Claude Sonnet 5

The next generation of the Sonnet family, offering what Anthropic calls the best combination of speed and intelligence. It is a capability upgrade over Sonnet 4.6 at a lower price and a drop-in replacement for it, with the largest gains in coding and agentic tasks. It is the model to reach for when you need more than Sonnet 4.6 without moving to an Opus-class price, and it supports the same low-to-max effort ladder.

Anthropic built Sonnet 5 to be the most agentic Sonnet it has shipped: a model that makes plans, drives browsers and terminals, and runs on its own at a level that a few months earlier needed a larger and dearer model. The lab tells the history deliberately. The agentic era began with Sonnet-class models, then the clearest gains moved up into the Opus class; Sonnet 5 is the release that pulls the line back down. It became the default model on Free and Pro the day it shipped and is available on every other plan.

Anthropic presents the results as cost-performance curves across the effort ladder rather than as single peak scores, which is the honest way to read a model whose thinking budget the caller sets. On the agentic search evaluation BrowseComp and the computer-use evaluation OSWorld-Verified, the lab reports Sonnet 5 as a strict improvement over Sonnet 4.6 at every rung, covering a wider range of cost-performance options than Opus 4.8 does, with the largest efficiency gain at medium effort and higher-effort runs matching Opus 4.8 on some tasks.

The table Anthropic published puts Sonnet 5 at 63.2% on SWE-bench Pro against 58.1% for Sonnet 4.6; 80.4% on Terminal-Bench 2.1 against 67.0%, within about two points of Opus 4.8; 81.2% on OSWorld-Verified against 78.5%; 43.2% on Humanity's Last Exam without tools and 57.4% with them, against 34.6% and 46.8%; and 1618 on GDPval-AA v2, just past Opus 4.8's 1615. Early-access testers described the change as follow-through rather than raw score: finishing multi-step jobs where earlier Sonnet models stopped halfway, and checking their own output without being asked to.

One practical catch is stated plainly in the post. Sonnet 5 uses an updated tokenizer, so the same input can map to roughly 1.0 to 1.35 times as many tokens depending on the kind of content, which is a real cost difference on top of the per-token rate. Anthropic also raised rate limits across its own surfaces to absorb the heavier token use of the higher effort rungs.

The lab is specific about where the model stops. It did not deliberately train Sonnet 5 on cybersecurity, and on the Firefox exploit-development evaluation the model never produced a working exploit, though its partial-success rate sits slightly above Sonnet 4.6's, which Anthropic attributes to general capability rather than targeted training. It shipped with cyber safeguards enabled by default, less strict than Fable 5's, and Anthropic recommends Opus 4.8 for cybersecurity work that needs reduced guardrails. On its automated behavioural audit Sonnet 5 is safer overall than Sonnet 4.6 but shows more misaligned behaviour than Opus 4.8 and Claude Mythos Preview.