AI Workflows / For Every Professional
Claude Sonnet 5: First Look at the New Free Agentic Model
Anthropic made Sonnet 5 the default model for Free and Pro users, priced well below Opus, and pitched it as the most capable Sonnet yet for real agentic work.
On June 30, 2026, Anthropic released Claude Sonnet 5 and made it the default model for Free and Pro users. Anthropic calls it its most agentic Sonnet yet, with performance close to Opus 4.8 at a much lower price: introductory API rates of $2 per million input tokens and $10 per million output tokens through August 31, 2026, then $3 and $15. It is available across all plans, Claude Code, and the Claude Platform as claude-sonnet-5.
The short version
On June 30, 2026, Anthropic released Claude Sonnet 5 and made it the default model for Free and Pro users. That last part is the news. Most people who open Claude today are already talking to Sonnet 5 without changing a setting. Anthropic calls it the most agentic Sonnet yet, meaning it is built to make plans, use tools like browsers and terminals, and finish multi-step tasks that used to require the larger and pricier Opus line.
For a working professional, the question is not which benchmark moved. It is simpler: does my everyday Claude get better, and does anything I built on top of it get cheaper or more reliable. The answer is mostly yes, with a couple of caveats worth understanding before you rewire a workflow around it.
What actually changed
Three things shipped at once.
First, Sonnet 5 replaced the previous default. Free and Pro users get it automatically, and Max, Team, and Enterprise users have it available too. It also runs in [Claude Code](/ai-workflows/) and on the Claude Platform for developers, where you call it as `claude-sonnet-5`.
Second, the price. At launch Anthropic set introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that it moves to $3 per million input and $15 per million output. For comparison, Anthropic prices Opus 4.8 at $5 per million input and $25 per million output. So Sonnet 5 lands well under the flagship: roughly 60 percent cheaper at the intro rate, and about 40 percent cheaper once standard pricing kicks in.
Third, the capability gap narrowed. Anthropic's own benchmarks position Sonnet 5 as approaching Opus 4.8 on some tasks while costing far less, and it calls the model a clear step up from its predecessor, Sonnet 4.6, on reasoning, tool use, coding, and knowledge work. TechCrunch, citing Anthropic's own numbers, reported one agentic coding benchmark where Sonnet 5 scored 63.2 percent against Opus 4.8's 69.2 percent and Sonnet 4.6's 58.1 percent, and a knowledge work benchmark where Sonnet 5 edged slightly ahead of Opus. Treat those as vendor figures, because they are. Anthropic reports them; no independent lab has confirmed them yet.
The cost math, done honestly
The headline is that a near-flagship model now sits at Sonnet prices. The footnote matters just as much.
Sonnet 5 uses an updated tokenizer, the same kind of change Anthropic made with Opus 4.7. It processes text differently to improve quality, and the tradeoff is that the same input can map to more tokens, roughly 1.0 to 1.35 times as many depending on the content. In practice, the lower rate plus the new tokenizer may be roughly cost neutral for many users moving from Sonnet 4.6. In plain terms: your per-token price went down, but your token count may go up, so your actual bill might barely move even though the rate card looks cheaper.
Here is a sober way to think about it if you run Claude through the API:
- Do not assume your bill drops 40 to 60 percent. Measure it. Run a representative batch of your real prompts through Sonnet 5 and compare the total token count and total cost against your Sonnet 4.6 baseline, not the advertised per-token rate.
- Watch the output side. Output tokens cost five times input tokens, and an agentic model that "keeps going" can produce more of them. More autonomy can mean more tokens.
- Use the intro window deliberately. The $2 and $10 rate holds only through August 31, 2026. If you have a large batch job or an evaluation you were going to run anyway, running it before that date is cheaper.
When to use Sonnet 5, and when to reach for Opus
The most useful thing Anthropic ships alongside Sonnet 5 is not a benchmark. It is the framing that you now choose a point on a cost-and-quality curve rather than just picking a model name.
Anthropic exposes an effort setting that lets you dial Sonnet 5 up or down. At medium effort it is far cheaper than Opus 4.8 for similar work. At higher effort it costs more but, on some tasks, approaches Opus quality. Opus 4.8 remains the pick when you need the highest accuracy on the hardest judgment calls and deepest research.
A practical routing rule for professional work:
- Reach for Sonnet 5 by default. Drafting, summarizing, first-pass research, code changes, data cleanup, routine analysis, and most agent automations. This is the workhorse, and for the vast majority of daily tasks it is enough.
- Raise the effort level when a Sonnet task is close but not quite landing, before you switch models. Often that closes the gap for less than an Opus call would cost.
- Escalate to Opus 4.8 for the small share of work where being wrong is expensive: a bet-the-company analysis, a subtle legal or financial judgment, deep multi-hour research where one bad inference poisons everything downstream.
