How to Choose the Right Claude Model: A Guide for Enterprise Teams
Choosing the right Claude model starts with understanding what you need AI to do, how quickly it needs to respond,…
Choosing the right Claude model starts with understanding what you need AI to do, how quickly it needs to respond, and what your budget can support at scale. Anthropic offers several Claude models, each designed for different workloads. Some prioritize speed and affordability. Others are built for complex reasoning, analysis, and autonomous work. The best choice is the model that reliably completes your task at a speed and cost that work for your organization.
To see how that works, we’ll follow a familiar IT scenario: a VPN outage affecting an entire office in Chicago. We’ll trace the incident through intake, investigation, and remediation to see how different Claude models can support each stage.
Claude is Anthropic’s family of large language models (LLMs), designed for work as simple as ticket classification and as complex as multistep autonomous workflows. The current lineup includes four primary models: Haiku, Sonnet, Opus, and Fable. While all four can process and generate text, analyze information, and use tools, they differ in speed, cost, and the complexity of work they’re best suited to handle.
| Model | Best for | Relative speed | Relative cost |
| Claude Haiku 5.5 | High-volume classification, data extraction, routing, and simple automated tasks | Fastest | Lowest |
| Claude Sonnet 5.5 | Everyday business tasks, summarization, writing, coding, and analysis | Fast | Moderate |
| Claude Opus 5.5 | Complex reasoning, enterprise workflows, coding, and autonomous agents | Moderate | Higher |
| Claude Fable 5.1 | Demanding reasoning, extended research, and long-running autonomous tasks | Slower | Highest |
Source: Anthropic’s official model documentation. Model availability, capabilities, and pricing may change. See Anthropic’s pricing page for current rates.
Anthropic currently recommends starting with Claude Opus 5.5 for most workloads. That gives you a baseline for what good output looks like, so you can test whether a faster, less expensive model can match it for each step.
Before picking a model, ask about capability, speed, and cost. Anthropic frames model selection as a balance among these three. Each task weighs them differently, so the right model depends on the step in front of you.
Capability: how hard is the thinking? Classifying a ticket into “Network” with priority P2 is a narrow task. Finding the shared root cause across 200 related incidents takes deeper reasoning. The harder the reasoning, the more a stronger model pays off.
Speed: how long can the user wait? Triage that runs inside an intake form needs to respond in about a second. A weekly problem review can take minutes without affecting the team’s workflow.
Cost: how many times will this run? A step that fires on every incoming ticket multiplies any per-call cost by your full ticket volume. A step that runs once a week has a much smaller impact on the budget.
When a step is high-volume and latency-sensitive, speed and cost dominate. When it’s rare and high-stakes, capability dominates.
Match the model to the stage: Haiku for intake, Sonnet for working the ticket, Opus for problem management, and Opus at higher effort or Fable for autonomous remediation. Treat these as starting points, and let your own evaluations make the final call.
| Stage | What happens | Model fit |
| Intake | Classify, set category and priority, route to the assignment group | Haiku |
| Working the ticket | Summarize work notes, draft the employee update, suggest KB articles | Sonnet |
| Problem management | Root-cause analysis across related incidents and CMDB relationships | Opus |
| Autonomous remediation | Multi-step agent that investigates, proposes a fix, drafts the change request | Opus at higher effort, or Fable for long-horizon runs |
Let’s return to our Chicago VPN outage.
Employees are reporting connection failures. The service desk is receiving tickets, IT teams are investigating, and leadership wants updates.
Here’s how different Claude models could support each stage of the response.
Claude Haiku is the starting point for high-volume incident classification and routing. Its speed and lower cost make it well suited to repetitive tasks that follow established rules, such as identifying incident categories, extracting relevant details, and recommending assignment groups.
The VPN ticket lands. A short, messy message from a Chicago employee (“can’t connect, third time this week, whole floor affected”) needs a category, a priority, and an assignment group. This is classification, and a small model is usually enough. You’d move up a tier if Haiku misroutes ambiguous tickets in your evaluations, or if priority depends on subtle business context. If accuracy holds, Haiku is the right fit.
Claude Sonnet is a good default for everyday ITSM tasks that require understanding context and generating useful written responses. It can summarize incident histories, draft employee communications, and help analysts locate relevant knowledge articles without requiring the capabilities of a more expensive reasoning model.
