Anthropic

Claude Sonnet 5

Provider

Anthropic

Status

available

Context

1,000,000 tok

SWE-bench

85.2%

Price

$3 / $15 /MTok

Knowledge

2026-03

Claude Sonnet 5 is Anthropic’s mid-tier model and, for most people, the one they actually use day to day. Released on 30 June 2026, it’s the default model on the Free and Pro tiers of claude.ai and the workhorse that sits beneath the Opus 5 flagship — Anthropic bills it as its most agentic Sonnet yet, pulling Opus-like coding, tool use and browser work down into a cheaper Sonnet-class model.

The reason it matters is price-to-performance. On Anthropic’s own benchmarks, Sonnet 5 scores 85.2% on SWE-bench Verified and 63.2% on the harder SWE-bench Pro — within about six points of Opus 4.8 on repository-scale coding, and comfortably ahead of Sonnet 4.6. The catch, which we’ll come back to, is a new tokenizer that changes how you should think about its cost.

Quick specs

TRY CLAUDE SONNET 5 →

What’s new in Claude Sonnet 5

Sonnet 5 is the successor to Sonnet 4.6, and the pitch is agentic capability at the Sonnet price point.

The most agentic Sonnet yet

Anthropic positions Sonnet 5 as bringing Opus-like coding, tool use, browser work and professional-task performance into a cheaper Sonnet-class model (Anthropic). On SWE-bench Pro — the memorisation-resistant, multi-language coding benchmark this site weights most — it scores 63.2%, up from Sonnet 4.6’s 58.1% and only about six points behind Opus 4.8 (69.2%). On the standard SWE-bench Verified it reaches 85.2%, and on SWE-bench Multilingual 78.3%. Anthropic also says it edges Opus 4.8 on at least one knowledge-work benchmark — the same “the mid-tier can beat last generation’s flagship” pattern Sonnet 4.6 established.

A 1M-token context window

Sonnet 5 ships with a 1-million-token context window and 128K maximum output, matching the Opus line, so whole-codebase and long-document work no longer needs to escalate to Opus purely for room.

The same effort controls as Opus 5

Like Opus 5, Sonnet 5 is one model shaped by adaptive thinking and an effort dial (low, medium, high, max), rather than separate model IDs. Left at max effort it reasons harder and scores better, at the cost of more tokens — which matters more than usual this generation because of the tokenizer change below.

A new tokenizer — the hidden cost

The most important practical change isn’t a capability, it’s the meter. Sonnet 5 uses a new tokenizer that encodes the same text into roughly 1.0-1.35x more tokens than Sonnet 4.6 (Data Science Dojo). Anthropic set introductory pricing low enough that, for now, the switch is roughly cost-neutral — but once standard pricing resumes (see Pricing), the same workload can cost meaningfully more than it did on 4.6. Re-run your own cost evals before migrating a high-volume pipeline.

Where Sonnet 5 sits in the Claude lineup

Within Anthropic’s ladder, Sonnet 5 is the balanced mid-tier: most of the capability, a fraction of the flagship price.

Sonnet 5 is the model most Claude users actually touch, because it’s the default on the two most popular tiers. You escalate to Opus 5 when a task genuinely needs the extra ceiling on reasoning or repository-scale coding.

Benchmark performance

All figures below are Anthropic-reported from the launch at adaptive thinking / max effort. Independent composite results (Artificial Analysis, Vals AI) had not stabilised at the time of writing, so treat the vendor numbers as a ceiling.

Coding

The story is how close Sonnet 5 runs to last generation’s flagship: on SWE-bench Pro it’s about six points behind Opus 4.8 and five ahead of Sonnet 4.6. For where this lands against the wider field — including the Opus 5 flagship (96.0% Verified) and GPT-5.6 Sol — see our best AI for coding rankings.

The Opus gap

Sonnet 5 narrows but doesn’t close the gap to the flagship. On the hardest scientific and abstract reasoning, on large multi-file refactors and on precise long-context retrieval, Opus 5 still pulls clearly ahead — which is exactly why it’s the paid-tier default and Sonnet 5 the everyday one. If your workload is dominated by novel-pattern reasoning or repository-scale software engineering, that’s the signal to escalate.

Pricing breakdown

Sonnet 5’s pricing has two phases, and the gap between them is the thing to plan around.

Two things matter here. First, the headline rate lands back at the familiar $3/$15 Sonnet price from 1 September. Second — and this is the part most coverage leads with — the new tokenizer counts ~1.0-1.35x more tokens for the same text than Sonnet 4.6 did (Data Science Dojo). During the introductory window the lower per-token rate offsets that, so costs are roughly flat; after 1 September, an unchanged workload can run ~20-35% more expensive than it did on 4.6. Cheaper-looking per token is not the same as cheaper per job — model the real per-task cost before you commit.

