DEV Community: Harshit Sharma

I Calculated What 1M Tokens Costs Across 50+ LLM Models

Harshit Sharma — Mon, 02 Feb 2026 12:13:30 +0000

I spent time compiling pricing data for 50+ LLM models across OpenAI, Anthropic, Google, Mistral, and others. Here's what I found.

The Price Range is Wild

Cheapest: Gemini 1.5 Flash 8B at $0.19/1M tokens
Most expensive: o1 Pro at $750/1M tokens
That's a 3,947x difference

Quick Comparisons

Frontier models (1M tokens):

Model	Input	Output	Total
GPT-5	$1.25	$10.00	$11.25
Claude Opus 4.5	$5.00	$25.00	$30.00
Gemini 2.5 Pro	$1.25	$10.00	$11.25

Budget-friendly options:

Model	Input	Output	Total
GPT-5-mini	$0.25	$2.00	$2.25
Claude Haiku 4.5	$1.00	$5.00	$6.00
Gemini 2.0 Flash	$0.10	$0.40	$0.50

Key Takeaways

Output tokens cost 3-8x more than input - optimize your max_tokens
Newer isn't always pricier - GPT-5 is cheaper than GPT-4 Turbo
"Mini" models are underrated - GPT-5-mini costs 80% less than GPT-
Google is aggressive on pricing - Gemini 2.0 Flash at $0.50/1M is hard to beat.

Full breakdown with all 50+ models:https://withorbit.io/blog/llm-pricing-comparison-50-models

What models are you using? Curious how others are balancing cost vs capability.

I spent $2k on OpenAI before realizing one feature was 70% of it

Harshit Sharma — Thu, 29 Jan 2026 10:48:50 +0000

Built a tool to tag LLM calls with feature names and track costs at that level.

Before: "You spent $2,147 on GPT-4"
After: "summarization: $1,503, chat: $412, search: $232"

SDK wraps your existing client with zero code changes to your API calls:

npm install @with-orbit/sdk
const openai = orbit.wrapOpenAI(new OpenAI(), { feature: 'chat' });

Works with OpenAI, Anthropic, Gemini. Free tier, no credit card.

https://withorbit.io/docs

5 Ways to Cut Your AI Spend (Without Downgrading Models)

Harshit Sharma — Sun, 25 Jan 2026 15:13:40 +0000

Your AI bill is 40% higher than it needs to be.

Zombie retries. Runaway loops. Prompts that "work" but burn 10x the tokens.

Most teams don't catch these until the invoice hits.

Here's how to find and fix them before your next bill.

→ withorbit.io/blog/llm-cost-optimization-5-ways-to-reduce-ai-spend

Built a small tool to understand AI cost & failures per feature — looking for early feedback

Harshit Sharma — Mon, 05 Jan 2026 11:34:18 +0000

Over the last few weeks, I’ve been working with AI features in production, and I kept running into the same problem:

Vendor dashboards (OpenAI, Anthropic, etc.) are great at showing model-level usage, but once AI is embedded across multiple product features, it becomes hard to answer basic questions like:

Which feature is actually driving AI cost?
Where is latency impacting users?
Which AI feature is failing in production?

API keys and model usage don’t map cleanly to how a product is structured.

What I built

I built a small MVP called Orbit to explore this problem.

It’s a lightweight SDK-based tool that captures real runtime data and shows:

AI cost per product feature
Latency per feature
Error rates per feature

The focus is on feature-level observability, not just infra or model analytics.

How it works (high level)

A simple SDK wraps AI calls in your code
Each call is tagged with a feature name
Runtime data (tokens, latency, errors) is sent to Orbit
The dashboard shows how AI behaves inside the product
No proxies, no request interception — just instrumentation.

Who this might be useful for

Engineers shipping AI-powered features
Founders running LLMs in production
Teams trying to understand where AI cost or reliability issues actually come from

*This is very early-stage and currently free.
I’m mainly looking for honest feedback, not signups or validation.
*

What I’d love feedback on

Is feature-level AI visibility something you’ve needed?
What metrics would actually matter to you?
Does this solve a real problem, or is it overkill?
What would make you come back to a tool like this?

If you’re curious, here’s the link:
👉 https://withorbit.vercel.app

Happy to answer questions here, and equally happy if the feedback is “this isn’t useful.”
Thanks for reading 🙏