top_p, or top_k. Sending those values can return a 400 error.

If an older app depends on those controls, remove them before switching models. Use prompts and effort settings instead.

Claude Sonnet 5 context window and token limit

Claude Sonnet 5 has a 1 million-token context window. It supports up to 128,000 output tokens on the standard synchronous Messages API.

For asynchronous batch requests, Anthropic supports up to 300,000 output tokens with the output-300k-2026-03-24 beta header. That higher limit applies to the Message Batches API, not the synchronous Messages API.

The 1 million-token context window gives it room for large codebases, long documents, extended conversations, and agent sessions with a lot of history. Still, there is no reason to fill the window just because it is available. Extra context can add cost and irrelevant material.

Sonnet 5 uses a new tokenizer

Anthropic says the same input text produces approximately 30% more tokens on Sonnet 5 than on Sonnet 4.6. The exact change depends on the content.

This affects both cost and context use. The same document can take up more of the 1 million-token window. An output limit tuned for Sonnet 4.6 may also be too small for a similar Sonnet 5 response.

If you're moving a live workload, recount real prompts and review limits set around the older tokenizer.

When choosing the right AI model, compare context size with speed, output quality, and cost on the same workload.

Claude Sonnet 5 adaptive thinking and effort levels

Sonnet 5 supports five effort levels:

• low
• medium
• high
• xhigh
• max

high is the default on the Claude API and Claude Code.

Lower effort can use fewer tokens and respond faster on simple work. Higher effort gives the model more room for hard reasoning, coding, and long-running agent tasks.

xhigh is aimed at the hardest coding and agentic work. max gives Claude the most room to reason, but it can use more tokens and take longer. Use it when your own tests show that the extra spend improves the result.

Effort is not a fixed thinking-token budget. It tells Claude how deeply to work on the request rather than setting a fixed number of thinking tokens.

Claude Sonnet 5 API

The Claude Sonnet 5 API model ID is claude-sonnet-5. Despite the dateless format, Anthropic treats this ID as a pinned model snapshot rather than an evergreen pointer that silently moves to a newer model.

A basic request still uses Anthropic's Messages API:

1curl https://api.anthropic.com/v1/messages \
2
3-H "content-type: application/json" \
4
5-H "x-api-key: $ANTHROPIC_API_KEY" \
6
7-H "anthropic-version: 2023-06-01" \
8
9-d '{
10
11"model": "claude-sonnet-5",
12
13"max_tokens": 4096,
14
15"messages": [
16
17{
18
19"role": "user",
20
21"content": "Review this deployment plan and identify the main risks."
22
23}
24
25]
26
27}'

If you're moving from Sonnet 4.6, check token counts first. Remove manual thinking budgets and non-default sampling values before testing the new model.

Tool definitions and response shapes stay the same as Sonnet 4.6.

How to use Claude Sonnet 5

There are three main ways to use Sonnet 5:

• Claude: For direct chat, writing, analysis, and knowledge work.
• Claude Code: For software development and coding tasks.
• Claude API: For custom apps, agents, and workflows.

Anthropic also makes Sonnet 5 available through Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Developers using the Claude API call it with the model ID claude-sonnet-5.

Claude Sonnet 5 pricing

Claude Sonnet 5 API pricing is $2 per million input tokens and $10 per million output tokens. Anthropic originally announced that rate as introductory pricing, then made it permanent on August 10, 2026.

The current per-token rate is lower than Sonnet 4.6, which is priced at $3 per million input tokens and $15 per million output tokens. Real task cost can still differ because Sonnet 5's tokenizer may count more tokens for the same text, and higher effort can use more output tokens.

Anthropic prices Message Batches at 50% of standard API rates. At Sonnet 5's current $2/$10 rate, that works out to $1 per million input tokens and $5 per million output tokens for batch jobs.

A lower token price doesn't always mean a cheaper workflow. In production, compare the total cost of finishing the same task at the quality and speed you need.

Is Claude Sonnet 5 free?

Yes. Claude Sonnet 5 is available on Anthropic's Free plan, subject to usage limits. API access is separate and billed by token usage.

Claude Sonnet 5 benchmarks

Anthropic reports the largest gains over Sonnet 4.6 in coding and agentic tasks.

Anthropic's launch tests compare Sonnet 5 with Sonnet 4.6 and Opus 4.8 across agentic search and computer use. Sonnet 5 improves on Sonnet 4.6 in those tests.

At higher effort, it can reach Opus 4.8-level performance on some tasks.

Those results don't make Sonnet 5 the best model for every workload. Performance changes with the task, effort level, tools, and test setup.

Before a live rollout, run the same prompts and workflows on the models you're considering. Compare task completion, speed, retries, tool use, and total cost rather than one benchmark score.

Claude Sonnet 5 coding and agentic use cases

Sonnet 5 is most useful when the model has to keep working across several steps instead of answering one prompt and stopping.

Coding is a strong fit. In an agent with the right tools, Sonnet 5 can inspect code, make changes, run tests, debug failures, and continue through a multi-step task. Outside coding, it fits workflows that need planning, tool calls, verification, and several steps.

Common use cases include:

• Software development and debugging
• Multi-step AI agents
• Long-document and data analysis
• Computer-use workflows
• Large-codebase analysis

Reliability still depends on the workflow around the model. In agentic AI workflows, the wider setup controls data access, tools, actions, and what happens after each step.

Claude Sonnet 5 vs Sonnet 4.6

Anthropic calls Sonnet 5 an upgrade from Sonnet 4.6, with the largest gains in coding and agentic work.

