Anthropic released Claude Sonnet 5 on 30 June 2026 as its most agentic Sonnet yet, with performance it puts close to Opus 4.8, at $2 and $10 per million tokens. Start with boring, checkable work such as enquiry triage, quote drafts and weekly reports, with a person approving what customers see.
Start with boring, checkable work. Claude Sonnet 5, released by Anthropic on 30 June 2026, brings performance that Anthropic puts close to its Opus 4.8 model into the mid-priced tier, at $2 per million input tokens and $10 per million output tokens. For a small business, that makes multi-step jobs such as enquiry triage, quote drafting and weekly reporting cheap enough to automate, provided a person approves anything a customer will see.
The interesting part of the launch was the price tier, not the benchmark chart. Agent-grade ability, meaning a model that can plan a task, use your tools and finish the job, used to be priced as a flagship feature. When the mid-tier can do it, automation stops being an enterprise luxury. Here is what shipped, what it costs in practice and where to point it first.
- price per million input / output tokens, made permanent on 10 August
- $2 / $10
- price per million input / output tokens, made permanent on 10 August
- lower list price than Sonnet 4.6, which costs $3 / $15
- 33%
- lower list price than Sonnet 4.6, which costs $3 / $15
- token context window, roughly 555,000 words by Anthropic's estimate
- 1M
- token context window, roughly 555,000 words by Anthropic's estimate
- as many tokens for the same text, at the top of Anthropic's range for the new tokeniser
- 1.35x
- as many tokens for the same text, at the top of Anthropic's range for the new tokeniser
01
What Anthropic shipped
Sonnet 5 is the cheapest of Anthropic's new 5 series models, below Fable 5, released on 9 June as Anthropic's most capable model for all customers, and Opus 5, released on 24 July. Across the full Claude range it sits in the middle, above Haiku 4.5. It reads text and images, holds up to 1 million tokens of context and can write up to 128,000 tokens in one response. It is also the default model on Claude's Free and Pro plans, so many people now use it without ever choosing it.
Anthropic says its performance is close to Opus 4.8, the top Opus model until July, which lists at two and a half times the price. It also says Sonnet 5 resists prompt-injection hijacks better and hallucinates less than Sonnet 4.6. The launch price was billed as introductory until 31 August, with a rise to $3 / $15 planned for 1 September. On 10 August Anthropic cancelled the rise and made $2 / $10 the standard price.
Claude Sonnet 5 is built to be the most agentic Sonnet model yet.
02
What agentic means, minus the marketing
Anthropic's own engineers draw a useful line. A workflow runs a model and tools along steps you have written in advance. An agent decides its own steps, and which tools to use, as it goes. Agents suit open-ended problems where you cannot predict the number of steps; anything with a fixed path is cheaper and more predictable as a workflow. Their advice is to find the simplest solution that works and add complexity only when it is needed.
Most small business wins sit in the first category: a fixed process with one model step that reads, classifies or drafts. That is good news, because a fixed process is easier to test and easier to explain to staff.
Agent features are also arriving inside everyday software. Anthropic's Managed Agents service, in public beta since April, added scheduled runs on 9 June. OpenAI's GPT-5.6, released on 9 July, added multi-agent orchestration in beta, and SpaceXAI pitched Grok 4.5 at office and clerical work as well as coding. The ability is becoming a standard feature, so the real question is where you let it act.
03
The new arithmetic
Here is how the Claude range was priced in August 2026, per million tokens of input and output.
| Model | Released | Input / output per million tokens |
|---|---|---|
| Claude Haiku 4.5 | October 2025 | $1 / $5 |
| Claude Sonnet 4.6 | February 2026 | $3 / $15 |
| Claude Sonnet 5 | 30 June 2026 | $2 / $10 |
| Claude Opus 4.8 | 28 May 2026 | $5 / $25 |
| Claude Opus 5 | 24 July 2026 | $5 / $25 |
| Claude Fable 5 | 9 June 2026 | $10 / $50 |
A hypothetical example shows the scale. Suppose a Kochi travel agency receives 1,500 enquiries a month by email and WhatsApp. If each triage run reads 3,000 tokens (the message, the agency's service list and its instructions) and writes 400 (a category, a priority and a draft reply), the month uses 4.5 million input tokens and 0.6 million output tokens: about $9 plus $6, or $15 at list price. Allow a third more for the new tokeniser and it is still about $20.
At that volume the model is rarely the expensive part; building, testing and reviewing the workflow is. Two settings cut the bill further: the Batch API halves the price for jobs that can wait, such as overnight report assembly, and prompt caching charges a tenth of the normal input price when the same instructions are reused.
04
Where to point it first
The best first agents are boring. Each of these has clear inputs, a checkable output and a person who already does the job today.
- Enquiry triage. Read each new email or WhatsApp message, classify it, pull out the name, need and urgency, and route it to the right person with a suggested reply.
- Quote drafting. Build a first-draft quote from your actual price list and the client's request, for a salesperson to check and send.
- Weekly report assembly. Pull figures from the sheets and dashboards you already use and write the summary your manager reads on Monday.
- Document checks. Compare an incoming purchase order or invoice against your records and flag mismatches for a person to resolve.
Keep a human on the send button for anything customer-facing. The goal is drafts a person can approve in seconds, not autonomous replies you discover a week later. Our guide to guardrails for customer-facing AI covers the controls, and when AI automation is worth it covers the payback test.
05
The risks that come with agents
The main new risk is prompt injection. When a model reads outside content, such as an email, a web page or an uploaded file, instructions hidden in that content can steer it. The OWASP Gen AI Security Project ranks prompt injection first in its 2025 Top 10 risks for large language model applications, and recommends giving the model only the access it needs, clearly separating untrusted content and requiring human approval for high-risk actions.
Anthropic says Sonnet 5 resists these hijack attempts better than Sonnet 4.6, but no model is immune. Design as if the model can be fooled: let it read and draft, and keep payments, deletions and emails to customers behind an approval step.
06
What to do this month
- Pick one workflow with clear inputs, a clear output and a person who checks it.
- Measure the baseline today: hours per week spent on it and how often mistakes happen.
- Pilot on Sonnet 5 with the model ID pinned, and log every input and output.
- Give the model the fewest permissions that work: read access first, and no sending or paying without approval.
- Count tokens on 20 real examples and add a buffer of about a third to your budget.
- Review a sample of outputs every week for the first month, then decide whether to expand.
Sources
Facts current as of 12 August 2026. Prices are list prices in US dollars; check Anthropic's pricing page before you budget.