Anthropic's new mid tier coding model undercuts its flagship Opus on promo rates, yet at standard pricing it runs about 15% more per task, a tiering play more than a price revolution.
Anthropic is selling two stories about Claude Sonnet 5 at once, and which one you hear depends on which price tag you read. The promotional headline rate genuinely undercuts the company's own flagship on token cost. The standard list rate does not. At $2.29 per agentic coding task, Sonnet 5 runs roughly 15% more than Opus 4.8 before any discount. "Cheaper agents" is less a price revolution than a deliberate tiering bet.
Sonnet 5 is the latest mid-tier release in Anthropic's Claude family, the line positioned between the smaller Haiku and the top-shelf Opus. Anthropic markets it as its "most agentic" Sonnet model, meaning it is built to chain tool calls, browse the web, run code in a terminal, and execute multi-step coding workflows with limited human supervision. That agentic capability, not raw intelligence, is the central pitch. Sonnet 5 is meant to be the workhorse an engineering team leaves running overnight.
The pricing architecture is where the launch stops looking simple. Anthropic's introductory rate is $2 per million input tokens and $20 per million output tokens. Several aggregators also cite a $3/$15 promotional tier. At those rates, running an agent loop on Sonnet 5 is meaningfully cheaper than on Opus 4.8, which is the headline most coverage has run with. The arithmetic flips at standard list pricing. Independent benchmarker Artificial Analysis puts Sonnet 5's standard cost per task at $2.29, about 15% above Opus 4.8 once you normalize for the work each model actually finishes. The promotional window is a launch discount, not a permanent shift.
Capability scores track the same two-tier pattern. On the Artificial Analysis Intelligence Index, Sonnet 5 lands at number five, two to three points behind GPT-5.5 (xhigh) and Opus 4.8. On agentic coding specifically, SWE-bench Pro puts Sonnet 5 at 63.2%, against 69.2% for Opus 4.8 and 58.1% for the prior Sonnet 4.6. That is a clean step up from its predecessor and a clean gap below the flagship. The honest read is "noticeably better than last generation, clearly behind the best available."
Two benchmark categories are worth flagging. On Terminal-Bench v2.1, a test of long-horizon command-line work, Sonnet 5 jumps nine points over Sonnet 4.6. On Humanity's Last Exam, a wide-ranging expert-difficulty test, it picks up ten points. Both are the kind of gains a developer would actually feel in day-to-day agent use. On SciCode, a scientific-coding benchmark, Sonnet 5 gains seven points over its predecessor. These are real improvements, and they are why Anthropic can credibly call Sonnet 5 an agentic upgrade rather than a price play.
The caveat sits in the bench Anthropic did not lead with. On CritPt, a physics-reasoning test that increasingly separates general code assistants from scientific-tooling-grade models, Sonnet 5 scores 17%. That is fourteen points above Sonnet 4.6 and still behind GLM-5.2, the Claude Opus/Fable pair, and GPT-5.5. If your agent is meant to reason about physics or scientific code rather than ship features, the tiering logic holds. Sonnet 5 is the workhorse, Opus is still the premium anchor.
Independent industry signal is thin but pointed. TechCrunch's launch coverage carries one named on-record reaction: Zapier senior engineer Daniel Shepard, who described Sonnet 5 as "the model we'd been waiting for" on cost and tool-use grounds. Most other early reactions are analyst-blog aggregations, several of which echo each other on benchmark deltas, so the human signal beyond Zapier is limited. A broader benchmarks and pricing roundup tracks the same numbers without adding new human validation.
Two watch items follow from the design of the launch. First, the promotional window: how long the $2/$20 rate holds determines whether "Sonnet 5 is cheaper than Opus" is the frame for the next quarter or just the launch week. Second, the deferred benchmarks: independent coding and agentic tests from labs outside Anthropic and Artificial Analysis have not yet landed in force. If a third-party agentic eval puts Sonnet 5 closer to Opus than the AA gap suggests, the tiering story tightens. If it widens, the workhorse framing survives but the "near-flagship" label drops.
The honest summary: Claude Sonnet 5 is a real agentic step up from Sonnet 4.6, a clean step down from Opus 4.8 on intelligence, and a step sideways in price once the promotional window closes. "Cheaper agents" is true for as long as the discount lasts, and a marketing frame after that.