Anthropic launches Claude Opus 5.5 with 40% lower cost and coding gains

Anthropic announced on Tuesday (22) Claude Opus 5.5, the new top-tier model in the Claude lineup that, according to the company, delivers performance equivalent to Claude Fable 5.1 (its most advanced model until then) while costing 40% less to run typical workloads than its predecessor, Claude Opus 5. The launch was reported by Reuters and reproduced by Economic Times.

For those already running Anthropic's API in production, the announcement hits the bill directly: Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, a 20% reduction from Opus 5 pricing. Combined with the claim of 40% savings in real workloads (which depends on how the model processes each type of task, not just the per-token price), Anthropic is positioning Opus 5.5 as an option for teams that currently avoid Claude's most expensive tier due to operational cost.

Code benchmark: Opus 5.5 versus GPT-5.6 Sol

The data point that matters most to those building software is the direct comparison with the competition: Anthropic says Opus 5.5 outperformed OpenAI's GPT-5.6 Sol on a software development benchmark, while costing about a third of the price to run the same task. The company did not publicly detail which specific benchmark was used nor publish the full methodology in this communication, which is worth flagging: AI lab marketing numbers tend to use the benchmark that most favors the model in question, and it's worth waiting for independent evaluations before switching vendors based on this data alone.

Even so, Anthropic explicitly targets code generation and review as Opus's competitive differentiator, in a market where GPT, Gemini, and Claude compete for preference among tools such as coding agents, IDE copilots, and CI pipelines that call an LLM for automated review.

Safety: tested by Frontier Design and METR

The launch comes at a moment of public tension over the pace of AI releases. Earlier this month, Anthropic CEO Dario Amodei asked the global AI community to slow down the release of new capabilities to allow more time for safety work. The company presents Opus 5.5 as a partial response to that very demand: according to Anthropic, the model underwent external testing by two independent AI safety research groups, Frontier Design and METR, before launch, and received safeguards previously reserved for the company's most capable systems.

In internal testing, Anthropic states that Opus 5.5 was about 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass containment limits in an evaluation dedicated to that behavior. For teams running autonomous agents with access to tools, files, or code execution, this kind of "sandbox escape attempt" metric is more practically relevant than general-knowledge benchmarks, because it is precisely the risk scenario when an AI agent has write permission in a production environment.

What changes for those running AI in production in Brazil

Opus 5.5 is already available on platforms such as Amazon Web Services, Google Cloud, and Microsoft Azure, giving Brazilian teams that already use Claude on these clouds a way to test the model within the same contract and billing infrastructure they already use, without needing to open a direct relationship with Anthropic. This lowers the adoption barrier for teams that already have contracts and billing set up on these clouds (relevant in Brazil, where dollar-denominated billing and service import tax weigh on the AI budget of any company that doesn't bill directly with Anthropic).

The reduction in per-token price, combined with the claim of lower cost per completed task, is the kind of announcement that matters to anyone managing an AI budget in reais: if the savings ratio actually holds up in production workloads (and not just in Anthropic's controlled benchmark), teams that currently limit Opus usage to critical tasks because of cost gain room to expand its use to more workflows, such as pull request review, test generation, or bug triage.

What's still to come

Anthropic said that Claude Sonnet 5.5 and Claude Haiku 5.5, the mid-tier and lightweight versions of the same generation, are expected to launch in the coming weeks, bringing much of the performance, speed, and safety gains of Opus 5.5. This matters because Sonnet tends to be the option used by teams seeking a balance between cost and capability in their API integrations, so those relying on it for high call volume still need to wait. Until these models come out, the decision to migrate entire workload stacks remains incomplete: you can test Opus 5.5 today, but those who depend on Sonnet for high call volume need to wait.

The company also said it is expanding programs that give previously vetted cybersecurity and life sciences researchers broader access to the model, which suggests Anthropic continues to treat certain capabilities of Opus 5.5 as sensitive enough to restrict by default. Neither Anthropic nor Reuters's coverage detailed exact timelines or eligibility criteria for these programs as of this writing.