Anthropic Cut Its Best Agent Model's Price In Half. What It Means For Your Business.

On June 30, Anthropic released Claude Sonnet 5. It is now the default model on the Free and Pro plans, priced at $2 per million input tokens and $10 per million output tokens through August 31, then $3/$15 after. Anthropic says higher-effort runs "can match Opus 4.8 on some tasks" at less than half the cost. Source: Anthropic's announcement.

You did not sign up for a new subscription. You did not change any settings. But if you use Claude, ChatGPT-style tools with Claude behind them, or a vendor that wraps a Claude API, the model doing the work under the hood just got better at running multi-step tasks and (usually) cheaper to run. That last part matters more than the headline.

#What actually changed

The old story: cheaper models were fine for chat, but you had to pay for Opus to get real agent behaviour, the kind where the model opens a browser, reads a spreadsheet, updates a record, and moves on. That gap just narrowed.

Anthropic's own benchmarks focus on agentic work, not chat quality:

• BrowseComp (agentic web browsing)
• OSWorld-Verified (using a computer like a person)
• Humanity's Last Exam (long-horizon reasoning)

Sonnet 4.6 already scored 78.5% on OSWorld-Verified. Sonnet 5 pushes further and closes the gap to Opus 4.8, at less than a third of Opus's price when you use standard pricing after August. The word Anthropic keeps using is "agentic." Translation: the model can be trusted to run a task on its own for longer without a human intervening.

#Why a small business owner should care

Two reasons, in this order.

1. Your vendors' costs just dropped. If your CRM has an "AI assistant," your helpdesk has "AI ticket triage," or your automation tool has an "AI agent" feature, there is a decent chance Claude is behind it. Vendors do not usually pass model price cuts to you automatically. Some do. Most raise their margin quietly. Ask.

2. Tasks you were told "AI cannot handle yet" are on the table again. Multi-step ops work (reading an invoice PDF, checking it against a PO, emailing the vendor if wrong) got cheaper AND more reliable in the same release. Six months ago the answer might have been "wait." Now it might not be.

#Three questions to ask your AI vendor this week

Send these, verbatim. If they cannot answer clearly, you have a signal about how much they actually know.

1. "Are you on Claude Sonnet 5 yet? If not, when?" The rollout is API-level. Any vendor using Claude can flip a config change. If they say "we are evaluating," ask what the evaluation criteria are.

2. "Did my per-transaction AI cost drop with this release? If yes, is that reflected in my pricing?" Most SaaS pricing is fixed per-seat, so the answer will be "no, but our margins improved and here is what we can do." That is a fair answer. It gives you a lever for a renegotiation or an upgrade at the same price.

3. "What is now possible that was not possible two months ago?" Vendors who cannot answer this are not paying attention to what they sell. The good ones will have a shortlist of new automations they are testing.

#The specific tasks worth revisiting

If you looked at these six months ago and decided "not quite reliable enough," reopen the file:

• Vendor invoice reconciliation. Read the PDF, match line items to the PO, flag mismatches, draft the reply.
• Inbound lead qualification. Read the form submission plus any linked website or LinkedIn, decide if it fits your ICP, assign a lead score, route to the right person.
• Customer support triage across channels. Email, chat, WhatsApp, and voice transcripts merged into one queue with proposed responses per ticket.
• Weekly financial roll-up. Pull the numbers from Stripe, Xero, and your bank, produce a one-page summary with a plain-English "here is what changed and why."

None of these are new ideas. They are the same tasks that were "almost there" and now cross the reliability threshold for more businesses.

#The date to circle

August 31, 2026. That is when Sonnet 5's introductory pricing ends and standard pricing kicks in ($3 input / $15 output per million tokens). If you are budgeting agent spend for Q4, use the standard number, not the intro number. Vendors who bake their pricing on intro rates and forget to update will get squeezed in September.

#Should you switch models?

Only if a specific job is not working, or you are paying for Opus and the task is squarely in Sonnet 5's range. There is no reason to switch for the sake of switching. The right move for most small businesses is not "change your setup" but "revisit the shortlist of tasks you shelved for reliability reasons and try them again."

If you want a second opinion on which of your current tasks are ready to hand to an agent now, book a call. We do a free technical analysis. No pitch attached.