Everyday workflows

Five concrete workflows on Llama: writing, research, coding, data and planning. For each, the model to use and what it costs on Meta AI, a hosted API and self-host.

In this lesson you will

Match five common tasks to the cheapest Llama model that handles them
Estimate what each workflow costs on the free, hosted and self-hosted routes
Build a repeatable multi-step workflow instead of one giant prompt

Here are five workflows that cover most of what people actually use an AI model for. For each one there is a model recommendation from the Meta lineup and a cost note for the three routes from Lesson 3. Prices referenced are OpenRouter rates from our catalogue at the time of writing; the price ladder in Lesson 2 is live.

1. Writing and editing

Drafting emails, rewriting paragraphs, adjusting tone. Model: Llama 3.3 70B or Llama 4 Scout. Meta AI: free, and genuinely good enough for most of this. Hosted API: a full day of editing is a few hundred thousand tokens, which is a few cents on Scout. Self-host: overkill unless you already have the hardware.

2. Research and summarising

Reading long documents and pulling out what matters. Model: Llama 4 Scout, because of its very large context window and low input price. Hosted API: a 100,000-token document costs about a cent in input on Scout at the time of writing. Meta AI: fine for short pastes, but no control over model or context. Self-host: only if the documents are confidential.

Figure 1.Use the cheapest model for the high-volume extraction step and a stronger model only for the final synthesis. Most of the tokens run through the cheap model.

3. Coding

Writing functions, explaining errors, generating tests. Model: Llama 4 Maverick for harder problems, Scout or 3.3 70B for routine snippets. Hosted API: coding is output-heavy, and output tokens cost more than input, so cap max output and ask for code only. Local runner: a strong option for developers because you can point editor tools at a local model with no meter.

4. Data tasks

Classifying tickets, extracting fields, cleaning lists. Model: Llama 3.1 8B, with a strict JSON output format from Lesson 5. Hosted API: this is where Llama shines. At $0.05 per million input tokens at the time of writing, a million short records is a small bill. Self-host: worth considering at very high volume or with private data, because the 8B model runs on modest hardware.

5. Planning

Breaking a project into steps, weekly plans, checklists. Model: any. This is low-volume, conversational work. Meta AI: free and ideal. Hosted API: negligible cost. Self-host: pointless for this alone.

Workflow summary. Hosted costs are order-of-magnitude estimates from OpenRouter rates in our catalogue at the time of writing.
Workflow	Suggested model	Meta AI (free)	Hosted API	Self-host
Writing	Llama 3.3 70B / Scout	Good	Cents per day	Overkill
Research	Llama 4 Scout	Short pastes only	About a cent per 100K tokens in	If confidential
Coding	Maverick / Scout	Basic	Output-heavy; cap tokens	Good for developers
Data	Llama 3.1 8B	Not practical	Very cheap at volume	At very high volume
Planning	Any	Ideal	Negligible	Pointless

Knowledge check

In the research workflow above, why does the extraction step use Llama 3.1 8B rather than Llama 4 Scout?

Lesson FAQ

Can Llama replace ChatGPT for daily work?

For writing, planning and summarising, usually yes, and Meta AI does it for free. For the hardest coding and reasoning tasks, test Llama 4 Maverick on your own work before switching.

Which Llama model is best for coding?

Llama 4 Maverick for difficult problems; Scout or 3.3 70B for routine code. Developers who code all day often run a local model to avoid a per-token bill entirely.

Is Llama good for processing lots of data?

Yes. Llama 3.1 8B with a strict output format is one of the cheapest ways to classify or extract from large volumes of short text.