Anthropic just released a model that makes a lot of AI budgets look different. Claude Sonnet 5.5 came out on September 28, 2026, six days after Opus 5.5, as the second model in the Claude 5.5 family. The headline everyone is sharing: Sonnet comes close to Opus at half the price. 

That headline is mostly true. But if you run AI agents for real business work, the more useful question is: which model should handle which job. 

What Changed With Sonnet 5.5

Anthropic calls Sonnet 5.5 a clear upgrade over Sonnet 5 that runs 30%+ faster, costs up to 30% less for most work, and serves as a faster, lower-cost complement to Opus 5.5.

The price per token didn’t change. Input stays at $2 per million tokens and output at $10 per million. The savings come from the model using fewer tokens and tool calls to finish tasks.

The biggest jump shows up in agentic coding. Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, compared to Sonnet 5’s 10.3%. Anthropic’s own benchmarks even show Sonnet 5.5 beating Opus 5.5 on agentic coding.

Sonnet 5.5 Costs Half of Opus 5.5

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 is exactly half of both.

But half the price per token doesn’t always mean half the bill. How much a model “thinks” before answering affects how many tokens it uses. Anthropic says Sonnet 5.5 costs less per task than Opus 5.5 at lower effort settings, but at higher settings it can reach comparable results at a similar cost.

In other words, if you push Sonnet to its maximum effort on every task, you may end up paying close to Opus prices anyway. But the savings are real when Sonnet is used for the right kind of work.

Where Each Model Fits

Anthropic is clear about what each model is for.

Sonnet 5.5 is strongest at everyday tasks with a clear scope, including fixing bugs and creating polished documents, slides, and spreadsheets.

Opus 5.5 is built for complex tasks that demand careful judgment.

Early customer feedback supports the Sonnet case for routine agent work. One tester said that without changing any prompts, Sonnet 5.5 did better than Sonnet 5 on almost all of their offline Slackbot evals, in fewer steps and with about 14% fewer output tokens.

A Simple Way to Split the Work

You don’t need a complex framework to decide. Ask three questions about each task your agents handle:

Is the task clearly scoped? “Pull last week’s pipeline from Salesforce and post a summary in Slack” has a clear start and finish. That’s Sonnet territory.

Can the output be checked easily? A weekly report, a ticket triage, or a formatted status update is easy to review. If mistakes are easy to spot and cheap to fix, the cheaper model makes sense.

What does a wrong answer cost? A vague request, a long multi-step investigation, or a decision that affects customers or revenue deserves the stronger model. Paying more per task is cheaper than cleaning up a bad call.

For most teams, the answer ends up being a mix. The high-volume, repeatable work goes to Sonnet. The smaller share of judgment-heavy work goes to Opus.

What This Means for WorkflowFiesta Teams

WorkflowFiesta works with Anthropic, OpenAI, and AWS Bedrock. You can connect your own API keys, and prompts go directly to your provider under your own account and data processing agreement. AI token usage is billed either through WorkflowFiesta or directly by your provider if you bring your own model.

That setup matters when model pricing shifts. When a provider releases a cheaper model that handles routine work well, your token costs are where you’ll feel it, and choosing the right model for each kind of task becomes a real lever on your spend.

It also fits how WorkflowFiesta runs work. A single plain-language request can dispatch a chain of specialized agents, each handling its part and passing results forward. Specialized agents doing clearly scoped jobs are exactly the kind of work Sonnet 5.5 was built for.

The Bottom Line

Sonnet 5.5 doesn’t replace Opus 5.5. It changes where Opus is worth the money. Use the faster, cheaper model for the repeatable work that fills most of your team’s week, and save the stronger model for the tasks where judgment matters most.