Claude Sonnet 5 is here.
It is Anthropic's most agentic Sonnet-tier model yet: frontier-level coding, tool use, reasoning, and knowledge work at mid-tier speed and price. It narrows the gap to Opus 4.8 while staying faster and cheaper. Here is what it is, how to start using it today, and what to actually do with it.
Plain-English explainer · written for operators and small teams · by the BrainVaultAI team
What it is and what changed
The everyday model just got a lot more capable
Sonnet is the tier most people use most of the time: fast enough for real work, cheap enough to run at volume, and now smart enough to handle tasks that used to need the top-tier model. Sonnet 5 is the new default model for Free and Pro users across the Claude apps, Claude Code, the API, and managed agents. It is also rolling out to partners including Cursor, Notion, GitHub Copilot, Devin, and OpenRouter, so you may already be using it without changing a thing.
Compared with Sonnet 4.6, the practical gains are about autonomy and follow-through. It handles complex, multi-step tasks better and keeps going to the end instead of stalling halfway. It checks its own output. It is more comfortable in messy, real-world ("brownfield") code: debugging, tracing a bug to its root cause, and working in systems that were not built cleanly. Tool use and computer interaction are stronger too, so it can drive a browser or terminal and run end-to-end automation, getting more done in fewer steps. Refusals are cleaner: it declines malicious requests more reliably with less hand-wringing.
1M token context
A native one-million-token context window. You can hand it large codebases, long documents, or months of notes in a single conversation.
Adaptive thinking
It reasons more on hard problems and less on easy ones, with a new “xHigh” reasoning-effort tier and high effort as the default in many interfaces.
High-resolution vision
Stronger image understanding for screenshots, diagrams, documents, and UI, which matters for computer-use automation.
Stronger safety profile
Better refusal of malicious requests, lower hallucination and sycophancy than Sonnet 4.6, with cyber safeguards on by default.
On benchmarks: Anthropic reports strong gains on agentic coding (SWE-bench), computer use (OSWorld), and reasoning (GPQA) evaluations. For the exact numbers, read Anthropic's official announcement and system card rather than secondhand figures.
We are not quoting specific benchmark scores here on purpose. We only publish numbers we can point to in a primary source.
How to start using it today
You probably already have access
For most people there is nothing to install and nothing to switch on. Sonnet 5 is the default, so the fastest path is simply to use the tools you already have and notice that the answers got better.
- 1
In the Claude apps
Open Claude on the web, desktop, or mobile. Free and Pro accounts now run Sonnet 5 by default. Start a normal chat. For harder tasks, raise the reasoning effort if your interface exposes that control.
- 2
In Claude Code
If you write or ship software, Claude Code uses Sonnet 5 as its default model. Point it at a real repository and ask it to fix a bug, add a feature, or explain a confusing file. It is built to work through multi-step tasks and check its own output.
- 3
Through the API
Developers call the model with the id claude-sonnet-5. Standard pricing is $3 per million input tokens and $15 per million output tokens, with an introductory rate of $2 / $10 per million tokens through August 31, 2026.
- 4
Pick an effort level
Sonnet 5 supports adaptive thinking, including a new xHigh tier for the hardest problems. Many interfaces default to high effort. Use more effort for thorny reasoning or debugging, and less for quick, well-defined tasks to keep things fast.
One thing to know about the bill
Sonnet 5 ships with a new tokenizer that can count roughly 1.0 to 1.35 times more tokens for the same input. The introductory pricing is set to offset that, so most users will not see a real cost increase, but it is worth knowing if you track token usage closely.
Use cases
What to actually do with it
You do not need to be technical to get value here. These are concrete jobs Sonnet 5 is well suited for, especially for solo operators and small teams.
Drafting
Turn a rough outline into a clean first draft: proposals, sales pages, job posts, SOPs. It follows instructions through long documents without losing the thread.
Research
Paste long reports, transcripts, or threads and ask for the signal: a summary, the open questions, the decisions you need to make. The 1M context window means you rarely have to chop the input into pieces.
Inbox and meeting prep
Have it triage a messy inbox, draft replies in your voice, and build a tight pre-read before a call: who you are meeting, what changed since last time, and the three things to cover.
A simple AI agent for a recurring task
Wire up one job you do every week (sorting leads, formatting a report, answering a common question) and let an agent handle it. Sonnet 5 is built for exactly this kind of cost-efficient, repeated automation.
Computer-use automation
With stronger tool use and vision, it can drive a browser or terminal to complete an end-to-end task: pull data from a site, fill a form, run a routine, and report back. Start small and supervised.
Coding and debugging
Even if you do not code, this matters: it can read an existing codebase, find the root cause of a bug, and fix it. For technical teams it is a capable coding-agent default.
Which model to use
Sonnet 5 vs Opus 4.8 vs Fable 5
A plain rule: reach for Sonnet 5 first. Step up to a top-tier model only when the task is genuinely hard, long-horizon, or high-stakes.
Best for everyday high-volume work, coding agents, computer use, professional and knowledge tasks, and cost-efficient autonomy. Fast, affordable, and now close to top-tier on a lot of real work.
Reach for it when the task is extremely complex, long-horizon, or high-stakes and getting it wrong is expensive. Slower and pricier, but the deepest reasoning when you genuinely need it.
Anthropic’s new flagship for the most demanding, high-stakes work. Like Opus, treat it as the model you escalate to, not your default.
The short version: default to Sonnet 5. If it stalls on something truly complex, long-running, or high-stakes, escalate to Opus 4.8 or Fable 5.
From BrainVaultAI
A new model is only useful if you actually use it
BrainVaultAI teaches non-technical operators to put models like Sonnet 5 to work on real business tasks: drafting, research, and building simple AI agents for the jobs you repeat every week. Start with the free guides, or apply for the 12-week AI Certification and build a working AI system with guidance.
Get the next model breakdown the day it ships
When Anthropic ships a new model, we send a plain-English breakdown like this one, plus a short series of our highest-leverage AI plays for operators. No spam. Unsubscribe in one click.
Free. No spam. Unsubscribe in one click. By submitting you agree to receive emails from BrainVaultAI.
More free AI guides in the full library · follow @brainVaultAI