What Is Claude Sonnet 5?
Claude Sonnet 5 is Anthropic's new mid-tier model, released on June 30, 2026. It replaces Sonnet 4.6 as the default model on Claude's Free and Pro plans, brings a 1 million token context window as standard, and performs close to the flagship Opus 4.8 on reasoning, tool use, and coding, at a lower price.
If you use Claude daily, this update landed in your account already. The question is not whether you have it. The question is whether you know which changes actually affect how you work.
Most model launches are benchmark theatre. This one changes three things practitioners touch every day: how the model decides when to think, how much text it can hold, and how usage is counted. Each has a practical consequence worth knowing.
What Actually Changed From Sonnet 4.6?
The headline change is agentic capability. According to Anthropic's launch announcement, Sonnet 5 plans multi-step tasks, uses tools like browsers and terminals, and runs autonomously at a level that previously required larger, more expensive models. Its performance approaches Opus 4.8 on reasoning, tool use, coding, and knowledge work.
In practice, that means the model finishes tasks that older Sonnet versions abandoned halfway. Early testers quoted by Anthropic described it completing two-part jobs end to end, checking its own output without being asked, and staying on plan through multi-step work.
For non-developers, the translation is simple: workflows that chain several steps together, such as "read these three reports, extract the pricing data, then draft a comparison table," are now noticeably more reliable on the default model you already have.
Anthropic also reports lower rates of hallucination and sycophancy compared with Sonnet 4.6. The sycophancy point matters for anyone using Claude as a feedback tool: it is now less likely to tell you your draft is brilliant when it is not.
How Does the 1M Token Context Window Help Your Daily Work?
Sonnet 5 supports a 1 million token context window by default, with 128,000 maximum output tokens. There is no smaller variant. Anthropic charges standard pricing across the entire window, which makes large-context work economical for the first time at this tier.
A context window is the model's working memory: everything it can see at once. One million tokens covers roughly the length of several full-length books, so you can load an entire project's documents into a single conversation.
Concrete uses worth trying this week:
--- Load a full year of meeting notes and ask for every commitment made to a specific client
--- Drop in a 200-page contract plus your company's standard terms and ask for a clause-by-clause deviation list
--- Paste an entire content archive and ask which topics you have never covered
One caveat: a bigger window does not mean the model weighs everything in it equally. Long-context research consistently shows models attend most strongly to the beginning and end of the input. Put your instructions and the most important material at the start, and repeat the key question at the end.
What Does Adaptive Thinking Mean for Your Prompts?
Adaptive thinking is on by default in Sonnet 5. The model decides for itself when a task needs slower, deeper reasoning and when a fast answer is fine. On earlier models, you had to manually switch on extended thinking for hard problems; per Anthropic's documentation, that manual mode is now removed entirely.
For everyday use, this removes a decision you probably were not making well anyway. Complex questions automatically get more reasoning effort. Simple ones stay fast.
The practical adjustment: stop writing "think step by step" as a magic incantation on every prompt. Sonnet 5 already allocates thinking effort based on task difficulty. Your prompt real estate is better spent on context, examples, and clear success criteria.
For those using the API or building automations, note that manually setting temperature or other sampling parameters now returns an error on this model. Behaviour steering happens through instructions, not dials.
How Much Does Claude Sonnet 5 Cost?
Sonnet 5 launched with introductory API pricing of US$2 per million input tokens and US$10 per million output tokens, valid through August 31, 2026. After that, standard pricing is US$3 and US$15, the same rates as Sonnet 4.6. On the Claude apps, it is included in existing plans and is the default for Free and Pro.
There is a catch worth understanding: the new tokenizer. According to Anthropic's documentation, the same text now produces approximately 30% more tokens than on Sonnet 4.6. Per-token prices are unchanged, but an equivalent request can cost more because it contains more tokens.
Anthropic states the introductory pricing was set so the transition is roughly cost-neutral. The practical takeaway for API users: recount your token budgets against the new model rather than reusing old estimates, and revisit any output limits you tuned for Sonnet 4.6.
What Should You Watch Out For?
Three gotchas are worth flagging before you rebuild workflows around this model.
The tokenizer affects text capacity. The window is 1 million tokens, but each token now covers less text on average. The same window holds somewhat less raw text than the number suggests.
Cybersecurity requests may be refused. Sonnet 5 is the first Sonnet-tier model with real-time cyber safeguards enabled by default. If your legitimate work touches security topics, expect occasional refusals, and note that Anthropic recommends its Opus models for sanctioned security work.
Adaptive thinking uses output tokens. Thinking counts toward the output budget. If an automation caps output length tightly, the model may spend the budget reasoning and truncate the answer. Give long tasks generous output room.
Try It Now: A 1M-Context Exercise
The fastest way to feel the difference is a cross-document task that older models handled poorly. Gather 5 to 10 long documents from one project, paste or upload them into a single conversation, then run this prompt:
Try This Prompt:
"You now have every key document from this project in context. Before answering, list the documents you can see so I can confirm nothing is missing. Then: (1) identify the three biggest inconsistencies between these documents, quoting the exact conflicting passages; (2) list every deadline or commitment mentioned, with its source document; (3) tell me which single document is most out of date and why. Present each part under its own heading. If information is ambiguous, say so rather than guessing."
The confirmation step at the start matters: it verifies the model actually ingested everything before you rely on its analysis. The instruction to quote exact passages keeps it honest, and the permission to flag ambiguity reduces confident guessing.
The Bottom Line
Sonnet 5 is the rare update where the default option became genuinely strong. Near-flagship capability is now the baseline on free and standard plans, and the 1 million token window turns whole-project analysis from a premium trick into an everyday habit.
The winners will be the people who adapt their workflows first: front-load context, drop the ritual prompt phrases, and start feeding the model whole projects instead of fragments. UD has been by your side for 28 years, making technology feel human. With UD, AI works for you, not the other way around.
Put Sonnet-Class AI to Work in Your Business
Knowing what a new model can do is step one. Turning it into a workflow that runs reliably every day is where the real gains are. We'll walk you through every step, from tool setup to workflow design and deployment, so the upgrade actually shows up in your output.