- For anything sensitive, keep a human on the final decision regardless of model. Cheaper autonomy makes it tempting to let the agent run unattended. Resist that on high-stakes work.
If you want the full decision framework for matching a model to a task and building repeatable workflows around it, that is the core of what we teach in [AI for Managers](/ai-for-managers). Not sure where to start? The [two-minute quiz](/quiz) points you to the right course.
What "most agentic Sonnet yet" means in practice
Agentic is an overused word, so here is the concrete version. Anthropic's early partners, quoted in its launch post and in TechCrunch, describe Sonnet 5 finishing tasks that earlier Sonnet models would abandon halfway, and checking its own output without being told to.
One example TechCrunch cited: a Zapier engineer said the model was handed a two-part job, update Salesforce account tiers and send a launch announcement to enterprise contacts, and it finished end to end where the work "used to stall halfway." Lovable's co-founder said it "refuses unsafe requests cleanly and consistently." These are testimonials from a vendor announcement, so read them as directional rather than proof. But the pattern they point to, follow-through across several steps without a human nudging it back on task, is the thing that actually changes how you use a model. A tool that reliably completes a five-step job is worth far more than one that does step one brilliantly and then asks what to do next.
For a professional, the payoff is delegation. You can hand Sonnet 5 a small end-to-end task, a research memo pulled from several sources, a spreadsheet cleaned and summarized, a draft email sequence, and get a finished first pass rather than a fragment. We cover concrete examples of that in [Claude use cases for professionals](/ai-workflows/claude-use-cases-professionals).
The safety notes worth reading
Anthropic reports that Sonnet 5 behaves better than Sonnet 4.6 on safety: it refuses malicious requests more reliably, resists prompt-injection hijack attempts better, and shows lower rates of hallucination and sycophancy. That last point matters for daily work. A model that flatters you and invents confident answers is a liability when you are relying on it for research.
Two honest caveats. Anthropic says Sonnet 5 still shows a somewhat higher rate of misaligned behavior than the more capable Opus 4.8 in its internal audits, so "safer than before" does not mean "the safest model available." And because Sonnet 5 is a bit stronger than its predecessor on cybersecurity tasks, Anthropic launched it with real-time cyber safeguards on by default. Neither of these should stop a normal professional from using it. Both are reasons to keep verifying important outputs yourself.
Where this fits
Zoom out and the story is about price, not raw capability. Every major lab now treats agentic behavior as the baseline expectation, so the competition has shifted to how cheaply and reliably a model can run without constant supervision. Sonnet 5 is Anthropic's move on that front: put near-flagship agentic quality at a mid-tier price and make it the default for a very large user base without asking.
For you, the takeaway is unglamorous and good. Your everyday Claude just got more capable at no extra cost, your automations may run more reliably, and if you were paying Opus prices for work that did not need them, you now have a cheaper option that is close enough for most of it. If you want to compare Claude against the other assistant most professionals also have open, we do that in [Claude vs Copilot for professionals](/ai-workflows/claude-vs-copilot-for-professionals).
Frequently Asked Questions
Do I have to do anything to get Claude Sonnet 5?
No. As of June 30, 2026, Sonnet 5 is the default model for Free and Pro plans, so if you open Claude you are most likely already using it. Max, Team, and Enterprise users have it available too, and developers can call it as claude-sonnet-5 through the Claude API and Claude Code.
Is Sonnet 5 actually cheaper for me, or just on paper?
The per-token price is lower than Opus and lower than the sticker on the previous Sonnet, but a new tokenizer can turn the same text into up to 1.35 times more tokens. In practice the lower rate and the higher token count may roughly offset for many users on the switch from Sonnet 4.6, so measure your real usage rather than assuming a large drop.
When should I still pay for Opus 4.8 instead?
Use Opus for the small share of work where the highest accuracy matters most: subtle judgment calls, high-stakes analysis, and deep research where one wrong inference is costly. For most daily drafting, research, and automation, Sonnet 5, sometimes at a higher effort setting, is enough and much cheaper.
Are the benchmark numbers independent?
No. The performance figures come from Anthropic and were relayed by outlets like TechCrunch. They are useful as a signal of direction, but no third-party lab has verified them yet, so weigh your own results on your own tasks above any published score.
Can I trust it to run tasks on its own?
It is built to finish multi-step tasks and to check its own work, and early testers report it follows through where older models stopped. Even so, keep a human reviewing anything sensitive or consequential. Cheaper autonomy is not a reason to skip the final check.
How long does the introductory pricing last?
Through August 31, 2026. After that, Sonnet 5 moves to $3 per million input tokens and $15 per million output tokens. If you have a large batch job planned, running it before the deadline costs less.
Sponsored Training
Informational tool analysis for working professionals, not legal, medical, or financial advice. AI tools do not replace your professional judgment.