Cost comparison with contemporaries

How to access Claude Sonnet 5

Via API

Sonnet 5 is generally available with no waitlist as claude-sonnet-5 on the Claude API, and on Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and GitHub Copilot. Effective context varies by platform, so check the limit where you deploy.

from anthropic import Anthropic
client = Anthropic()

message = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(message.content)

Via the Claude apps

Sonnet 5 is the model most people on claude.ai are using, because it’s the default on both the Free and Pro tiers (Anthropic pricing).

It also powers Claude Code and Claude Cowork for lighter, high-volume work, with Opus 5 available when a task needs more.

How Claude Sonnet 5 compares

vs Claude Sonnet 4.6

Sonnet 5 leads its predecessor on every published coding benchmark — SWE-bench Pro 63.2% vs 58.1%, Verified 85.2% vs 79.6% — and adds the full 1M context and stronger agentic behaviour. The one real trade-off is the tokenizer: after the introductory period, the same job can cost more than it did on Sonnet 4.6, which remains available via the API for teams that want to hold their current cost profile.

vs Claude Opus 5

This is the escalation decision you’ll make constantly. Opus 5 is the plan-included flagship — it tops the independent intelligence index, leads on the hardest reasoning, and scores 96.0% on SWE-bench Verified — but it costs more per token and per task. Choose Sonnet 5 as your default for high-volume coding, tool use, computer use and customer-facing agents. Choose Opus 5 for deep reasoning, large multi-file refactors and anything where getting it exactly right beats getting it cheaply.

vs GPT-5.6 Terra

Against OpenAI’s balanced tier, GPT-5.6 Terra ($2.50/$15) is the closest competitor on price and positioning. Sonnet 5 leads on the coding benchmarks this site weights, while Terra brings OpenAI’s broader multimodal ecosystem. For the full field, see best AI models.

Known limitations

The tokenizer changes the cost maths. Sonnet 5 counts ~1.0-1.35x more tokens than Sonnet 4.6 for the same text, so after standard pricing resumes on 1 September 2026, an unchanged workload can cost ~20-35% more. This is the single most important thing to model before migrating.

Trails Opus 5 on the hardest reasoning. On novel-pattern reasoning, the hardest science, and repository-scale software engineering, Opus 5 is clearly ahead — Sonnet 5 narrows the gap but doesn’t close it.

Vendor-reported headline figures. Every launch benchmark is Anthropic-run; independent composite scores (Artificial Analysis, Vals AI) had not stabilised at the time of writing, so treat the numbers as a ceiling.

No fast mode. The research-preview fast mode is an Opus-only feature; Sonnet 5 runs at standard speed.

Version history

Sonnet 5 arrived on 30 June 2026, near the end of a two-month stretch in which Anthropic also shipped the Mythos-class Fable 5 and Mythos 5 (9 June) and, three weeks later, the Opus 5 flagship (24 July).

FAQ

Is Claude Sonnet 5 better than Sonnet 4.6?

Yes on capability — it leads Sonnet 4.6 on every published coding benchmark (SWE-bench Pro 63.2% vs 58.1%, Verified 85.2% vs 79.6%) and is more agentic. The caveat is cost: a new tokenizer counts more tokens per request, so once introductory pricing ends on 31 August 2026 the same workload can cost ~20-35% more than on 4.6. Re-run your cost evals before migrating.

How much does Claude Sonnet 5 cost?

Introductory pricing is $2 input / $10 output per million tokens through 31 August 2026, reverting to the standard $3 / $15 from 1 September — the same headline rate as Sonnet 4.6, with no long-context surcharge on the 1M window. Because of the new tokenizer, model your real per-task cost rather than relying on the per-token rate.

Can I use Claude Sonnet 5 for free?

Yes. Sonnet 5 is the default model on the free tier of claude.ai, subject to usage limits, and also the default on Pro ($20/mo), which adds access to the Opus 5 flagship.

How does Sonnet 5 compare to Opus 5?

Opus 5 is the stronger, pricier flagship — it tops the independent intelligence index and leads on the hardest reasoning and repository-scale coding (96.0% SWE-bench Verified). Sonnet 5 is the better default for high-volume coding, tool use and agents. Escalate to Opus 5 when a task genuinely needs the extra ceiling.

What is the context window and knowledge cutoff?

A 1,000,000-token context window with up to 128,000 output tokens, and a knowledge cutoff of March 2026.

Last verified 27 July 2026. Benchmark figures are Anthropic-reported from the Claude Sonnet 5 launch (30 June 2026); independent composite scores were still landing at the time of writing. Pricing is introductory ($2/$10) through 31 August 2026, reverting to $3/$15 on 1 September; the new tokenizer affects real per-task cost. Details subject to change — confirm against Anthropic’s official pages.