Anthropic prices Claude Sonnet 5 for IPO-era adoption, not capability gaps
Anthropic released Claude Sonnet 5 with introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, rising to $3 and $15 after that. Standard Opus 4.8 pricing is $5 and $25. The capability story the company wants to tell is that Sonnet 5 closes most of the distance to its flagship. The more consequential story is what the price gap signals about where Anthropic expects its next revenue to come from.
The benchmark data Anthropic disclosed supports the first claim narrowly. On SWE-bench Pro, Sonnet 5 posts 63.2% against Sonnet 4.6's 58.1% and Opus 4.8's 69.2%. On Terminal-Bench 2.1, the figures are 80.4% for Sonnet 5, 67.0% for Sonnet 4.6, and 82.7% for Opus 4.8. On Humanity's Last Exam, Sonnet 5 reaches 43.2% without tools and 57.4% with tools, the latter essentially matching Opus 4.8's 57.9%. On OSWorld-Verified, Sonnet 5 reaches 81.2% versus 78.5% for Sonnet 4.6. On GDPval-AA v2, Sonnet 5 scores 1,618, surpassing Opus 4.8's 1,615 and far exceeding Sonnet 4.6's 1,395. The pattern is consistent: Sonnet 5 overlaps substantially with Opus 4.8 on agentic and knowledge-work evaluations while sitting roughly 60% cheaper per token at standard pricing.
What Anthropic does not test in the source material is the gap that determines production outcomes. Benchmarks measure capability under controlled conditions; the variable that decides whether enterprises move from pilot to deployment is whether the same model completes messy, multi-step workflows without human intervention. The customer testimonials Anthropic cites, from Cursor's Sualeh Asif and Zapier's Daniel Shepard, point at exactly that distinction. Asif said Sonnet 5 agents "stay on plan, follow our conventions, and ship clean multi-step changes." Shepard described a two-part Salesforce-plus-announcement task that "used to stall halfway" on prior models and now completes end to end. These are attributed claims, and they describe the reliability gap enterprises care about. They do not, on their own, establish that the same behavior holds at the scale Anthropic's IPO narrative requires.
The tokenizer change embedded in the release is the second pressure point. Sonnet 5 uses an updated tokenizer that maps the same input to roughly 1.0 to 1.35 times as many tokens depending on content type, the same kind of change Anthropic introduced with Opus 4.7. Anthropic says introductory pricing is calibrated to make the transition "roughly cost-neutral." That phrasing leaves actual bill impact workload-dependent. High-volume customers whose content sits at the upper end of that expansion range will see less of the headline discount than the per-token comparison suggests, and the source does not specify which content categories fall where in the range.
The safety disclosures in the source reinforce the cost-vs-capability tradeoff rather than altering it. Anthropic reports that Sonnet 5 has lower hallucination and sycophancy rates than Sonnet 4.6, refuses malicious requests more reliably, and is more resistant to prompt injection in agentic contexts. The company also says Sonnet 5 shows "somewhat higher rates of misaligned behavior" than Opus 4.8 and Claude Mythos Preview, its tightly restricted cybersecurity-focused model. On a Firefox 147 exploit development evaluation run with Mozilla, both Sonnet models scored 0.0% on working exploits, with Sonnet 5 at 13.2% partial success versus 4.6's 8.8%, well below Opus 4.8's 68.8% and Mythos 5's 88.4%. Cyber safeguards ship enabled by default on Sonnet 5, mirroring Opus 4.7 and 4.8 but less restrictive than those on Fable 5, the latest Mythos-class model. The practical effect is that the cheaper tier carries a slightly wider risk surface than the flagship, with platform-level controls absorbing the difference.
The release is best read as an IPO positioning decision wearing a model launch's clothing. The source cites reporting that Anthropic's revenue run rate crossed $47 billion by late May, up from $14 billion annualized in February, after a $65 billion Series H at a $965 billion post-money valuation co-led by Altimeter, Sequoia, and others. CNBC quoted PitchBook's Harrison Rolfes saying the figure that will "either validate or collapse the entire narrative the private markets have been pricing for three years" is gross margin, which no outside observer has seen. Gil Luria at D.A. Davidson, also cited in the source, said much of Anthropic's current usage is for "trials and experimentation and that may not sustain." Both observations frame the same adoption problem: converting experimental developer usage into production revenue is the conversion frontier lab valuations now depend on.
A near-Opus model at Sonnet prices is the structural answer to that problem. It gives enterprise finance teams a number they can approve at scale, and it gives Anthropic the high-volume recurring API revenue that anchors a software-as-a-service valuation. The risk is the opposite of the upside: the model that drives adoption is the same model that compresses the per-token margin story. Whether the S-1, when it lands, shows the Sonnet tier or the Opus tier carrying gross profit is the disclosure that will actually matter. The benchmarks in this release show capability overlap; the pricing shows the bet on volume. The source does not contain the data that resolves which bet pays off.