Claude Sonnet 5 Review: Near-Opus Quality at a Sonnet Price

Anthropic shipped Claude Sonnet 5 on June 30, and unlike the last couple of model launches everyone got excited about, you can actually use this one right now. No waitlist, no preview gate, no "coming weeks." It's the default model on Free and Pro, it's in Claude Code, and developers can call claude-sonnet-5 on the API today.

The pitch is simple and, for once, mostly true: near-Opus 4.8 quality at a Sonnet price. Sonnet 5 gets close to Opus 4.8 on agentic coding and computer use, slightly beats it on knowledge work, and costs a fraction of what Opus does to run. For a solo dev or a small SaaS team watching an API bill, that's the interesting part, not the benchmark bragging.

Here's the short version. If you're on Sonnet 4.6, switch. If you're paying for Opus 4.8 on work that doesn't truly need it, Sonnet 5 will cut your bill hard. If you're doing the genuinely hard stuff, gnarly reasoning, high-stakes accuracy, stay on Opus 4.8. Below is the real pricing, the actual monthly cost math, and the honest benchmark picture.

What's actually new in Sonnet 5?

Anthropic is calling this "the most agentic Sonnet model yet," and for once the marketing lines up with what testers are reporting. The headline behavioral change is autonomy: Sonnet 5 finishes multi-step tasks that earlier Sonnets would stall halfway through, and it checks its own output without being told to. A Zapier engineer described handing it a two-part job, update Salesforce tiers then send a launch announcement, and watching it run end to end where the old model would have stopped.

On paper, the gains over Sonnet 4.6 are real across reasoning, tool use, coding, and knowledge work. It ships with a 1M token context window and up to 128,000 output tokens. It also hallucinates less, is less sycophantic, and refuses malicious requests more consistently than Sonnet 4.6, which matters if you're running it as an agent with access to real systems.

One thing to keep honest: most of the benchmark numbers so far come from Anthropic itself. They're plausible and they match early tester reports, but I'd treat the exact figures as directional until independent evals land. That's true of every model launch, not just this one.

How much does Claude Sonnet 5 cost?

This is where it gets good for indie hackers. Sonnet 5 launched with introductory pricing of $2 per million input tokens and $10 per million output tokens, running through August 31, 2026. After that, it moves to $3 per million input and $15 per million output.

That later rate is exactly what Sonnet 4.6 charged. So the way to read this: during the intro window you get a real discount, and after it, you get a much better model at the same price you were already paying.

Here's how that plays out on a realistic workload. Take a solo SaaS running about 1,000 API calls a day at 1,500 input and 800 output tokens each, roughly 45 million input and 24 million output tokens a month.

At intro pricing, that workload runs about $330 a month on Sonnet 5 versus $825 on Opus 4.8. That's roughly 60% cheaper for work where Sonnet 5 is close enough in quality to not matter. Against Sonnet 4.6, you're saving a third during the intro window for a better model.

One honest caveat that most launch coverage is skipping: Sonnet 5 uses a new tokenizer that counts tokens higher than before, by anywhere from 1x to about 1.35x depending on your content. So the same text costs more tokens than it did on Sonnet 4.6. Anthropic says the intro pricing accounts for this, which is likely why the discount window exists at all. The practical takeaway: run your own workload for a day and check the real token counts before you trust the sticker math exactly. This is the kind of thing that quietly erodes a "60% cheaper" claim into "45% cheaper."

How does Sonnet 5 compare to Opus 4.8?

This is the decision most people actually care about, and the honest answer has nuance.

On agentic coding, Sonnet 5 scores meaningfully below Opus 4.8 but well above Sonnet 4.6. On knowledge work, it edges slightly ahead of Opus 4.8. On computer use and browser tasks, it lands close. Push Sonnet 5 to its maximum reasoning effort and it can roughly match Opus 4.8's mid-to-high setting on some benchmarks, but at that effort level it also gets expensive to run, sometimes more than Opus 4.8 at a comparable setting. So maxing out Sonnet 5 to chase Opus-level results is usually the wrong move. If you need Opus-level output, just use Opus.

Anthropic is refreshingly direct about this: Opus 4.8 is still the model of choice for higher accuracy on hard tasks, and Sonnet 5 is the lower-priced option that's much better than what was there before. That maps cleanly onto how you should actually pick. I broke the up-tier down in the Claude Opus 4.8 review, and the full tier logic is in when to use Sonnet vs Opus vs Haiku.

Against the wider field, Sonnet 5 undercuts Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on price while landing in a similar quality band for agentic work. It's still pricier than Gemini 3.5 Flash, which remains the budget option if raw cost is all you care about. The Gemini 3.5 Flash vs Sonnet 4.6 comparison covers that end of the market.

Who should switch to Sonnet 5?

Clear picks by situation.

On Sonnet 4.6 right now: switch. There's no real downside. It's cheaper during the intro window, matches your old price after, and it's a better model on every axis Anthropic measured. Update your model string to claude-sonnet-5 and move on.

Paying for Opus 4.8 on everyday work: switch most of it. If you've been running Opus as your default because it was the best, audit which tasks actually need that accuracy. Most day-to-day coding, drafting, summarizing, and tool-calling runs fine on Sonnet 5 at a fraction of the cost. Keep Opus for the hard 10%.

Doing genuinely hard, high-stakes work: stay on Opus 4.8. Complex multi-repo refactors, tricky root-cause debugging, anything where a wrong answer costs real money, Opus still has the edge on the hardest reasoning. Sonnet 5 narrows the gap, it doesn't close it.

Just using Claude in Claude Code or the chat app: you're already on it. Sonnet 5 is the new default, and you can switch models with /model if you want Opus for a specific session. If you're new to the workflow, the Claude Code guide for solo devs walks through it.

What else shipped alongside it?

Two things landed the same day, worth a line each. Anthropic also launched Claude Science, a research workbench for scientists, which is niche for most indie hackers. And Claude Code and Claude Design now sync both ways through a /design-sync command, so you can push design components between your repo and the Claude Design canvas. If you build UI, that one's worth a look on its own.

The bottom line

Claude Sonnet 5 is the rare launch that's both genuinely good and genuinely available on day one. For most indie hackers, it's the new default Claude model: close enough to Opus 4.8 on the work you do all day, cheaper than the Sonnet you were already using, and live everywhere right now.

Switch from Sonnet 4.6 without thinking twice. Move your non-critical Opus workloads over and watch the bill drop. Keep Opus 4.8 for the genuinely hard stuff. And test your real token usage during the intro window, because the tokenizer change means the savings are real but slightly smaller than the headline numbers suggest.

Running Sonnet 5 on something interesting? Tell me what you're building on Twitter @devtoolpicks.