Anthropic launches Claude Sonnet 5 as a cheaper agent model
TL;DR
- Sonnet 5 scores 63.2% on agentic coding benchmarks vs. Opus 4.8's 69.2%, narrowing but not closing the gap with Anthropic's top model.
- A new tokenizer can inflate token counts by up to 35%, meaning real-world cost savings will run smaller than the $2/$10 headline rates imply.
- The introductory rate expires September 1, when input cost rises from $2 to $3 per million tokens, making the launch window a land-and-expand play.
Anthropic dropped a new mid-tier model today, and the interesting part is the price, not the leaderboard line. Claude Sonnet 5 is live as of June 30, with introductory pricing of $2 per million input tokens and $10 per million output through August 31, after which the standard rate moves to $3 and $15. The company is pitching it as 'the most agentic Sonnet model yet,' and is making it the default for Free and Pro users rather than gating it behind the top tier.
The framing matters, because Anthropic itself is not claiming Sonnet 5 beats its own flagship. The company says performance is 'close to that of Opus 4.8, but at lower prices,' tested against Sonnet 4.6 and Opus 4.8 on the BrowseComp agentic search evaluation and OSWorld-Verified computer use. According to TechCrunch, on one agentic coding measure Sonnet 5 scores 63.2% against Opus 4.8's 69.2% and Sonnet 4.6's 58.1%, closer to the top tier than the previous generation at a fraction of the run cost.
Why a small-business developer should care: a lot of useful agent work, things like desktop automation, multi-step research, and code-edit-test loops, burns tokens. If you can keep most of the accuracy of the expensive model while paying roughly a fifth as much on the input side, the economics of running long-horizon agents in production change. TechCrunch quoted a Zapier senior engineer saying Sonnet 5 'finished end to end' on tasks that 'used to stall halfway,' which is the use case Anthropic is clearly aiming at.
The honest caveats are worth keeping in view. Anthropic explicitly notes Sonnet 5 has 'substantially weaker cybersecurity capabilities' than Opus 4.8, and the company's framing of 'close to' Opus accuracy is its own claim, not an independent eval. The announcement does not publish detailed BrowseComp or OSWorld numbers in the body, and the hallucination improvement is described qualitatively rather than quantified. Take the specifics as reported, not settled.
What is worth watching is whether the default-model swap on Free and Pro changes how people build with Claude at the hobby end, and whether the two-month introductory window pulls agent workloads off Opus 4.8 in production before the standard rate kicks in on September 1.
What others are reporting
-
TechCrunch Read →
Frames the launch as a competitive cost race in agentic AI, benchmarking Sonnet 5 explicitly against GPT-5.6 Sol and Gemini and arguing cost-per-task is now the primary AI differentiator.
It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.
-
Yahoo Finance Read →
Situates the launch inside broader enterprise retrenchment: Meta, Amazon, and Uber cutting AI spend after costs spiraled. Also ties the release to Anthropic's regulatory friction and pending IPO.
Testers described how [Sonnet 5] finishes complex tasks where previous Sonnets would stop short
-
The Next Web Read →
Adds a key caveat: Sonnet 5's new tokenizer can increase token counts by up to 35%, meaning headline per-token savings are materially softer in practice than the $2/$10 rate suggests.
Sonnet 5 offers near-flagship performance at a mid-tier price. It lands close to Opus 4.8...and costs far less than Opus to run.
-
The New Stack Read →
Focuses on the benchmark delta between Sonnet 5 and Opus 4.8 and frames the August 31 pricing cliff as a deliberate go-to-market tactic rather than a permanent rate.
Shared on Bluesky by 4 AI experts
Originally reported by anthropic.com
Read the original article →Original headline: Anthropic Launches Claude Sonnet 5 — Approaches Opus 4.8 Performance at $2/$10 per Million Tokens, Marketed as 'Most Agentic Sonnet Yet'