Today, June 30th, Anthropic released Claude Sonnet 5 (codename Fennec). It’s not an Opus, it’s not the most powerful model in the lineup, and that’s exactly why it matters: it’s the middle model, the one balancing speed, intelligence and price. And for anyone building on top of these models —agents, tools, coding assistants— the middle model is usually the one you actually end up running in production.

Here’s an analysis using the official data (pulled from Anthropic’s docs and the System Card, not from headlines), focused on what matters if you’re going to build with it.

Where it sits in the lineup #

Anthropic keeps three tiers, and it helps to be clear about them:

• Opus 4.8 — the ceiling. The most capable for complex reasoning and long-horizon agentic work. The most expensive.
• Sonnet 5 — “the best combination of speed and intelligence”, per Anthropic. Fast, very capable, at a mid-range price.
• Haiku 4.5 — the fastest, with near-frontier intelligence, for high-volume tasks.

Sonnet 5 becomes the default model on the Free and Pro plans, and the default in Claude Code for Pro users. It replaces Sonnet 4.6, which was from February.

The price, which is half the story #

If you bill by tokens, this is the first thing you check. Sonnet 5 price per million tokens (input / output):

And there’s an important detail: launch pricing of $2 / $10 through August 31, 2026. After that it goes up to the $3 / $15 in the table. So for these two months you get a high-end model at a noticeably lower cost.

The underlying read is the usual one in this race: the “good enough” bar keeps rising while the price keeps dropping. What forced you to reach for Opus last year is now probably solved by Sonnet for considerably less.

The tech, on one screen #

• Context: 1 million tokens. Max output of 128k (up to 300k with a beta header).
• Adaptive thinking: the model decides for itself how much to “think” based on difficulty. The effort parameter defaults to high on the API and in Claude Code.
• Knowledge cutoff: January 2026.
• API ID: claude-sonnet-5 (a pinned snapshot, no date in the identifier).
• “Fast” latency, below Opus.

The numbers (official) #

These are figures from the Sonnet 5 System Card, compared with its predecessor Sonnet 4.6. I’m including them because there’s a lot of blog content quoting percentages that don’t match each other (I saw an “82.1%” on SWE-bench floating around that isn’t the official one):

• SWE-bench Verified (resolving real software issues): 85.2%.
• SWE-bench Pro (harder, multi-file variant): 63.2% vs 58.1% for Sonnet 4.6.
• Terminal-Bench 2.1 (terminal tasks): 80.4% vs 67.0%.
• OSWorld-Verified (computer use): 81.2% vs 78.5%.
• Humanity’s Last Exam with tools: 57.4% vs 46.8%.
• FrontierCode v1: 38.8% vs 15.1% — the most striking jump over 4.6.

And, to be fair, it doesn’t win at everything: on Terminal-Bench, GPT-5.5 with its own harness (Codex CLI) comes out ahead at 83.4%; and on one automation test, Gemini 3.5 Flash edges it by a few tenths. It’s not “the best at absolutely everything”, it’s very strong at agentic work and code, which is where it’s used most.

On safety, Anthropic reports fewer undesirable behaviors, fewer hallucinations and less sycophancy (agreeing with everything you say) than Sonnet 4.6, plus better resistance to prompt injection. In the offensive cybersecurity tests, it never managed to develop a working exploit.

When to pick Sonnet 5 #

How I’d decide, now that this is on the table:

• Agents and automations that make many calls → Sonnet 5. This is where price rules and where its agentic performance is already more than enough. Cost per task drops without giving up almost anything.
• Very complex reasoning, long chains, the hardest stuff → Opus 4.8 is still the ceiling. When failure is expensive and you need the maximum, pay the difference.
• High volume, latency-critical, simple tasks (classify, extract, route) → Haiku 4.5. You don’t need Sonnet for that.

Put another way: Opus when the error is expensive, Haiku when the volume is huge, and Sonnet 5 for almost everything else —which, if you’re honest, is most of the real work—.

What I take away #

The interesting thing about Sonnet 5 isn’t a specific benchmark, it’s the move: the middle model brings its performance closer to the top one and lowers the price. For anyone building, that’s what moves the needle —not the “most powerful” headline, but the “almost as good for a lot less”. The cost barrier to building something serious on AI keeps dropping on its own, month after month. And that, whatever you’re building, is good news.

(The figures and prices in this post come from Anthropic’s model documentation and the Claude Sonnet 5 System Card, dated June 30, 2026. Launch pricing applies through August 31, 2026.)