Item — HN Reader

Ever since the insane discount with GPT-5.6 Luna, not much excites me anymore. I mean just look at the benchmarks, even though Gemini 3.7 Flash performs well on the DeepSWE 1.1, Luna (Max) still performs way better. I personally have stuck to Luna (Xhigh) because its been more than enough and does not bloat up the context window too fast with reasoning tokens.

https://deepswe.datacurve.ai

> Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply.

Compare this to Luna which is at $0.2/1M input ($0.02 cached) and $1.2/1M output.

https://developers.openai.com/api/docs/models/gpt-5.6-luna

6 comments

GPT-5.6 Luna is an insanely powerful model for its price. It's been great for coding workflows where I guide the LLM's hand step by step. It's also insane to see my weekly limit drop by than 2% after an hour of coding ever since the discount.

However, I've noticed 2 drawbacks with Luna. Context rot is much more palpable than Terra and Sol. It tends to get confused and go into rabbit holes when it's context gets filled up. In addition, when instructions are vague, it performs poorly and tends to write way to more code than necessary, but that is to be expected of smaller models. In all, for clearly defined, bite-sized coding tasks, Luna's price-to-performance has been insane. It might have very well commanded the price tag of Sol if it came out just a year ago.

Yes it is cheap, but per task DeepSeek v4 Flash is a bit more expensive and lands between Terra and Gemini 3.6 Flash in quality. Closer to Gemini than Terra...

Fable orchestrating DeepSeek v4 Flash to implement a plan is my new favorite thing.

It's so freaking fast, but you gotta tell Fable to watch Deepseek like a hawk or it'll go off the rails.

Yes. It works very well for simple tasks. When I know the context grows over 200k, I implement with Kimi.

We run an agent company and we do a bunch of different things with agents. Where we used Gemini before Deepseek v4 Flash is taking the lead on price. It's like 5x cheaper than 3.6 and well 2.5x cheaper than 3.7 "introductory price". Comparable quality.

Interesting - how are you interacting and orchestrating this?

Not the parent, but:

https://omp.sh/

You define roles for different agents like this:

  modelRoles: 
    task: fireworks/kimi-k3-fast:high
    plan: fireworks/kimi-k3-fast:max
    slow: fireworks/kimi-k3-fast:max
    smol: fireworks/deepseek-v4-flash-0731:low
    tiny: fireworks/gpt-oss-20b
    vision: fireworks/qwen3.7-plus:high
    designer: fireworks/qwen3.7-plus:high
    advisor: openai-codex/gpt-5.6-sol:high
    main_worker: fireworks/kimi-k3-fast:high
    fast_worker: fireworks/deepseek-v4-flash-0731:low
    vision_worker: fireworks/qwen3.7-plus:high
    research_worker: fireworks/glm-5.2:medium
    code_worker: fireworks/kimi-k2.7-code-fast:high
    review_worker: anthropic/claude-fable-5:high
    security_review_worker: fireworks/kimi-k3-fast:max
    minimal_worker: fireworks/gpt-oss-20b
    default: fireworks/kimi-k3-fast
  task: 
    agentModelOverrides: 
      task: "@main_worker"
      sonic: "@fast_worker"
      scout: "@fast_worker"
      designer: "@vision_worker"
      librarian: "@research_worker"
      reviewer: "@review_worker"
      security-reviewer: "@security_review_worker"

Then you first say /plan and use some big model like K3. Finally the harness shows you a markdown you approve, and in approval you switch to a smaller model and reset the context. The smaller model gets the full plan and starts working on it. When done, you say /review and it spawns N review agents and returns the change suggestions. And you iterate on that.

I practically switched to doing everything with Luna or DeepSeek V4 flash. I haven't feel the need for the more expensive models.

I am in the same boat as you. I am using Luna and DeepSeek Flash. Both super fast, super cheap, and I have not felt need for anything more capable in few weeks.

I'd like to try DS4 if Cursor adds it

I'll use it locally too, but we use Cursor for work

Of the two, which do you find better?

I really, really like DSv4 Flash because you see the full, real thinking text. That’s been so useful for helping steer the model; as well as seeing its thoughts and correcting any errors, or expanding on it. It’s so difficult for me to use closed models with no or summarised thinking now — it feels so painful and gimped.

You don’t know what you’re missing until you’ve seen it. For me it’s almost like going from standard def to HD for the first time.

(This applies to other open models too — Kimi K3 in real world feels below Opus 5 in terms of raw intelligence, but significantly above Opus 5 in usability and personality. And no silly refusals — the model feels like it’s working for me; not working for Anthropic who’s always holding a leash over the model while I pay for it).

I prefer V4 flash, but Luna is ok and included in the OpenAI plan I'm already paying, so... that is the reason I use it.

I gave DeepSeek $50 around june, and I haven't been able to exhaust them yet. The model is super cheap and more than enough for my needs.

I'm my opinion, Flash V4 is less pedantic than OpenAI models, less prone to unsolicited prescriptions and less prone to "helpfully" reinterpreting my instructions (wrongly, of course).

luna is the first model that has outdone gpt-5-mini on the pareto frontier for some of my high value, cost sensitive ai product workflows. it's both cheaper (by about 60% in real world use) and higher quality based on my test harnesses. I was really worried that costs would go up since there wasn't a replacement as of a few weeks ago and gpt-5-mini is scheduled to be sunset toward the end of the year. So long as they don't randomly sunset this model anytime soon, that worry has now subsided.

