Claude Sonnet 5 Review: The Cheap Model That's Suddenly Pun
ching Up
So Anthropic just dropped Claude Sonnet 5, and honestly, this one's got people talking for a reason that isn't just "new model, bigger numbers." It's priced way below Opus 4.8 while apparently getting close to it on a bunch of agentic tasks. If you've been paying for Opus just to get work done reliably, this release is worth your attention.
What Actually Happened
Anthropic announced Claude Sonnet 5 on its official blog, and it's now the default model for Free and Pro plans, with availability across Max, Team, and Enterprise too. It's also live in Claude Code and on the Claude Platform for developers pulling it through the API.
Here's the part that matters for your wallet: introductory pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that it moves to the standard $3/$15 per million tokens. That undercuts both Opus 4.8 and, according to reporting from TechCrunch, OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro too.
Okay But Is It Actually Better?
To be fair, "better" depends on what you're doing with it. Anthropic says Sonnet 5 is a real jump over its predecessor, Sonnet 4.6, specifically on agentic work: reasoning, tool use, coding, and knowledge tasks. TechCrunch reported one benchmark where Sonnet 5 scored 63.2% on agentic coding versus Sonnet 4.6's 58.1%, with Opus 4.8 still ahead at 69.2%.
That gap matters less than it sounds. On a knowledge-work benchmark, Sonnet 5 reportedly edges out Opus 4.8 slightly, which is a weird thing to see from the "cheap" model. If you've used earlier Sonnet versions for coding, you'll notice this one behaves less like a chatbot that occasionally writes code and more like something that stays on task without you babysitting every step.
The Agentic Angle Is the Real Story
Here's the thing, Anthropic isn't just claiming this model runs autonomously, it's leaning hard into that framing. The company says Sonnet 5 can make plans, use tools like browsers and terminals, and keep working through multi-step jobs without constant hand-holding, at a level that previously needed a bigger, pricier model.
According to Anthropic's writeup, early testers reported that Sonnet 5 finishes complex tasks other models would stall on, and checks its own output without being told to. A senior engineer at Zapier described giving it a two-part job (updating Salesforce tiers, then sending an announcement email) and having it complete the whole thing without stopping halfway, something that reportedly used to be a common failure point.
Lovable's co-founder also commented that the model handles unsafe requests cleanly and consistently, which lines up with Anthropic's own safety numbers.
On the Safety Side
Anthropic's system card and blog post both note that Sonnet 5 shows a lower rate of undesirable behaviors than Sonnet 4.6, including less hallucination and less sycophantic behavior, and it's supposedly more resistant to prompt-injection style hijacks. It also has real-time cybersecurity safeguards for the first time at the Sonnet tier, meaning certain risky requests get refused outright.
To be clear though, Anthropic itself says Sonnet 5 isn't in the same league as Opus 4.8 or the newer Mythos-tier models when it comes to alignment robustness. It's safer than its own predecessor, not the safest thing Anthropic makes.
Why This Matters If You're Actually Using AI Day to Day
If you're running agent-style workflows, coding assistants, or anything that chains multiple steps together, the price-to-capability ratio here is the headline. Getting most of Opus's practical usefulness at roughly a third of the cost changes the math for a lot of people who were rationing Opus usage or avoiding it because of cost.
For casual chat use, you probably won't feel a dramatic difference day to day, but if you lean on Claude for coding, research chains, or automation, this is the upgrade that actually shows up in your workflow instead of just in a press release.
What's Not Confirmed Yet
Worth flagging: some of the more specific benchmark comparisons and tester quotes are coming from Anthropic's own announcement and press coverage, not independent third-party testing that's had time to fully verify things in the wild. Real-world performance across different coding stacks and use cases will take a few more weeks to shake out properly.
One more thing to note: the new tokenizer Anthropic shipped with Sonnet 5 reportedly produces about 30% more tokens for the same text compared to Sonnet 4.6. Per-token pricing didn't change, but that means your actual cost per request could shift depending on what you're doing, so it's not a pure like-for-like comparison on cost.
Bottom Line
Claude Sonnet 5 looks like Anthropic's answer to "why would I pay for the expensive model" for a lot of everyday agentic and coding use cases. It's not replacing Opus for the hardest problems, but for most day-to-day work, it might be the smarter default.
Have you tried it yet? Curious if it's holding up the same way in your actual workflow.
