Anthropic Launches Opus 5.5, Priced 60% Below Flagship Model
nashnova research
Anthropic released Claude Opus 5.5 on September 22 — matching its top-tier Fable 5.1 on performance at roughly 40% of the cost. The aggressive pricing, weeks before a planned IPO, is a direct bet that lower unit prices will drive enough volume to lift total revenue.
How much cheaper, and where?
Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, an overall ~40% cut from Opus 5.
The discount varies by use case: input and output each drop 20%, while cache reads drop 60%. This means → customers running heavy agent workloads — where AI autonomously executes multi-step tasks — see savings closest to the 60% ceiling.
Early adopters confirm the gains: cloud-storage firm Box says Opus 5.5 uses roughly one-third the tokens of Opus 5 for equivalent tasks; trading firm Optiver reports agent-coding costs down 40–50%.
How does it stack up against OpenAI?
On Terminal-Bench 4.0, Opus 5.5 default mode scores 66.4%, beating OpenAI GPT-6 Astra at 57.9% and Anthropic's own Fable 5.1 at 55.8%.
On CursorBench it leads OpenAI Sol by 11 points at roughly one-third the cost; on FrontierCode it tops Astra's best result at about one-fifth the per-task price.
Pricing is pinned exactly to OpenAI Sol's promotional rate ($4/$20), with cache reads cut in half: Opus 5.5 charges $0.20 per million, versus Sol's $0.40. In plain terms = performance at parity or ahead, headline price matched to the rival's floor, caching half the cost.
A 73% price drop in one year — who's forcing the pace?
Flagship pricing has fallen from $15 input / $75 output to $4 / $20 in just over a year — a cumulative ~73% decline.
Open-source models are the main pressure: Vercel data shows open-source models — largely from China, including DeepSeek and Moonshot AI — accounted for 56% of token traffic on its AI gateway in August, up from just 7% in December 2024. Yet they represented only 14% of estimated spend.
This means → open-source captures volume but barely generates revenue. Anthropic's response is deliberate self-cannibalization: the half-price Opus 5 surged to 22.5% of Vercel gateway spend in August, displacing the company's own premium Fable 5. Opus 5.5 extends the same logic — flagship performance at a 60% discount, keeping revenue in-house.
What's the IPO story behind the pricing?
Anthropic is preparing a fall IPO at a reported valuation of roughly $96.5 billion. Annualized revenue topped $65 billion by late July, up from $9 billion at end-2025; investors expect full-year revenue between $100 billion and $120 billion.
The company confidentially filed its IPO draft with the SEC on June 1. This reflects a timing play: Opus 5.5 needs to prove the "cut price → grow volume → grow total revenue" thesis before the roadshow begins.
In plain terms = this is a Jevons bet — the cheaper the unit, the more units sold, and total revenue rises. Whether that math holds will determine if the IPO valuation stands.
What about safety and the broader ecosystem?
Opus 5.5 is 85% less likely to bypass its safety boundaries than Opus 5 or Mythos 5.1, following alignment testing, external pre-release reviews, and safeguards in high-risk domains such as cybersecurity and biology.
Subscriber usage caps will rise 20% within the five-hour window, with a new banked-reset mechanism added.
The model is live on Amazon Web Services, Google Cloud, and Microsoft Azure; Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks.
市场有风险，内容仅供研究参考，不构成投资建议。