Claude just released Opus, Sonnet, and Haiku as distinct code models, and most project managers don't realize this changes what they can actually automate in their delivery work. The models are not interchangeable. One costs five times more than another. One runs three times faster. One hallucinates less on complex logic. And choosing wrong means either bleeding budget on unnecessary compute or betting critical work on a model that will fail.
Here is the real problem: you probably have recurring PM tasks that eat your week. Status report compilation. Resource capacity checks. Timeline risk flagging. Stakeholder deck assembly. Most of these live in spreadsheets, Jira, and email threads today because they require logic, conditional thinking, and accuracy that simple templates cannot deliver. You could automate them with Claude's code generation capability. But you need to pick the right model, or you waste money or deliver garbage.
The gap is not whether Claude code models can help. It is whether you match the model to the actual complexity of your workflow. Haiku is fast and cheap. Opus is powerful but expensive. Sonnet sits between them. Most PMs reach for the biggest model because they assume bigger is better. Then they pay for compute they did not need and wonder why their monthly API costs spiked.
Start here: what PM task is currently manual, repeatable, and painful enough that you would automate it if you could?
If it is simple, Haiku handles it. A simple task means: take data from one place, apply one or two rules, output the result. Example: extract all overdue tasks from Jira, format them into a bulleted list, send a message to Slack. Haiku reads the data cleanly, applies the logic, writes the code you need. It costs roughly $0.80 per million tokens. It runs fast. For high-frequency tasks like daily status drafts or weekly task summaries, Haiku is your default.
If it is moderately complex, Sonnet is the right call. Moderate complexity means: integrate data from multiple sources, handle conditional logic, produce formatted output that requires structure and reasoning. Example: pull active projects from Jira, compare them against the resource capacity sheet in Notion, flag conflicts, highlight which projects are understaffed, generate a prioritized list for the steering committee. Sonnet can hold multiple data threads in context, reason across them, and produce reliable output. It costs about $3 per million tokens. It is still fast enough for weekly runs. Most PM automation lives here.
If it is genuinely hard, Opus earns its cost. Hard complexity means: work across three or more data sources, apply sophisticated logic like risk scoring or dependency analysis, handle edge cases and exceptions, produce output that feeds executive decision-making. Example: ingest project timelines, budget data, resource constraints, and past delivery metrics, then generate a risk-adjusted schedule and flag which milestones are most vulnerable to slip. Opus has the reasoning depth to handle this without hallucinating intermediate steps. It costs roughly $15 per million tokens. Use it for work that happens infrequently but matters enormously, or for automations that run against sensitive delivery data where failure is expensive.
The honest limitation is this: none of these models can learn from your delivery history on their own. You have to feed them context. If you ask Haiku to draft a status report without giving it actual project data, it will sound like a status report but contain no actual information. The model is not your team's memory. You are. Your job is still to prepare the data and specify the rules. The model just handles the logic and formatting work.
Here is a framework to choose: list three PM tasks you do weekly. For each one, answer these questions. First, how many data sources does it touch? If one or two, lean toward Haiku. If three or more, consider Sonnet. Second, how much conditional logic is involved? If the rule is "flag anything over X threshold," that is Haiku territory. If the rule is "flag anything over X threshold relative to historical patterns or peer projects," that is Sonnet. Third, what happens if the output is wrong? If it is a rough draft that you review before sending, cheaper models work fine. If it feeds directly into a steering committee decision, invest in Opus.
The quick win is this: start with Haiku on your highest-frequency task. Most PMs have one task they do more than once per week. Status drafts. Capacity checks. Task compilations. Build the automation with Haiku first. It is cheap enough that the cost is trivial. It is fast enough that you get feedback quickly. If it works and you want to upgrade to a more complex task, move to Sonnet. If you discover you need better reasoning for the task itself, you know you have found a Sonnet candidate.
Run this experiment for four weeks. Track two things: the actual output quality and the monthly API spend. Do not optimize for cost alone. The real question is this: which model gave you output reliable enough to use without editing it every time, at a price you felt was worth your time back? That ratio, not just the model name, tells you what to use next.
Practical AI intelligence for project managers. Weekly, free. Get frameworks, tools, and decisions that help you stay ahead of AI adoption on your projects. No hype. No filler. Subscribe free →
Stop writing from scratch. Get the 20 prompts PMs actually use for status reports, stakeholder updates, and retro summaries. Download free →