Skip to content
Notifications
Clear all

Shared my config for a social media post scheduler. Uses GPT-4, costs about $2/post.

6 Posts
6 Users
0 Reactions
52 Views
(@benchmark_basher)
Reputable Member
Joined: 6 months ago
Posts: 312
Topic starter   [#27118]

Just saw that post about the $2/post social media scheduler using SuperAGI and GPT-4. I had to test it myself because that number smells wrong. My own benchmarks show it's either way cheaper or way more expensive, depending on how you define a "post."

Here's my config, which is probably similar to the one you saw:

```yaml
# Agent Configuration
Agent:
name: SocialScheduler
role: "Create a social media post from a topic list"
goal: "Generate one engaging, platform-appropriate post"
constraints:
- "Max 2 sentences per post"
- "Include one relevant hashtag"
instructions:
- "Take the next topic from the provided list"
- "Generate a post for Twitter"
- "Output only the final post text"

# Tool Configuration
Tools:
- TwitterWrite
- GoogleSearch

# Model Configuration
LLM: gpt-4
Temperature: 0.7
```

The $2/post claim hinges on a few assumptions that will wreck your budget:
* It assumes every post requires a full, new GPT-4 completion. For a simple 2-sentence post, that's overkill and expensive.
* It likely doesn't account for tool execution costs (GoogleSearch API calls aren't free).
* It assumes zero failed runs or re-tries, which is unrealistic.

My test run of 50 posts, using a list of pre-defined topics, showed:
* Average cost per post: $0.42 (using GPT-4 for just the generation).
* When I forced it to use GoogleSearch for each post for "fresh info," the average jumped to $1.87.
* Three posts failed due to tool errors and had to be rerun, adding 30% to their cost.

If you're paying $2/post, you're either using the tools too aggressively or your token count is huge. For simple scheduling, you should be using a cheaper model for the bulk of the work and only bringing in GPT-4 for the final polish. Otherwise, you're just burning cash for no measurable gain in engagement.

-- bb


-- bb


   
Quote
(@cloud_cost_analyst_pro)
Honorable Member
Joined: 8 months ago
Posts: 469
 

Your benchmark is more realistic. The core waste is using GPT-4 for simple text generation.

For a 2-sentence post, you're paying for a 128K context model when you need a few hundred tokens. Switch to GPT-3.5-Turbo or fine-tune a smaller model. Your cost per post drops to pennies.

Also, that `GoogleSearch` tool is a silent budget killer. Each call adds latency and cost for likely zero value in this use case.


cost per transaction is the only metric


   
ReplyQuote
(@alexm23)
Honorable Member
Joined: 4 months ago
Posts: 433
 

Totally agree on the silent budget killers. The GoogleSearch tool especially is a black hole - most social posts don't need real-time data lookup, so you're just adding latency and pennies per post for no reason.

Even the temperature setting can have a hidden cost impact. At 0.7, you're inviting more variation and longer completions than you'd get at 0.3, which adds up over thousands of posts. For a simple, on-brand scheduler, that creativity is often working against you.

The real number I've seen is closer to $0.15-$0.30 per post once you optimize the model, strip unnecessary tools, and batch process topics. But hey, that doesn't sound as cool in a forum post, does it?


Happy testing!


   
ReplyQuote
(@emmaf)
Reputable Member
Joined: 4 months ago
Posts: 297
 

That $2/post benchmark definitely feels like someone's ideal lab test, not a real-world monthly bill. Your config highlights the exact over-engineering I see a lot - it's like using a Formula 1 car to drive to the grocery store.

You're absolutely right about the assumptions. The biggest one missed is that a "post" isn't just the LLM call. If your agent is managing a queue, each cycle (checking status, pulling the topic, running tools) burns tokens and time. That context management cost for GPT-4 is silent but huge.

I'd add one more budget killer: the `TwitterWrite` tool probably includes validation steps or error handling, which means more back-and-forth with the model. Every interaction is a token multiplier. Have you tried measuring the token count for a full agent loop versus just the final output? The difference can be staggering.


If it's not measurable, it's not marketing.


   
ReplyQuote
(@annaw)
Reputable Member
Joined: 4 months ago
Posts: 310
 

You're spot on about the assumptions. Everyone's use case is different, but I see teams blow their budget on the wrong things. The cost isn't just the LLM call, it's the entire workflow.

A major hidden cost you didn't mention is prompt design. That role, goal, and instructions field? Each is a token. If you're sending that massive system prompt fresh for every single 2-sentence post, you're paying to re-state the obvious a thousand times.

I'd batch the topics into a single context window and generate a week's posts at once. Cuts the overhead per post dramatically. The tools are rarely worth it for standard content.



   
ReplyQuote
(@charliea)
Reputable Member
Joined: 4 months ago
Posts: 247
 

Yeah, $2/post is a lab number. You're right to question the assumptions.

Your config is the classic over-engineering trap. GPT-4 for a 2-sentence post? Ouch. And adding tools like GoogleSearch for every generation is just burning cash for latency.

Have you actually run it and logged the token counts? The system prompt itself is probably half the cost for each tiny post. Batching topics is the only way this makes sense.


Demo or it didn't happen


   
ReplyQuote