When companies purchase large models, they usually first look at the quote per million token. However, the actual cost of completing a task also depends on how many steps the model takes, how many times tools are invoked, and whether there is any need for repeated rework. On September 28th, Anthropic launched Claude Sonnet 5.5, making these costs more transparent: officially, it claims that its generation speed is over 30% faster than that of Sonnet 5. Although the listed price has not been reduced, in the company's tests, the cost of completing similar tasks can be as much as about 30% lower. This is not a promise to reduce each customer's bill by 30%, but rather a result of the new model being more efficient in output and completion under certain tasks and test conditions.

Sonnet 5.5 is already available on the Claude platform and major cloud channels. Developers can use `claude-sonnet-5-5`. Anthropic It is positioned under the newly released Opus 5.5 for tasks with clear daily scopes, such as fixing program errors, creating documents, and making presentations and tables. This division of labor is noteworthy: in the past, manufacturers liked to describe each update as "the strongest model setting a new record," but this time it's more like treating the model as part of a company's human resource allocation—complex, open tasks with low fault tolerance are still assigned to the more capable and more expensive team; for tasks with clear boundaries and large volumes, it depends on whether Sonnet can complete them stably and at a lower cost.

The price hasn't gone down, so why is it said to be cheaper?

The official price list shows that Sonnet is 5.5 per million inputs, token is 2 US dollars, and the output cost is 10 US dollars; the cost for cached reads is 0.20 US dollars, which is the same as the previous generation Sonnet at 5. If one simply compares the new and old prices side by side, there appears to be no “price reduction.” The so-called cost improvement mentioned in Anthropic comes from another aspect: the token used, tool calls, and time required to complete the same task. It cites feedback from early customers; for example, Slack states that the output of their internal robots has increased by approximately 14%, and Zendesk claims that test ticket processing has accelerated by about 20%. These are results specific to certain customers and workflows and should not be generalized as average market gains.

For corporate purchasers, this difference is quite real. A model that costs a few cents less to produce the first version of an answer but requires repeated manual revisions may not necessarily be cheaper; a model that is slightly more expensive per response, if it can reduce errors and duplicate communications, might actually be more cost-effective for individual tasks. Anthropic also mentioned in the announcement that the default “thinking effort” of Sonnet at 5.5 is set to Medium in Claude Code and applications, and to High in Claude Platform. If the same model is configured differently, the delays and costs will also vary. A truly fair comparison should take into account the task success rate, manual review time, interruption rates for long tasks, as well as the final quality of output, rather than just looking at the price list.

Performance data also needs to be put back into the testing scenario. Anthropic reports that Sonnet 5.5 achieved 70.6% in the Terminal-Bench 4.0 proxy programming assessment, while the previous generation Sonnet 5 scored 10.3%. These results are specific to that assessment and its setup, and cannot be directly generalized to mean that "ordinary programmers' efficiency has increased sevenfold." In another assessment, Sonnet 5.5's high-effort setting was close to Opus 5.5's, but Anthropic itself acknowledges that for complex, open tasks that require continuous judgment, Opus is still more effective. It is particularly noteworthy to note its footnote: in some tests, Sonnet 5.5 saw a decrease in scores when the effort level was set to the highest, as additional reviews and out-of-bound modifications were penalized. This indicates that the notion of "the longer the model thinks, the better" is not always true; process constraints and task scope are more important than setting parameters to their maximum values.

When models are put into the production environment, it's not just about their capabilities; their limitations also need to be taken into account.

Sonnet 5.5 also brings more explicit security configurations. Anthropic states that since the new model's network security capabilities have significantly improved compared to the previous generation, it will continue to use some of the same network security protections and rollback mechanisms as the higher-level models: routine vulnerability fixing tasks can still be carried out, but high-risk requests may be rolled back to the older version Sonnet. The biosecurity protections are the same as those in Sonnet 5. The company has also for the first time used a classifier to prevent large-scale "distillation" of the model's inference capabilities in the Sonnet series. For enterprises, this means that migration is not just about changing the model name; the existing workflow needs to be re-examined to determine which requests may be intercepted, whether the output remains consistent during rollback, and how compliance retention and auditing should be handled.

The official also reminded that users who previously disabled thinking will need to use the new `between_tools` settings during migration. Such technical details may not make for a suitable headline, but they could determine whether the production environment can switch smoothly. Especially in high-throughput processes such as automated customer service, code review, and table generation, a change in response format or a change in the rollback path can have a greater impact than a few percentage points in benchmark tests. Teams should first create samples using real work orders to measure error rates, latency, cost per task, and the proportion of tasks that require manual intervention, before deciding which tasks to transfer from Opus or the old Sonnet.

The signal released by this update is not that "inexpensive models are finally replacing flagship models." On the contrary, Anthropic is refining the product line even further: Opus is designed to handle difficult problems, Sonnet is for daily high-frequency tasks, and in the future, Haiku 5.5 will be aimed at more sensitive cost and scale requirements. What is most meaningful for users is not just the addition of another name to the model family, but rather a shift in large model procurement from "buying the smartest" to "assigning different tasks to the most suitable models." Whether Sonnet 5.5 can truly save companies money can only be determined by the completion rates and review costs of their own businesses.

Cover material: The Dario Amodei portrait from the Anthropic official leadership page. It has been cropped for use on the cover, but this does not imply any direct personal involvement of the individual in this release.