By mid-morning the VPN ticket has thirty work notes from three engineers. Someone needs a clean summary for the incident manager and a plain-language update for the affected office. Sonnet handles both well at a speed that keeps the agent in flow. You’d move up to Opus if summaries miss critical technical details in your evaluations, or try Haiku for simple, templated updates where quality is already high.
Claude Opus is a strong candidate for complex problem management and root-cause analysis. These tasks often require examining large amounts of information, identifying relationships between systems, and evaluating multiple explanations before recommending a course of action.
The Chicago outage is one of several. Over the past quarter, 200 incidents mention VPN drops, certificate warnings, or a particular gateway. Problem management asks: is there a single underlying cause, and what in the CMDB connects the affected services? A model that holds a lot of context and reasons carefully across it is worth the extra cost, because this step runs rarely and the payoff of a correct diagnosis is large. Before moving up to Fable, try Opus at higher effort (more on that below). You’d move down to Sonnet if your evaluations show it finds the same root causes.
Effort settings control how thoroughly a Claude model works through a request, which makes them one of the most useful ways to tune quality, speed, and cost without switching models. Higher effort gives the model room to reason more carefully. Lower effort keeps responses leaner and faster.
The setting affects every token the model produces, including its reasoning, its written response, and any tool calls it makes.
There are five levels: low, medium, high, xhigh, and max. Each model has its own default, which matters more than many teams realize. Claude Fable 5.1 and Claude Sonnet 5.5 default to high effort, while Claude Opus 5.5 and Claude Haiku 5.5 default to medium. So when we suggest trying “Opus at higher effort” before moving to Fable, that means raising Opus above its default.
In our VPN scenario, effort gives you a second dial at every stage. Intake classification may hold its accuracy at low effort, which cuts cost on a step that runs for every ticket. Sonnet’s summaries run at high effort by default, so testing medium could lower cost without a noticeable drop in quality. And for the quarterly root-cause review, raising Opus to xhigh may close the gap you’d otherwise try to solve with a more expensive model.
The approach is the same as model selection. Test each level against your evaluation set, and keep the lowest setting that consistently meets your quality bar.
Yes, and most production systems do. A common pattern is routing: a fast model handles every request and escalates only the hard ones. Another is orchestration, where a stronger model plans the work and delegates pieces to cheaper sub-agents.
In our VPN incident workflow, both appear. Haiku classifies every incoming ticket. Routine tickets flow to Sonnet for summaries and updates; tickets that look like they’re part of a larger pattern get flagged for Opus-level problem analysis. In an agentic version, an Opus or Fable orchestrator plans the investigation while Haiku or Sonnet sub-agents handle bulk lookups, log summarization, and drafting. You pay for top-tier reasoning only where it changes the outcome. This is also where implementation details start to matter: escalation thresholds, fallback behavior, and monitoring. Those are worth designing deliberately.
There are other ways to manage cost, too. Anthropic’s Batch API offers a 50% discount for work that doesn’t need an immediate response, making it a good fit for something like a weekly problem review. Prompt caching can also reduce costs for high-volume intake workflows, where the same instructions are used across thousands of tickets.
The most reliable way to select a Claude model is to test it against representative work from your own environment. Published benchmarks can tell you a lot about general model capabilities, but they can’t tell you whether a particular model will correctly categorize your incidents, interpret your terminology, or meet your standards.
Start with a representative test set. For an incident management use case, that might mean selecting 50 real, anonymized incidents that include straightforward service requests, common technical issues, ambiguous descriptions, high-priority incidents, and complex cases involving multiple systems. Remove personal information, credentials, sensitive business data, and anything else that shouldn’t be shared with the model.
Next, define what a successful result looks like before you run the test. If you’re evaluating incident classification, you might score category selection, assignment group recommendations, extraction accuracy, handling of incomplete information, response time, cost, rework rates, and escalation rates. If you’re evaluating root-cause analysis, you might look at whether the model identifies relevant evidence, separates facts from assumptions, and recommends reasonable areas for further investigation.
Run those same examples through the models you’re considering using consistent instructions and comparable context. Test different effort settings, too. Then compare the results against the criteria you established.
The principle is straightforward: choose the least expensive model that consistently produces work you’d be comfortable putting into production. If Haiku meets your requirements, use Haiku. If the quality improves significantly with Sonnet, evaluate whether that improvement justifies the difference in cost. If only Opus can reliably handle the task, you’ve learned something important about the complexity of the workflow. A more capable model can sometimes cost less per finished task if it requires fewer retries, corrections, or escalations.
