What is Claude Sonnet 5?
Claude Sonnet 5 is Anthropic's most agentic mid-tier model, released on 30 June 2026. It can plan, use tools like browsers and terminals, and run tasks autonomously at a level that previously needed larger models, while costing far less than the flagship Opus 4.8.
If you juggle Claude, ChatGPT, and Gemini and never quite know which to open for which job, this release changes the maths. Sonnet 5 is built to be the default workhorse, not the compromise.
The short version: near top-tier capability, mid-tier price, with adaptive thinking always switched on. The rest of this guide is about when that trade is worth it.
What's actually new in Claude Sonnet 5?
Three things stand out. Sonnet 5 ships with a 1 million token context window, a maximum output of 128k tokens, and adaptive thinking that is always on, defaulting to high effort on the API and in Claude Code. Anthropic also reports fewer undesirable behaviours than Sonnet 4.6.
The 1M context window means you can drop an entire project, a long report, or months of notes into a single conversation without splitting it.
Always-on adaptive thinking means the model decides how hard to reason per request, rather than you toggling a setting. For practitioners this removes a step you used to manage by hand.
Anthropic's own launch page notes Sonnet 5 shows an overall lower rate of undesirable behaviours than Sonnet 4.6, which matters when you let a model run tools on its own.
How does Claude Sonnet 5 compare to Opus 4.8 and Sonnet 4.6?
On agentic coding, Sonnet 5 scores 63.2%, above Sonnet 4.6 at 58.1% but below Opus 4.8 at 69.2%. On Terminal-Bench 2.1 it hits 80.4%, actually beating Opus 4.8's 74.6%. On a knowledge-work benchmark, Sonnet 5 slightly edges out Opus 4.8.
Read that carefully. Sonnet 5 is not simply "worse Opus." It trails on some coding tasks yet leads on terminal work and knowledge tasks.
Other reported figures include 81.2% on OSWorld-Verified, close to Opus 4.8's 83.4%, and 84.7% on BrowseComp agentic search, per MarkTechPost's 13 July 2026 comparison.
The practical takeaway: for most day-to-day knowledge and browsing work, Sonnet 5 is at or near Opus level. Reach for Opus 4.8 only when a hard coding task justifies the premium.
How much does Claude Sonnet 5 cost?
Sonnet 5 launched at an introductory $2 per million input tokens and $10 per million output tokens through 31 August 2026. After that it moves to standard pricing of $3 per million input and $15 per million output, per Anthropic's pricing.
There is a catch worth knowing. Sonnet 5 uses a new tokenizer, and the same text can map to 1.0 to 1.35 times more tokens than on previous models.
So the headline price looks cheaper, but your real bill depends on token counts that may run higher. Compare on a real task, not on the per-token rate alone.
For a marketer or ops manager, the honest move is to run one representative job and read the actual token usage before committing a workflow to Sonnet 5.
When should you use Sonnet 5 instead of another model?
Use Sonnet 5 as your default for agentic and multi-step work: browsing, tool use, terminal tasks, and long-document analysis where its 1M context and strong terminal scores shine. Escalate to Opus 4.8 only for the hardest coding, and keep a fast, cheap model for trivial lookups.
Here is a copy-paste prompt to benchmark Sonnet 5 on your own real task:
You are being evaluated for a repeated task in my workflow.
TASK: {paste one real task you do weekly, with real inputs}
Please:
1. Complete the task to production quality.
2. At the end, list which tools or steps you used.
3. State one thing that would have made the task easier or more reliable.
Do not ask clarifying questions unless the task is impossible without them.
Run that same block on Sonnet 5 and on whatever you use now. Compare output quality, then compare the token usage each reported. That side-by-side, on your own work, beats any benchmark table.
What are the gotchas to watch for?
The biggest is the tokenizer change: budgeting on the old per-token intuition will understate your spend by up to a third. Always measure real token counts on a live task before rolling Sonnet 5 into a recurring workflow.
The introductory price ends 31 August 2026. Any cost case you build now should also pencil out at the standard $3 and $15 rates.
Always-on adaptive thinking defaults to high effort, which is great for quality but can add latency and cost on simple requests. For trivial lookups, a lighter model is still the smarter pick.
At UD, we understand AI. We understand you better. With UD by your side, AI doesn't feel cold, and choosing the right model for each job stops being guesswork and becomes a decision you can actually defend.
Find the Right Model for Every Job
New models launch almost monthly, and picking the right one for each task is its own skill. UD's team will walk you through every step, from benchmarking models on your real workflows to setting up a simple decision framework your whole team can follow.