What Is New in Claude Sonnet 5?
Claude Sonnet 5 is Anthropic's mid-tier model released on 30 June 2026. It pairs a 1 million-token context window with adaptive thinking that is on by default, and Anthropic positions it as its most agentic Sonnet yet, performing close to the larger Opus 4.8 on reasoning, coding and knowledge work while costing far less to run.
The short version: you get near flagship-level output from a model priced like a workhorse.
The three changes that matter most for daily work are the 1M context window, adaptive thinking, and stronger tool use. Each one removes a friction point you have probably hit already.
Worth noting on the technical side: Sonnet 5 keeps the same tools and platform features as Sonnet 4.6, so nothing you already built breaks. The API model ID is simply claude-sonnet-5, and adaptive thinking is on without you enabling anything.
If you have used an earlier Sonnet or a competing mid-tier model and felt it was capable but a step behind the flagship, this release is aimed squarely at closing that gap. The rest of this guide walks through what each change unlocks and where it still trips up.
Why Does the 1 Million Token Context Window Matter?
A 1 million-token context window means Sonnet 5 can hold roughly 750,000 words in a single conversation without forgetting the start. For you, that means dropping an entire quarter of meeting notes, a full contract, or a 300-page report into one chat and asking questions across all of it at once.
Before this, long documents forced you to chunk material and paste it in pieces. The model would lose track of section three by the time it reached section nine.
Now you can load the whole thing. Anthropic confirms Sonnet 5 carries the full 1M window with no long-context pricing premium, so a big document does not cost extra per token to process.
A concrete example: paste a full year of customer feedback, say 5,000 short comments, and ask the model to cluster them into the ten most common complaints with a representative quote for each. That used to require a separate data pipeline. Now it is one prompt.
The practical payoff is consistency. When the model sees the entire source at once, its answers stop contradicting each other across a long working session.
It also changes how you research. Instead of summarising five reports one at a time and losing the thread, you load all five and ask for the points where they disagree, which is usually where the interesting insight sits.
How Do You Use Adaptive Thinking Effectively?
Adaptive thinking lets Sonnet 5 decide how much internal reasoning a task needs before it answers, and it is switched on by default. Simple questions get fast replies; a multi-step analysis triggers a longer reasoning pass. You do not toggle a setting, you shape it through how you phrase the request.
The lever you control is instruction clarity. If you want deeper reasoning, tell the model to work through the problem step by step and to check its own logic before giving a final answer.
If you want speed, say so directly: ask for a one-line answer with no explanation. The model matches its effort to your stated expectation.
The difference is visible in practice. A vague prompt like "is this contract okay?" gets a shallow scan. A framed prompt like "check this contract for three specific risks: auto-renewal terms, liability caps, and termination notice periods, and quote the exact clause for each" forces a deeper reasoning pass and a far more useful answer.
This is the same principle behind context engineering: the quality of what you get out depends on how precisely you frame what goes in.
What Can You Actually Build with Sonnet 5 Today?
Sonnet 5 is strongest at long-document analysis, multi-step drafting, and tool-driven workflows where it reads, plans and acts across several steps. In practice that covers reviewing a full RFP, turning a research folder into a briefing, or drafting a report that pulls from many sources at once.
Here is a concrete workplace scenario. You have a 40-page client proposal and need a risk summary before a meeting in an hour.
Paste the full document, then use this prompt:
Try This Prompt:
"You are reviewing the attached 40-page proposal for risks. Work through it section by section. For each risk you find, give: (1) the exact section and page, (2) the risk in one sentence, (3) how serious it is on a low or medium or high scale, and (4) one suggested mitigation. List the top five risks only, ordered by severity. Do not summarise the whole document, only surface risks."
Because Sonnet 5 holds the full document in context, it can cite the exact section rather than guessing, and adaptive thinking gives it room to reason before ranking.
A second common scenario is turning scattered inputs into one clean output. Suppose you have a folder of six research links, two transcripts, and a rough outline. You can paste all of it and ask for a single two-page briefing that reconciles the sources and flags any contradictions, in one pass rather than six.
The pattern to learn here is loading everything, then asking a narrow question. The model is strongest when it has full context but a tightly scoped task, not when it has to guess what you care about.
Where Does Claude Sonnet 5 Still Fall Short?
Sonnet 5 is strong but not flawless. A 1M window does not mean it weighs every token equally, information buried in the middle of a very long input still gets less attention than material at the start or end. This is a known pattern across long-context models, not a Sonnet-specific bug.
So do not assume that loading a huge document guarantees the model noticed one buried clause. If a detail is critical, call it out explicitly in your prompt.
It can also still produce confident but wrong statements when a source is ambiguous. Adaptive thinking reduces this on reasoning tasks, but it does not remove the need for you to verify facts and figures.
Finally, the introductory API price of $2 per million input and $10 per million output tokens runs only through 31 August 2026. After that it moves to $3 and $15, so budget on the standard rate if you are planning ongoing use.
How Do You Decide When to Use Sonnet 5 Over a Cheaper Model?
Use Sonnet 5 when a task needs real reasoning, long context, or multi-step tool use, and drop to a smaller model when you only need short, simple replies. The decision comes down to whether the cost of a wrong or shallow answer is higher than the cost of the extra tokens.
For high-stakes drafting, contract review, or analysis that feeds a decision, the stronger model earns its price.
For quick reformatting, short replies, or bulk low-risk tasks, a lighter model is the sensible call. Matching the model to the job is itself a skill, and one you can pressure-test with UD's AI IQ Test.
A rough cost check makes this concrete. At the introductory rate, processing a 50,000-token document and getting a 2,000-token answer costs about 12 US cents. If that answer saves you an hour of manual review, the model is not the expensive part of the task, your time is.
The mistake to avoid is defaulting to the strongest model for everything. That habit quietly inflates your token bill on tasks a cheaper model would have handled identically.
Try It Now
Take a document you already know well, a report or proposal you wrote, and paste it into Claude Sonnet 5. Ask it to find the three weakest arguments and explain why each is weak. Because you know the material, you can judge the quality of its reasoning immediately, and you will learn fast where the model is sharp and where it needs a tighter prompt.
That single exercise teaches you more about the model in ten minutes than any feature list.
Then push it further. Ask it to rewrite the weakest argument so it holds up, and watch whether the fix is genuinely stronger or just reworded. That is the fastest way to feel where the model reasons well and where it merely sounds confident.
The real shift with Sonnet 5 is not a new button to press. It is that near flagship reasoning is now cheap enough to use on ordinary daily work, not just the occasional big task. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Put Claude Sonnet 5 to Work in Your Team
Knowing the model is one thing. Building it into a workflow your whole team can rely on is another. UD helps you deploy AI that works like a dedicated employee, and we'll walk you through every step, from choosing the right model to designing the workflow and putting it into daily use.
Reviewed by the UD AI team.
\n
