This weekend brought two big pricing stories for anyone running AI agents. OpenAI cut GPT-5.6 Luna API prices by 80%, and DeepSeek V4 Flash went generally available. Both are now the cheapest agent-capable models you can run in Hermes. In the walkthrough below I set up both APIs from scratch, run live research with each model, compare them side by side on real benchmarks, and finish with a mixture of agents setup where both models answer together.
The two headlines
OpenAI cut GPT-5.6 Luna API prices by 80% on July 30. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens. Terra also dropped 20% to $2 in / $12 out, and Sol is unchanged at $5 in / $30 out.
DeepSeek V4 Flash went generally available on July 31. Same model, re-post-trained into the 0731 snapshot, with agent benchmarks far above the preview. It costs $0.14 in and $0.28 out per million tokens.
What each model is, in plain English
GPT-5.6 Luna is OpenAI’s fastest, cheapest model tier. It can use tools and complete multi-step tasks, so it is built for high-volume work. OpenAI says Luna beats their previous frontier model, Fable 5, on professional work tasks at roughly 99% lower cost per task. It is about as capable as frontier models from a year ago at about 6 cents on the dollar per task, nearly 9x faster, with a 1.05M token context window.
DeepSeek V4 Flash is DeepSeek’s small, fast model: 284B total parameters with 13B active, a 1M token context window, 384K max output, and thinking and non-thinking modes. The GA update added a big jump in agent capability, with vendor-reported scores of 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE, 70.3 on Toolathlon, and 25.2 on Agents’ Last Exam. It natively supports the Responses API and is adapted for Codex.
The pricing table
| GPT-5.6 Luna | DeepSeek V4 Flash | |
|---|---|---|
| Input per 1M tokens | $0.20 (was $1.00) | $0.14 ($0.0028 on cache hit) |
| Output per 1M tokens | $1.20 (was $6.00) | $0.28 |
| Context | 1.05M tokens | 1M tokens |
| Max output | 128K | 384K |
| Price change | 80% cut, July 30 | GA July 31 |
| Where it runs | API, Codex, ChatGPT Work, Hermes | API (DeepSeek, OpenRouter), Hermes |
Why this matters in Hermes
Both models run in one Hermes profile, so you can route each task to the cheaper or faster option without switching tools. On the subscription side, OpenAI says ChatGPT and Codex subscription prices and quota budgets stay the same, but Luna and Terra usage now consumes fewer credits. Practically, the same subscription buys roughly 80% more Luna work. In Codex and ChatGPT Work, Plus, Pro, Business, and Enterprise users can pick Luna, Terra, and Sol and now get way more usage.
Watch the full walkthrough
What’s in the video
You’ll see the full OpenAI and DeepSeek API setup, how to hand Hermes your API keys with one prompt across all profiles, live research with each model, benchmark comparisons on Agents’ Last Exam, the coding agent index, and Terminal Bench, the cache-hit mechanic that makes V4 Flash cost nearly nothing on repeat reads, and a mixture of agents demo where Luna and V4 Flash work in parallel with a V4 Pro aggregator.
Setting up both APIs, the short version
Go to platform.openai.com, create an API key, and give it to Hermes with one prompt. Same flow on platform.deepseek.com for V4 Flash. Hermes sets the rest up across your profiles.
Here is my API key: [your key]Please set this up across all profiles to use.
One note on rate limits: starting with $50 in OpenAI credits bumps you to tier two, which raises your tokens-per-minute and requests-per-minute limits. The video shows the exact numbers.
Mixture of agents, coming up
Mixture of agents (MOA) means asking more than one model the same question and merging the answers for a stronger result. In Hermes you can set Luna and V4 Flash as the working models and DeepSeek V4 Pro as the aggregator that synthesizes both answers. I’m putting together a dedicated video on MOA with proper examples, so check back next week.
Who should use what
Luna is the pick for high-volume work that needs to stay on OpenAI infrastructure: Codex, ChatGPT Work, or the API. V4 Flash is the pick for cheap agent automation in Hermes, especially long-running or high-token jobs. Both work in one profile, so the answer is usually “both” and route by task.