That's interesting, we've had the same results as you at my company. gpt-5-mini was the clear pareto frontier leader for our in-house LLM benchmarks -- benchmarks that we built and tailored to our specific use cases. Then gpt-5.6-luna came along with the price cuts and immediately supplanted gpt-5-mini.

I always thought it was a little odd that gpt-5-mini was the leader for so long when more popular benchmarks placed gpt-5-mini further down the roster, but it seems you had the same result too.

I'm curious how much people are manually curating context these days; I'm increasingly feeling for myself that it being auto-managed inside a front-end like claude code is not ideal, and I'd rather have more control over what exact files and pieces of discovery go into a particular prompt, and the ability to more easily "fork" a session and ask asides or make notes/todos in a way that doesn't disrupt or confuse a more focused task going on.

I don't think I want a gastown-style "just yolo everything" approach, in fact I really want more control over how decisions are made and with what info. Does this exist?

I am not sure this enlightens you with anything but I have a TODO.md file with three headlines. Todo, Doing and Done. The agent is aware of it and knows on which task we are on.

On complete, it moves the user story from Doing to Done. I also have a MEMORY.md file that the agent read and writes in the beginning of a new conversation and at the end of our conversation to update stale information. These files are referred to every time I start a new conversation.

Regarding forking, I know Codex has such button underneath each message that lets you fork the whole conversation. I usually do that when I want to sidetrack and discuss something.

I use no SKILLS or commands like /goal. I’ve come a long way with just prompts and markdown files. Its all a different way of encapsulating instructions anyways.

I like to call them TODO, TODOING. and TODONE. Keeps them in alphabetical order and is not in the slightest OCD...

Benchmarks mean very little. The difference between Luna and Sol in the real world is massive.

It’s a big difference but Luna is very usable. I’ve plugged it into the slot I used to have GLM 5.2 in; I think it’s just as good. And it is less costly. I have Sol do planning and design but do most task execution with Luna now.

> does not bloat up the context window too fast with reasoning tokens

How much does that matter if it's reset at every turn?

what do you mean by reset at every turn? context stays until compaction. if you remove the reasoning tokens after every turn you will be constantly blowing cache which is far worse than filling up context.

That's not my understanding of how most agents work. This is what a chain of request/response looks like:

  Your Prompt 1: Prompt Content 1 -> cache-1
  LLM Response 1: <Thinking>Thinking Content 1</Thinking> Response Content 1
  Your Prompt 2 (client side): prompt-1 + response-without-thinking-1 + Prompt Content 2
  Your Prompt 2 (server side): cache-1  + response-without-thinking-1 + Prompt Content 2 -> cache-2
  LLM Response 2: <Thinking>Thinking Content 2</Thinking> Response Content 2
  Etc...

So reasoning gets dropped from context and you still get cache from the accumulating requests.

Edit:

I've realised I was incorrect, the thinking doesn't get passed back and forth but the latent snapshot does which result in using memory just the same.

Modern protocols loop back the reasoning tokens in raw textual form via an encrypted parameter. You can't see them (modulo the recent attack), but you do resubmit them.

Yeah I've done more research and that's what I meant in the edit you replied to.

But that's not the full reasoning token context, just a snapshot of the latent state at the end of it, no?

Have a look at the gemini ones they're pretty small.

There was an attack that convinced models to leak the contents of the reasoning tokens, and it came back as text that matched the length of the encrypted data very well. So it's probably still tokens.

Looping back latent state isn't that easy. The hidden data is the entire contents of the KV cache which can be massive. I don't think any provider is trying to loop the KV cache through the client, and neural compressions of the reasoning would be lossy / an advanced technique that is still firmly in the realm of research papers, as I understand.

Does it reset at every turn? From my experience in Codex for example, Luna (Max) fills the 256k token window relatively quick. The only thing lowering the context window again is the compaction.

I mean the thinking does not bloat the context window because it gets dropped at the next request.

It doesn't. It's called preserved reasoning and every recent reasoning model does it

Sorry, I've realised I was only partially correct.

Gemini[0] for example passes along a snapshot of the reasoning state but it's not the equivalent to keeping all the reasoning tokens in the context.

[0] https://ai.google.dev/gemini-api/docs/thinking#signatures

Edit: Apparently it does take the same space in the LLM latent space so I was wrong.

Yeah, I remember reading about Anthropic's experiments with dropping out different things, and dropping the thinking is ok for compaction but pretty awful while it still fits in the context window. I could imagine doing it adaptively -- take a snapshot before the thinking, have a lesser model detect if there's a lot of spinning going on and reset to snapshot + result if so, otherwise accumulate. It'd also be interesting to have a lesser model rewrite thinking to be more streamlined (let's pretend the AI went directly down the right path on the first try). But it may be a disaster anyway -- isn't the thinking section something like a 1d best-logit slice of something like a 2d process, where the real processing could easily be happening under the visible surface?

I'm not sure if that's right. I only just recently learned that the "snapshots" (aka cached tokens) necessarily contain the entire context history, not just an image of a "state as of the final token" (well, it is the state as of the final token, but that state contains the whole history). So I'm not confident of my grasp of the structures here.