Sonnet 5: the middleweight closes the gap

01 What shipped

The launch itself was almost drowned out. Sonnet 5 arrived June 30 — the same day Washington lifted the export controls on Fable 5 and Mythos 5, the story I covered yesterday. By July 1 it was the default model for every Free and Pro user on claude.ai, the default in Claude Code, and live on Bedrock, Google Cloud, and Foundry. Anthropic calls it its “most agentic Sonnet model yet,” which is marketing, and then backs it with numbers, which isn’t.

Three things matter operationally. It ships with a 1M-token context window by default — no beta header, no long-context surcharge. It becomes the model most people get without choosing anything. And it lands at an introductory price with a fuse: $2 per million input tokens and $10 out through August 31, then $3/$15 — nominally identical to Sonnet 4.6.

Read the chart honestly and the story is “close, not equal.” Six points on SWE-bench Pro is a real gap — on hard, multi-file agentic coding you will still feel the flagship’s edge. Two points on OSWorld is a coin toss. On agentic search Sonnet 5 posts 84.7% on BrowseComp (no published Opus figure to pair it with), and on GDPval knowledge work the two models are, for any practical purpose, the same model.

02 The terminal upset

The number that made me sit up isn’t any of the near-misses. It’s Terminal-Bench 2.1: 80.4% against the flagship’s 74.6%. The middleweight doesn’t approach Opus 4.8 at driving a shell — it beats it, by nearly six points, at a third of the standard price.

That specific win matters because the terminal is where agentic work actually happens: package managers, test runners, git, deploys, the unglamorous plumbing that fills an agent’s day. A model that is cheaper andbetter at the highest-volume lane of agent traffic isn’t a budget option. It’s the correct default — which is presumably why Anthropic made it one.

The pattern rhymes with what I wrote when sizing up Fable 5: capability tiers aren’t a ladder, they’re a toolbox. Each generation, one “smaller” model quietly takes a lane from the model above it. Last time it was long-context retrieval. This time it’s the shell.

So the “same price as 4.6” framing survives only until September. After the promo, an unchanged workload costs roughly a third more in real money for — to be fair — a meaningfully better model. That may be a trade worth making. It is not the trade the sticker advertises, and if your finance team budgets in tokens, the budget line will move even though the price list didn’t.

03 My read

The obvious reading is competitive: Anthropic, days from the most anticipated IPO in tech, ships a model that undercuts its own flagship for most work and prices it to make switching irresistible for exactly two months. That reading is probably right and slightly boring.

The more useful reading is architectural. With Fable 5 restored at the frontier, Opus 4.8 holding the hard-problem tier, and Sonnet 5 now owning the default lane, the “which model?” question has stopped having one answer. It has a routing table — and the teams that treat it that way, with explicit lanes, explicit escalation, and effective-cost math, will quietly pay less for better work than the teams still picking a favorite model like a football club.