Claude Sonnet 5

Claude Sonnet 5 shipped 30 June 2026 at $2/$10 per million tokens with a 1M-token context. The tokenizer change, the cancelled price rise, and how to migrate from Sonnet 4.6.

Claude Sonnet 5 is Anthropic's mid-tier model, released on 30 June 2026. It costs $2 per million input tokens and $10 per million output tokens, has a 1,000,000-token context window as standard, produces up to 128,000 output tokens per request, and answers to the API ID claude-sonnet-5. For most workloads it is the model to default to.

The headline story is price. Sonnet 4.6 cost $3/$15; Sonnet 5 costs a third less on paper. The story underneath it is the tokenizer: Sonnet 5 counts the same text as roughly 30% more tokens, so the real saving is closer to 13% than 33%. This page works that through with numbers, then covers migration and cost control.

#What is Claude Sonnet 5?

Sonnet 5 sits between Haiku 4.5 below it and Opus 5 above. Anthropic's documentation describes it as "the best combination of speed and intelligence" and calls it the most agentic Sonnet so far. In practice it is the tier most production traffic should sit on: fast enough for interactive use, cheap enough to run at volume, capable enough that escalating is the exception.

Three specification changes matter more than the marketing:

• 1M context is the default. No beta header, no long-context price premium. A 900,000-token request bills at exactly the same per-token rate as a 9,000-token one. That was not true when 1M context first appeared in beta on Opus 4.6 in February 2026.
• Thinking is on by default. Sonnet 5 uses adaptive thinking rather than the manual budget_tokens approach it replaced. Depth is set with an effort parameter that defaults to high. See how adaptive and extended thinking work.
• Sampling parameters are gone. temperature, top_p and top_k return a 400 error if you pass a non-default value. Assistant message prefilling also returns 400.

Sonnet 5 is also the first Sonnet-tier model with real-time cybersecurity safeguards, which means it can return stop_reason: "refusal" as a successful HTTP 200 rather than an error. If your client treats any non-end_turn stop reason as a failure, fix that before you ship. The full model index lists which behaviours apply to which model.

#How much does Claude Sonnet 5 cost?

$2 per million input tokens and $10 per million output tokens, and as of 10 August 2026 that is the permanent price. Sonnet 5 launched with $2/$10 described as introductory pricing that would end on 31 August 2026, with a rise to $3/$15 scheduled for 1 September. Anthropic cancelled that rise and made the introductory rate standard. There is currently no announced successor price.

Two multipliers stack on top of everything above. Setting inference_geo: "us" for US-only inference applies a 1.1× multiplier to every token category. Regional and multi-region endpoints on Bedrock and Google Cloud carry a 10% premium over global endpoints. The full rate card, including subscription plans that have nothing to do with these rates, is on the pricing page.

#The tokenizer caveat: why $2 is not a 33% saving

Anthropic's own pricing documentation states that Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text". Sonnet 5 uses that tokenizer. Sonnet 4.6 and earlier use the previous one. So a price comparison of $2 against $3 compares two different units of measurement.

Here is the same workload priced both ways. Assume 1,000 requests, each sending 8,000 tokens of input and producing 800 tokens of output as counted by the Sonnet 4.6 tokenizer — the same prompts, the same documents, the same answers.

The saving is real but it is roughly a seventh, not a third. Anthropic notes that the exact increase "depends on the content and workload shape", so treat 30% as a planning figure and measure your own traffic before you commit a budget. Code, structured data and non-English text will not all move by the same amount.

Three consequences follow from the same fact, and all three catch people out:

• max_tokens now buys less prose. A limit of 4,096 that comfortably held a full answer on Sonnet 4.6 holds roughly 23% less text on Sonnet 5. Truncated outputs after a migration are usually this, not a model regression.
• The 1M window holds less text than it did. Both models advertise 1M tokens, but a million new-tokenizer tokens is around 23% less source material. If you were filling the window on Sonnet 4.6, you will overflow on Sonnet 5.
• Thinking tokens amplify the effect. Adaptive thinking is on by default and thinking tokens are billed as output at $10 per million. A workload that generates short answers can spend more on reasoning than on the answer. Set effort deliberately rather than accepting high.

#Migrating from Sonnet 4.6 to Sonnet 5

The model ID change is a one-line edit. Everything else on this list is a real behavioural difference that will bite in production if you skip it. Claude Code ships a /claude-api migrate skill that applies the ID swaps and breaking parameter changes across a codebase, which is the fastest route for a large repository — see Claude for developers.

A minimal Sonnet 5 request, with nothing in it that will 400:

from anthropic import Anthropic

client = Anthropic()  # reads ANTHROPIC_API_KEY

msg = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=2048,
    output_config={"effort": "low"},   # low | medium | high | xhigh | max
    messages=[
        {"role": "user", "content": "Summarise this contract clause in two sentences."}
    ],
)
print(msg.content[0].text)
print(msg.usage.input_tokens, msg.usage.output_tokens)

More request shapes, including streaming, tool use and error handling, are on the Claude API page, and the Sonnet-specific endpoint history is archived at the Sonnet API page.

#Cost-control architecture for Sonnet 5

Model choice is the smallest lever on your bill. The three that actually move it, in order of effect:

Combined, these compound. Take a support-assistant workload: a 3,000-token system prompt and knowledge base shared across every call, 600 tokens of unique user input, 400 tokens out. At list price that is $7.20 per 1,000 conversations. Cache the 3,000-token prefix and it falls to about $2.20. Move the intent-classification step that precedes it onto Haiku and the classification line costs cents. None of that required a different model for the hard part.

One more lever worth knowing: cache reads do not count toward your input-tokens-per-minute rate limit. Heavy caching therefore raises your effective throughput as well as lowering your bill. Patterns for orchestrating this across a fleet of calls are covered under automation and orchestration.

#Where Sonnet 5 loses

Stated plainly, because you should know before you standardise on it.

• Long-context retrieval is the weakest of the big three. On GDM-MRCR v2, Google scores 97.0% against Sonnet 5's 81.5%. A 1M-token window is not the same thing as reliable recall inside it. If your product is retrieval over very large corpora, test carefully — see Claude versus Gemini.
• No generative media, at all. Sonnet 5 produces text and code. It reads images and documents but cannot generate images, video, audio or music, and cannot accept video or audio as input.
• Its knowledge stops in January 2026. Counter-intuitively, Opus 5 has a newer training cutoff of May 2026 despite being the higher tier. If recency matters more than price, the more expensive model is also the better-informed one.
• No Priority Tier. Capacity guarantees are unavailable on this model.
• Exact benchmark scores are not published as numbers. Anthropic releases Sonnet 5 comparisons as chart images rather than machine-readable tables. Third-party figures circulating for Sonnet 5 on SWE-bench and similar are not traceable to a primary source, and this site does not repeat them.

If you are weighing Sonnet 5 against the tier above it, Fable 5 is five times the input price and five times the output price, and materially slower, at roughly 75 output tokens per second. It is a choice you make when a task failure costs more than the tokens, not a default. The predecessor's specification is archived at Claude Sonnet 4.6, and the wider version history at the Sonnet archive.

#Frequently asked questions