It adds adaptive thinking by default and uses the new tokenizer. It also supports xhigh effort and improves performance on longer tool-driven tasks.

Sonnet 5 now has a lower API rate than Sonnet 4.6: $2 per million input tokens and $10 per million output tokens, compared with $3 and $15 for Sonnet 4.6. Real task costs can still differ because Sonnet 5 may use more tokens for the same text.

Before routing production traffic from Sonnet 4.6 to Sonnet 5, run the same prompts and workflows on both models.

Claude Sonnet 5 vs Opus 4.8 and Opus 5

Anthropic's Sonnet 5 launch benchmarks compare it directly with Opus 4.8. At higher effort settings, Sonnet 5 can reach Opus 4.8-level performance on some of Anthropic's tests, which makes Opus 4.8 a useful reference for teams deciding whether to move to Sonnet 5.

The price difference is clear. Sonnet 5 costs $2 per million input tokens and $10 per million output tokens, while Opus 4.8 costs $5 and $25. Both support a 1 million-token context window and up to 128,000 output tokens.

For a current Sonnet-versus-Opus decision, Opus 5 is the higher-tier model to test. It keeps the $5/$25 API pricing and is aimed at harder agentic coding and enterprise workflows.

If Sonnet 5 handles your workload reliably, it offers the lower token cost. If it struggles with your hardest tasks, test Opus 5 on the same prompts, tools, and workflows before deciding.

Claude Sonnet 5 vs GPT-5.6

Sonnet 5 and GPT-5.6 are both built for complex coding and agentic work, but their API economics differ. Sonnet 5 is priced at $2 per million input tokens and $10 per million output tokens. OpenAI lists GPT-5.6 Sol at $5 and $30.

Both offer roughly a 1 million-token context window and up to 128,000 output tokens.

For coding agents, test more than first-pass answers. Check whether the model completes the task, uses tools reliably, and recovers from failed tool calls, bad edits, or test failures. At scale, latency, retries, and total task cost matter too.

For background on OpenAI's model lineup, the GPT models overview traces the major GPT releases and their tradeoffs.

If xAI is also on your shortlist, the Grok 4 guide covers its pricing, benchmarks, and access options.

Claude Sonnet 5 vs Gemini

Gemini is a model family, so the comparison changes depending on which Gemini model you plan to use.

For long-context work, test how well each model keeps track of the material you provide. For coding agents, focus on tool use, staying on task, and recovery after mistakes. High-volume workloads put more weight on speed and cost.

Run the same prompts, tools, and workflows on the models you're considering. The ChatGPT vs Claude vs Gemini comparison gives more context on the three ecosystems.

Limitations to know before switching

Sonnet 5 improves on Sonnet 4.6, but a live switch still needs testing.

A prompt that fit inside an old token budget may use more tokens on Sonnet 5. Adaptive thinking can also increase token use on hard tasks.

Using xhigh or max for every request can waste time and budget.

If you're migrating an older integration, remove manual thinking budgets and non-default sampling values first. Sonnet 5 also doesn't support Anthropic's Priority Tier, which matters for teams that depend on reserved API capacity or priority service levels.

Sonnet 5 can refuse some prohibited or high-risk cybersecurity requests. Anthropic returns these refusals as successful API responses with stop_reason: "refusal", so security-related apps should handle that response explicitly.

What shouldn't Claude Sonnet 5 automate without human review?

A model that can plan and use tools still shouldn't get unrestricted control over high-impact actions.

Keep approval steps or human review for cases such as:

• Irreversible production changes
• Financial transfers, large refunds, or purchases above set limits
• Account-security, identity, or permission changes
• Medical, legal, or compliance decisions
• Actions based on missing, conflicting, or unverified data

For customer support, escalation is useful when the agent can't verify an answer or a customer disputes a policy or charge. It also matters when account access is involved or a request falls outside the configured rules.

How Claude Sonnet 5 fits into a multi-model AI stack

In a customer-facing AI agent, the model handles reasoning, but the surrounding system decides what data it can use, which actions it can take, and when a person should take over.

Chatbase lets teams build AI agents with supported models from leading AI providers. Its AI model comparison guide breaks them down by credit cost, technical capability, communication quality, speed, actions, and complex-task handling.

For a support team, the fastest model isn't always the right one. A high-volume FAQ flow may favor speed and cost. Troubleshooting with several tool calls may need more reasoning and a clear path to human handoff.

If you're evaluating Sonnet 5 for a customer-facing chatbot, you still need to connect the knowledge it can use, define what actions it can take, and decide when a conversation should move to a person. The build an AI chatbot walkthrough shows how those pieces fit together from setup to deployment.

Once you've mapped the knowledge sources, actions, and handoff rules, you can sign up for Chatbase to build the agent. Then connect it to your business data and workflows and test it against real support questions.

FAQs

Does Claude Sonnet 5 support images?

Yes. Claude Sonnet 5 supports text and image input and produces text output.

What is Claude Sonnet 5's knowledge cutoff?

Anthropic lists January 2026 as both the reliable knowledge cutoff and training data cutoff for Sonnet 5.

Can Claude Sonnet 5 use tools?

Sonnet 5 is designed for tool-heavy and agentic work. It can use browsers, terminals, code tools, and other tools provided by the application. What it can do depends on the tools and permissions you give it.

What is the Claude Sonnet 5 model ID?

The Claude API model ID is claude-sonnet-5.