Anthropic / Wikimedia Commons (Public domain)
The first model in the Claude 5.5 family cuts per-token pricing and reportedly resists safety workarounds far better than earlier versions
Anthropic released Claude Opus 5.5 on September 22, 2026. It is the first model in the new Claude 5.5 family, and the pitch is simple: more capability for less money.
The company claims Opus 5.5 performs on par with Fable 5.1, its higher-tier model, while costing noticeably less to run than Opus 5.
Cheaper tokens, cheaper cache, cheaper everything #
Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens. That works out to approximately 20% less than what Opus 5 charged.
The bigger cut came on cache reads. Those fell 60% to $0.20 per million tokens.
Combine the lower list prices with the cache discount, and typical workloads on Opus 5.5 come out approximately 40% cheaper overall than on its predecessor.
The model is available through the Claude API, Amazon Web Services, Google Cloud, and Microsoft.
Early usage data suggests the discount is landing. Token volume rose 3.6 times following the release, a sign that lower per-task costs are pulling in more customer activity rather than simply shrinking the bill.
The safety numbers #
Anthropic paired the launch with external safety evaluations from organizations including Frontier Design and METR. Those assessments point to improved alignment and containment compared with earlier models.
The headline safety figure: Opus 5.5 is reportedly about 85% less likely to attempt to bypass predefined safety protocols than prior versions.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Opus 5.5 is built for complex agentic workflows, with better performance on coding migrations and multi-step jobs. An agent is software that does not just answer a question but carries out a sequence of actions, like moving a codebase from one framework to another.
How Anthropic builds Claude with Claude #
In a conversation published by Latent Space, Katelyn Lesse, Anthropic’s head of engineering, and Angela Jiang, its head of product, walked through how the company ships its models. A recurring theme: Anthropic uses Claude to build Claude products.
The engineers also described a design choice with clear safety implications. Anthropic keeps the systems that control agents separate from the systems that execute code.
Latent Space also noted that Anthropic credits its performance to two core loops, which it calls Improvement and Investment. The publication described the previous flagship as having held the position of generally accepted top large language model in the world for a month.
The financial backdrop #
The launch lands alongside Anthropic’s S-1 filings, the paperwork a company submits ahead of going public.
Anthropic reported $16.3 billion in revenue for the first half of 2026. That is an increase of approximately 1,300% year over year.
In the second quarter of 2026, the company posted an operating profit of $974 million and an adjusted operating margin of 19%.
Latent Space characterized the swing as a turnaround from a -126% loss to a 19% GAAP profit over the past year, describing Anthropic as having some of the highest revenue and research efficiency in history.
What this means #
The 3.6 times jump in token volume suggests Anthropic is betting that cheaper tasks unlock new use cases, and that volume more than makes up for lower unit prices.
For the cloud providers distributing the model, more token volume means more workloads flowing through their platforms. Amazon Web Services, Google Cloud, and Microsoft all carry Opus 5.5, so the growth is not concentrated in a single partner. As enterprises push models into agentic roles, with access to codebases and internal systems, a model that is reportedly about 85% less likely to slip its constraints becomes easier to justify to a risk committee. External evaluations from groups like METR give that claim more weight than a self-reported score would.
Anthropic is expected to follow Opus 5.5 with Sonnet 5.5 and Haiku 5.5, its mid-tier and lightweight models. With the S-1 now public, investors will be tracking whether the usage surge after this launch shows up in the next round of reported revenue and margins.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our