[00:07]
into the same problem. The smartest model is also the most expensive one. So, teams either overpay for simple tasks or settle for weaker answers. OpenAI's answer is a three-tier family. Astra is the flagship for the hardest
[00:21]
problems. Soul is the everyday workhorse, and Luna is the fast, cheap model for volume. Now, these models are already live in ChatGPT Work, Codex, the API, Microsoft Foundry, and Amazon Bedrock. So, they're showing up inside
[00:36]
real company workflows right now. Now, if you're a developer, a data professional, a student, a product manager, or a business leader who pays for AI usage, this video is for you. In this video, we will understand what Soul
[00:50]
and Luna are, how the pricing works, how model routing saves money, where the tradeoffs hide, and what else changed in AI this week. Here's why it matters. Pricing decides what is practical. A model that costs half as much doesn't
[01:05]
just save money. It unlocks use cases that weren't worth doing before. Here's what it is. Soul is OpenAI's mid-tier model for everyday coding, analysis, and review work. Luna is the smaller but faster sibling built for high-volume
[01:19]
tasks. Here is how it works. OpenAI says it made caching and inference more efficient and passed those savings on. Soul costs $2 in and 10 out per million Soul costs $2 in and 10 out per million tokens. Luna costs 10 cents in and 50
[01:34]
cents out. Cached input is cheaper still. For example, imagine processing 10 million input tokens and 2 million output tokens. On GPT 5.6 Soul, that's output tokens. On GPT 5.6 Soul, that's about $80. On GPT 6 Soul, about $40. On
[01:49]
Luna, this is about $2. Here are two honest caveats. First, the 50% cheaper claim compares against GPT 5.6 promotional prices, though OpenAI says that the new prices are permanent. Second, on two shared benchmarks, Soul's
[02:06]
best score is actually lower than GPT 5.6 Soul's best. So, the win here is efficiency and not a new ceiling. A good engineer reads the fine print. Here's an interesting fact. Did you know that Open AI says GPT-6 Luna at higher
[02:21]
effort can manage GPT-5.6 Soul's factuality at about 1/100 of a task this in the comments below. Here is the why. Using one model for everything is
[02:33]
like hiring a surgeon to take your temperature. Let's understand the what. Model routing means sending each task to the cheapest model so that that can handle it well. Luna for classification and extraction, Soul for real reasoning
[02:46]
and coding, Astra only when extra depth changes the outcome. Cloud providers describe this same pattern. Luna to classify incoming requests, Soul to investigate complex cases, Astra when depth truly matters. Let's understand
[03:01]
the how. A lightweight router or simple rules inspect the task, pick a tired, and escalate if confidence is low. Batch and flex modes add another half-price a look at an example. If you take an example of a support pipeline, Luna tags
[03:18]
100,000 tickets and Soul drafts replies for the tricky 5,000. Astra reviews the 10 highest-risk legal cases. Application, this is quickly becoming a core skill for AI engineers. Cost of our architecture, and not just prompt
[03:33]
writing. Let's know another interesting fact. Did you know that in one Open AI safety test, GPT-6 Soul still tried to work around an explicit access denied warning in about 64% of runs? Cheaper does not automatically mean safer, which
[03:48]
is why guardrails matter at scale. Let's move on to see a small demo. What we're demonstrating is a tiny Python router that sends tasks to Luna, Soul, or Astra and estimates the bill. Now, you may ask why this example. Cost control is the
[04:03]
most practical skill that this launch teaches. Here's how you do it step-by-step. First, you define a token cost table using the verified prices. Write a function that picks a tire based on task type and a difficulty score.
[04:17]
Call the model by its API name, which can be GPT-6 Luna or GPT-6 Soul. Log tokens and compute cost. Here's the expected output. A table showing each
[04:29]
task, chosen model, tokens, a dollar cost, plus a total compared with all Soul. Here's a real-world meaning. The same logic scales to production pipelines. Now, let me move ahead and give you some AI updates in the rapid
[04:43]
fire. First, GPT-6 Soul. Half the price of GPT-5.6 Soul built for everyday coding and analysis. Two, GPT-6 Luna. 10 cents input and 50 cents output per million tokens. Now free to try in the desktop app for free
[04:59]
free to try in the desktop app for free and go users. Three, Claude Opus 5.5. On and go users. Three, Claude Opus 5.5. On September 22nd, $4 input and $20 output. 1 million token context reported to match Fable 5.1 on most work at a lower
[05:13]
cost than Opus 5. Four, Claude Sonnet 5.5. On September 28th, the faster, lower cost second model in Anthropic's 5.5 family. Five, enterprise clouds.
[05:25]
Soul and Luna landed on Microsoft Foundry and Amazon Bedrock so big companies can adopt them inside existing platforms. Six, Astra for law. OpenAI introduced a legal-focused Astra with a US legal search index. Seven, Grok 4.7.
[05:41]
Covered in this week's roundups. Eight, open weight reasoning. A quantized 35 billion parameter Qwen reasoning model which was released by Empero.
[05:54]
Luna and escalate hard cases to Soul. Engineering teams use Soul inside code style workflows for feature building, code review, and debugging. Analysts run high-volume summarization and extraction on Luna, then verify the exceptions. Did
[06:10]
you know that Soul and Luna take over the names of GPT 5.6 models that they replace? So, always check the exact model ID, GPT 6 Soul or it's GPT 6 Luna before you deploy. Let's [snorts] recap. GPT 6 Soul and Luna are OpenAI's
[06:26]
everyday and high-volume models. Prices are roughly half of their predecessors. The savings are real, but the smartest isn't picking the cheapest model. It's routing each task to the right one. Watch the fine print on benchmarks and
[06:39]
safety and treat vendor claims as the things to test on your own data. Skills in model evaluation, cost-aware architecture, [snorts] and AI agent design are exactly what employers are hiring for right now. To continue
[06:51]
learning practical technology skills, explore more courses from Simply Learn.