And don’t stop evaluating once the application goes live. Models, prompts, workflows, and business requirements change. Your evaluation set gives you a repeatable way to test those changes rather than relying on assumptions.
The simplest approach is to match the complexity of the work to the least expensive model likely to handle it, then validate that assumption with an evaluation (eval), meaning a structured test against your own data.
| If your task involves… | Start by testing… |
| High-volume classification, extraction, or routing | Haiku |
| Everyday writing, summarization, or analysis | Sonnet |
| Complex reasoning, problem solving, or autonomous workflows | Opus |
| Demanding, long-running reasoning or agentic work | Opus at higher effort, then Fable |
If the task is repetitive, structured, and high volume, start with Haiku. If it requires writing, summarization, or moderate analysis, Sonnet is a logical candidate. For complex reasoning, extensive investigation, or workflows involving multiple tools, start with Opus. If Opus still falls short after you’ve optimized the prompt, context, and effort setting, evaluate Fable. For security-adjacent work, such as vulnerability investigation, Fable’s additional cybersecurity safeguards may affect its behavior, making it especially important to test against your actual tasks.
Use the table to decide what to test first, and let your results confirm the fit.
Choosing a Claude model is one decision inside a larger design. You also need to know where AI can make a meaningful difference, which processes are worth automating, what information each model needs, and how it will work with the systems your teams already use.
Governance belongs in that design from the start. Someone should own the escalation thresholds that decide when a ticket moves from Haiku to Opus. Each step should have clear limits on what data the model can see, and you’ll want monitoring that shows when a workflow’s quality or cost starts to drift. Those decisions are what turn a promising pilot into work your organization can rely on.
For some workflows, Haiku may be all you need. Others will benefit from Sonnet, Opus, or Fable, and many enterprise applications will use a combination. What matters most is that each choice reflects how your organization works and what your testing shows.
At Beyond20, we help organizations find the right opportunities for AI, evaluate model performance against real work, and design governed workflows around their business requirements. Our experience across IT service management, enterprise workflows, and the ServiceNow platform helps connect Claude’s capabilities to the processes your teams depend on. If you’re earlier in the journey, our AI Readiness Assessment and Putting AI to Work white paper are good places to start.
Ready to put Claude to work in your organization? Talk with our team about model selection, cost optimization, and where AI can help your teams most.
There isn’t one Claude model that’s best for every ServiceNow-related workflow. Haiku may suit high-volume ticket classification, Sonnet may work well for incident summaries and employee communications, and Opus may be appropriate for complex problem management. The best choice depends on the task, available data, performance requirements, and implementation architecture.
Claude Opus may justify its higher cost for complex ITSM tasks requiring advanced reasoning, such as analyzing related incidents, interpreting CMDB dependencies, or investigating recurring problems. For routine classification and summarization, less expensive models may deliver the quality you need. Testing against representative workloads is the best way to determine whether the additional capability provides enough value to justify the cost.
Claude Sonnet balances speed, capability, and cost, making it a strong candidate for many everyday enterprise tasks. Claude Opus is designed for more demanding reasoning and agentic workflows. Sonnet may be sufficient for summarizing incident records, for example, while Opus may perform better when investigating complicated relationships across incidents, changes, configuration items, and other sources.
Yes. Applications can use different Claude models for different tasks within the same workflow. Haiku might classify incoming incidents, Sonnet might generate employee updates, and Opus might investigate complex technical issues. This can help control costs by reserving more capable models for the tasks that require them.
Consider evaluating Claude Fable for particularly demanding reasoning and long-running autonomous workflows, especially when Opus doesn’t meet your requirements after you’ve optimized the prompt, context, and effort settings. Because Fable involves different speed and cost tradeoffs, the decision should be based on whether it meaningfully improves results in your evaluations.
Anthropic regularly introduces new models and updates its recommendations, capabilities, and availability. That’s why your model-selection strategy shouldn’t depend entirely on today’s model names. Keep a repeatable evaluation set and use it to test new model versions before changing a production workflow.
Beyond20 is a ServiceNow Elite Partner dedicated to changing work life for our clients through rapid time to value across the ServiceNow platform. We help organizations make ServiceNow work in practice and for people, combining platform-spanning expertise, practical guidance, and hands-on delivery to help teams move with confidence and see value faster.
