A cheaper flagship, with safety measures attached
Anthropic launched Claude Opus 5.5 on September 22, introducing the first model in its new Claude 5.5 family. The company says it performs at the level of its Claude Fable 5.1 model on most work while costing 40% less to run than Opus 5. That combination—not a claim of a sweeping leap over every rival—is the release’s clearest practical pitch.
Opus 5.5 is available through Claude and the Claude Platform, as well as Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic lists prices of $4 per million input tokens and $20 per million output tokens, below Opus 5’s $5 and $25. It also says the new model generates output more than 30% faster and uses fewer tokens on typical tasks, contributing to the claimed cost reduction.
You may also like
For companies and developers weighing AI model pricing, the key takeaway is to compare the cost of completing a task, not just the price of a single token. Anthropic’s figures are company estimates, and actual costs will depend on workload, model settings and how much human review is required.
Benchmarks point to coding and workplace tasks
Anthropic reported gains across several evaluations, including agentic coding, computer use and knowledge work. On Terminal-Bench 4.0, which tests multi-step tasks in a command-line environment, Opus 5.5 scored 66.4%, compared with 55.8% for Claude Fable 5.1 and 57.9% for OpenAI’s GPT-6 Astra in the company’s comparison. On CursorBench 4.0, a test built around ambiguous, multi-file coding tasks, it scored 57.8%, versus 51.8% for Fable 5.1.
The results are useful evidence, but they are not a universal ranking. Anthropic notes that some comparison scores come from other providers or evaluators, and that effort settings differ across models. For example, its Terminal-Bench table uses Opus 5.5 at “xhigh” effort and Astra at “high” effort. The company also cautions that benchmark margins may not predict differences in actual work; it says its own experience shows a narrower gap between Opus 5.5 and Fable 5.1 than some scores imply.
Anthropic highlighted a test in which Opus 5.5 completed a 680,000-line code migration in less than a day. That is an example from an early tester, not a guarantee that similar projects will finish as quickly elsewhere. Still, it signals the intended use: long-running AI coding agents that can take on large tasks with fewer steps and less costly back-and-forth.
Capability arrives alongside access controls
The launch also underscores how safety rules are becoming part of the product itself. Anthropic says external evaluators, including METR and Frontier Design, tested Opus 5.5 before release. The company reports that it achieved its best result to date on Anthropic’s automated behavioral audit, which tests model behavior across simulated scenarios. These are Anthropic’s descriptions of the evaluations; they do not establish that the system is risk-free.
Anthropic says Opus 5.5 has stronger defenses against prompt injection and is less likely than recent models to take hard-to-reverse actions or exceed its assigned boundaries. It also says the model has significant cybersecurity and biology capabilities, and is applying safeguards similar to those used for its more capable systems. Some cybersecurity and biology requests may be routed to other models, while vetted researchers can apply for expanded access through verification programs.
The company has also carried forward a measure it calls “preserved thinking,” intended to make it harder to extract model capabilities through large-scale misuse of API accounts. Such controls may matter to enterprise customers deploying agents in sensitive environments, though their effectiveness depends on implementation and ongoing testing.
The practical test is cost per reliable result
Opus 5.5’s release follows Anthropic CEO Dario Amodei’s call for the AI industry to pace frontier development. The new model therefore arrives with a tension at its center: Anthropic is expanding access to a more capable system while emphasizing safeguards, external testing and gated access for some high-risk work.
For buyers, the useful question is not simply whether Claude Opus 5.5 tops a benchmark. It is whether the model can reliably complete their own work at lower total cost—including retries, tool use, oversight and correction. Anthropic’s pricing and benchmark claims offer a reason to test it, but organizations should evaluate representative tasks under comparable settings before switching systems or granting agents greater autonomy.
That makes efficiency the more durable story than any single leaderboard position. If the lower cost per task holds up in independent, real-world use, Opus 5.5 could make capable coding and knowledge-work agents more practical. Until then, its launch is a strong company-reported case for testing—not proof that benchmarks alone settle which model is best.
Comments
0 comment