Claude Sonnet 5 Is Live: 90% of Opus Performance at a Fraction of the Cost

It Plans, Browses, and Executes — Without You Watching Every Step

The biggest shift with Sonnet 5 isn't raw intelligence — it's autonomy. Previous Sonnet releases were great at answering questions and generating drafts, but they needed you to hold their hand through multi-step work. Sonnet 5 can browse websites, run terminal commands, and chain tasks together into a real workflow without constant check-ins. That's the kind of behavior that used to require reaching for Opus or a purpose-built agent framework.

For builders, this means you can hand it a goal — 'research these five competitors, summarize their pricing pages, and draft a comparison table' — and actually walk away. For creators, it means automating research loops, outline generation, and publishing prep without stitching together five different tools. The model handles the sequencing. You handle the direction.

This level of agentic behavior used to be the reason people paid for the biggest, most expensive models. Sonnet 5 brings it down to a price tier that makes sense for everyday use cases, not just high-stakes enterprise deployments.

63% vs 69%: The Performance Gap That Actually Matters

On agentic coding benchmarks, Sonnet 5 scores 63.2%. Opus 4.8 scores 69.2%. The previous Sonnet 4.6 scored 58.1%. So Sonnet 5 isn't just marginally better than its predecessor — it's a meaningful jump forward, and it's sitting much closer to Opus than to the model it replaced.

For context on what that 6-point gap means in practice: for most coding tasks — writing functions, debugging, generating boilerplate, scaffolding a new feature — you will not feel the difference. The gap shows up on the hardest, most complex multi-file reasoning problems. If you're building a SaaS product or automating internal tooling, Sonnet 5 will handle the vast majority of your actual workload without hitting that ceiling.

The honest framing is this: Opus is still the right call when you're working on something genuinely hard and the cost of being wrong is high. For everything else — which is most things — Sonnet 5 gives you output quality that would have seemed like a flagship model six months ago.

$2 Input, $10 Output: The Pricing Math That Makes This a No-Brainer

Through August 31st, Sonnet 5 is priced at $2 per million input tokens and $10 per million output tokens. That launch pricing beats both GPT-5.5 and Gemini 3.1 Pro on cost while delivering near-Opus results. If you're running any kind of volume — automated pipelines, batch content generation, API-driven features in a product — this is a significant line item difference.

The way to think about it: you're paying a fraction of Opus pricing to get roughly 90% of Opus performance. For most builders and creators, that math closes the case. You'd need a very specific reason to reach for the more expensive option when the gap is this narrow and the price difference is this wide.

If you've been hesitant to build API-based workflows because the token costs felt unpredictable, this pricing tier makes it much easier to model your costs upfront. Run the numbers on your actual use case — you'll likely find Sonnet 5 fits comfortably within a budget that would have been tight with Opus.

It's Already Your Default — No Setup Required

There's no waitlist. There's no upgrade prompt. If you have a Claude Free or Pro account, you're already using Sonnet 5. Anthropic made it the default model across both tiers at launch, which is a statement: they built this to be the everyday workhorse, not a premium add-on.

For builders, that means any projects or workflows you've already set up in Claude are running on Sonnet 5 automatically. If you're using the API, check your model parameter — switching to claude-sonnet-5 is a one-line change and your existing prompts should carry over cleanly.

For creators using Claude.ai directly, just log in and start prompting. The interface is the same. The model is better. The biggest thing you can do right now is take whatever task you've been doing manually or with a weaker model and try handing it to Sonnet 5 with a more ambitious prompt than you'd normally write. It can handle more than you expect.

Safer by Design — Especially in Agentic Workflows

Autonomy without guardrails is a liability, not a feature. Anthropic built Sonnet 5 to be better at refusing malicious requests and resisting prompt injection attacks compared to other agentic models. That matters more than it might sound when you're giving a model access to a browser, a terminal, or an API.

Prompt injection is a real attack surface for agentic systems — it's when a model gets fed instructions embedded in external content (a webpage, a document, an email) designed to hijack its behavior. Sonnet 5's improved resistance to this is a practical safety property, not just a marketing checkbox. If you're building any kind of automated pipeline that touches external data, this is the kind of hardening that makes deployment feel less risky.

For creators, the upside is simpler: the model is better at staying on task and not being manipulated by weird inputs. For builders shipping products where users might try to jailbreak your AI features, Sonnet 5's safety improvements are doing real work in the background. It's a model you can deploy with more confidence than its predecessors.