Claude Sonnet 5: Pricing, Context Window, Benchmarks, API

What Is Claude Sonnet 5?

Last updated August 2026.

Claude Sonnet 5 is Anthropic's mid-tier model, released June 30, 2026. The model ID is claude-sonnet-5. It has a 1M-token context window, 128K max output on the Messages API, and costs $2 per million input tokens and $10 per million output tokens.

The pricing is the part worth pausing on. Sonnet 4.6 lists at $3 input and $15 output. Sonnet 5 replaced it at $2 and $10. Anthropic shipped those rates as introductory pricing, then made them permanent and dropped the $3/$15 standard tier it had originally announced. A generation upgrade that cuts the rate by a third is not the usual direction.

Anthropic's positioning is agentic: Sonnet 5 "is built to be the most agentic Sonnet model yet. It can make plans, use tools like browsers and terminals, and run autonomously." The docs describe it as the best combination of speed and intelligence, and rate its comparative latency as Fast against Opus 5's Moderate.

Claude Sonnet 5 Pricing

Sonnet 5 sits between Haiku 4.5 and Opus 5 on price, at $2 input and $10 output per million tokens. Against the current lineup:

Opus 5 costs 2.5x Sonnet 5 on both input and output. Haiku 4.5 costs half on input and half on output, but carries a 200K context window instead of 1M and caps output at 64K. The jump from Sonnet 5 to Fable 5 is 5x.

One number that does not appear on the price sheet: Sonnet 5 uses the tokenizer introduced with Claude Opus 4.7. Compared with models older than Opus 4.7, the same text produces roughly 30% more tokens. Moving a workload from Sonnet 4.5 to Sonnet 5 changes the token count as well as the rate, so re-measure with the token counting endpoint rather than assuming your old counts hold.

Context Window and Output Limit

Sonnet 5 has a 1M-token context window, which Anthropic annotates as roughly 555,000 words or 2.5M unicode characters. Max output is 128K tokens on the synchronous Messages API.

The Message Batches API raises that ceiling. Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 support up to 300K output tokens there using the output-300k-2026-03-24 beta header. If you are generating long documents or large diffs and latency is not the constraint, that path more than doubles the output budget.

A 1M window is a budget, not a plan. Filling it with retrieved files is how teams burn input tokens on context the model never needed. This is the case for narrowing what goes in: agentic code search that returns the right files beats bulk retrieval that returns many. See context rot for what happens to agent quality as the window fills.

Claude Sonnet 5 Benchmarks

Anthropic's launch post publishes most of its Sonnet 5 comparison as chart images, captioned "Scores for Sonnet 5 on a variety of evaluations compared to those of Sonnet 4.6 and Opus 4.8." The numbers stated in the body text are these:

SWE-bench Verified, SWE-bench Pro, Terminal-Bench, GPQA Diamond, and BrowseComp are all referenced in the launch material, but their Sonnet 5 figures appear only inside chart images and not as text. We are not reproducing numbers we cannot quote from a first-party source. The qualitative claim Anthropic does make in text is that agentic performance is a "substantial improvement over its predecessor, Sonnet 4.6," and that Sonnet 5 "reasons in tighter steps and gets our users to answers noticeably faster."

Sonnet 5 vs Opus 5 vs Haiku 4.5

Anthropic's own guidance is to start with Opus 5 for complex agentic coding and enterprise work, and to reach for Fable 5 when a workload needs the highest available capability. Sonnet 5 is the speed-and-intelligence tier.

Two differences decide most of these choices. Haiku 4.5 has a 200K window, so long-context agent loops rule it out regardless of price. Opus 5 carries a May 2026 knowledge cutoff against Sonnet 5's January 2026, which matters when the task depends on recent library or API behavior.

The routing question is the practical one. A coding agent session mixes trivial edits with hard multi-step reasoning, and pinning one model to the whole session either overpays on the easy calls or underperforms on the hard ones. See LLM routing and Sonnet vs Haiku for the cost math.

API Access

Pass claude-sonnet-5 as the model to the Messages API. The model ID is a pinned snapshot, not an evergreen pointer, even though it carries no date suffix.

Three API details differ from older Sonnet models. Adaptive thinking (thinking: {type: "adaptive"}) replaces the extended thinking parameter, and the older budget_tokens form is not supported. The effort parameter defaults to high on the Claude API and Claude Code, so a workload that never set it explicitly is running at high. Non-default sampling parameters are rejected.

For per-request pricing on the Anthropic API including batch discounts and prompt caching rates, see our Anthropic API pricing guide.

Pros and Cons

Open-Weight Alternatives on Morph

Morph does not serve Claude models. It serves open-weight models on custom kernels through an OpenAI-compatible API, and for coding workloads several of them undercut Sonnet 5 substantially. The weights are public, so a model you prototype on the API is a model you can self-host later.

GLM-5.2 is the closest substitute for coding work: 753B MoE under an MIT license with a 1M-token context, at roughly half Sonnet 5's input rate and 41% of its output rate. MiniMax M3 and DeepSeek V4 Flash sit an order of magnitude below that for high-volume pipelines where per-call quality matters less than throughput.

See GLM-5.2, Morph open source models, and pricing.

Frequently Asked Questions

What is Claude Sonnet 5?

Anthropic's mid-tier model, released June 30, 2026. Model ID claude-sonnet-5, 1M-token context, 128K max output, $2/M input and $10/M output. Anthropic calls it the most agentic Sonnet model yet.

How much does Claude Sonnet 5 cost?

$2 per million input tokens and $10 per million output tokens. That is below Sonnet 4.6's $3/$15. Anthropic launched at $2/$10 as introductory pricing and then made those rates permanent, dropping the $3/$15 standard tier it had planned.

What is Sonnet 5's context window?

1M tokens, roughly 555,000 words. Max output is 128K on the Messages API and up to 300K on the Message Batches API with the output-300k-2026-03-24 beta header.

Is Sonnet 5 better than Opus 5?

They target different jobs. Anthropic recommends Opus 5 for complex agentic coding and enterprise work; Sonnet 5 is the speed-and-intelligence tier at 2.5x lower cost. Opus 5 also carries a May 2026 knowledge cutoff against Sonnet 5's January 2026.

What did Anthropic publish for Sonnet 5 benchmarks?

In text: 34.6% on Humanity's Last Exam without tools, 46.8% with tools, and 78.5% on OSWorld-Verified. The comparison against Sonnet 4.6 and Opus 4.8, including BrowseComp and agentic coding evaluations, was published as chart images rather than text, so those numbers are not quotable from the announcement.

Is Sonnet 5 cheaper than Sonnet 4.6?

Yes on the rate: $2/$10 against $3/$15. But Sonnet 5 uses the tokenizer introduced with Opus 4.7, so the same text produces roughly 30% more tokens than on pre-4.7 models. Measure with count_tokens before projecting the saving.

Can I self-host Claude Sonnet 5?

No. Claude models are closed-weight and API-only. If self-hosting or fine-tuning on your own data is a requirement, an open-weight model such as GLM-5.2 (MIT license, 1M context) is the substitute.

Does Sonnet 5 support extended thinking?

Not the older form. thinking.type: "enabled" with budget_tokens is unsupported on Sonnet 5. Use adaptive thinking (thinking: {type: "adaptive"}) and control depth with the effort parameter, which defaults to high.