Anthropic Just Made Its Mid-Tier Model Almost as Smart as Its Flagship
Claude Sonnet 5 lands right behind the far pricier Opus 4.8 on the benchmarks that matter for real work — and at launch it costs less than half as much.
For the last couple of years, the pattern with AI models was easy to describe: if you wanted the smart one, you paid for the expensive one. Anthropic just blurred that line. Its newest release, Claude Sonnet 5, is the middle child of the lineup — and on several tests it comes within a hair of Opus 4.8, the flagship that costs more than twice as much to run.
That is the whole story in one sentence: near-top-tier intelligence, at a mid-tier price. If you have been quietly choosing which model to build on for coding or automation, the math just changed underneath you.
Why this release actually matters
Sonnet has always been Anthropic’s workhorse — the model most people and apps default to because it is fast and cheap enough to run all day. The problem was that the truly hard, multi-step jobs still got handed up to a bigger, costlier model. Sonnet 5 is Anthropic’s attempt to close that gap.
It is built to act, not just answer. Anthropic calls this its most agentic Sonnet yet, meaning it can make a plan and then carry it out using tools like a browser or a terminal on its own. Work that a few months ago only the largest models could reliably finish is now landing in the mid-tier. That shift is the quiet engine behind the whole “AI agents” wave — models that don’t just talk, but do.
The headline number
On agentic coding (Terminal-Bench 2.1), Sonnet 5 scores 80.4% — against Opus 4.8’s 82.7%. Nearly the flagship, at a fraction of the cost.
The receipts, side by side
Benchmarks are not gospel, but Anthropic’s own comparison chart tells a consistent story. Sonnet 5 sits clearly above its predecessor, Sonnet 4.6, and shadows Opus 4.8 closely. On a couple of measures it even edges ahead of the flagship.
| Benchmark | Sonnet 5 | Sonnet 4.6 | Opus 4.8 |
|---|---|---|---|
| Agentic coding (SWE-bench Pro) | 63.2% | 58.1% | 69.2% |
| Agentic coding (Terminal-Bench 2.1) | 80.4% | 67.0% | 82.7% |
| Reasoning w/ tools (Humanity’s Last Exam) | 57.4% | 46.8% | 57.9% |
| Computer use (OSWorld-Verified) | 81.2% | 78.5% | 83.4% |
| Knowledge work (GDPval-AA v2) | 1618 | 1395 | 1615 |
Look at the reasoning row. With tools switched on, Sonnet 5 hits 57.4% on Humanity’s Last Exam — basically tied with Opus 4.8’s 57.9%. And on the knowledge-work score, Sonnet 5 actually nudges past the flagship, 1618 to 1615. Those are the moments that make the price gap feel absurd.
The part your budget will care about
Here is where it gets loud. Per million tokens, Sonnet 5 launches at a special rate of $2 for input and $10 for output. Opus 4.8 runs $5 and $25. That is less than half the cost for output — the expensive half of most real workloads.
One catch worth noting
That $2/$10 rate is a launch promotion that reportedly runs through August 31. After that, standard pricing moves to $3 input and $15 output per million tokens — still well under Opus, but factor the change into any long-term budget.
There is also a dial worth understanding: effort level. You can tell Sonnet 5 to work harder on a task, and the cost chart shows exactly what happens — push the effort up and its results climb toward Opus territory, while the price per task stays far to the left of the flagship. In plain terms, you get to trade money for quality on a slider, instead of committing to the pricey model up front.
Smarter, and reportedly safer too
Raw capability isn’t the only thing Anthropic is claiming. Compared with Sonnet 4.6, the company says Sonnet 5 is better at refusing malicious requests and resisting prompt-injection attacks — the trick where hidden instructions try to hijack a model mid-task. It also reportedly hallucinates less and is less sycophantic, meaning it pushes back instead of just agreeing with you. In Anthropic’s own red-team testing, it says the model failed to produce a working exploit in every attempt — a clean 0%.
Treat those as vendor claims until independent testing weighs in, but the direction is the point: the cheaper model isn’t cutting corners on safety to get there.
Who should actually care
If you build or automate things — coding agents, research pipelines, customer workflows — this is the release to re-evaluate your defaults around. The people who benefit most are the ones running high volumes of agentic tasks, because that is exactly where the price-per-token gap compounds into real savings.
If you are a casual chat user, the upgrade is quieter but still free: Anthropic says Sonnet 5 is available across every plan from launch day, and it is the default model on both the Free and Pro tiers, plus Claude Code and the API. You are likely already using it without changing a thing.
TL;DR
- Sonnet 5 gets close to Opus 4.8 on agentic coding, reasoning and computer use.
- Launch pricing is $2/$10 per 1M tokens (vs Opus at $5/$25), rising to $3/$15 after Aug 31.
- An effort dial lets you trade cost for quality per task.
- Available on all plans now; default on Free, Pro, Claude Code and the API.
The cheap model is the story now.
If you’ve been defaulting to a flagship for coding or automation, run one real task on Sonnet 5 first and watch what it costs. The gap between “good enough” and “top tier” just got a lot cheaper to cross.
