GPT-6 Soul & Luna: The Cheaper Models That Still Deliver

OpenAI just dropped GPT-6 Soul and GPT-6 Luna, two models built to give you near-flagship performance at a fraction of the cost. They're not trying to beat GPT-6 Astra, the top of the family. They're built to handle the work that doesn't need a flagship, and to do it cheap. I dug into the benchmarks the same day they landed, and the cost-to-performance math is the real story here.

Here's the thing most people miss when a new model drops: they line it up against the biggest, most expensive model available and call it a loser if it scores lower. That's the wrong comparison. Soul and Luna were never meant to out-muscle Astra. They exist to fill the gap between "good enough" and "why am I paying flagship prices for this." And they landed the same day Anthropic shipped Claude Opus 5.5, so the timing forces a real head-to-head.

What Are GPT-6 Soul and GPT-6 Luna?

GPT-6 is a family of models, not a single release. At the top sits GPT-6 Astra, which has been out for a few weeks and is the main event — the model you throw your hardest problems at. Below Astra sit the two new arrivals:

• GPT-6 Soul — the mid-tier. Close to Astra on capability, but at roughly half the cost.
• GPT-6 Luna — the cheapest in the family. A step down in raw score, but priced so low it changes what tasks are economically worth automating.

The point isn't that Soul and Luna are "worse" than Astra. The point is they hit a sweet spot Anthropic has historically struggled to reach: cheap, highly efficient models for tasks that aren't complicated. Anthropic gives you Opus and Fable — both excellent, both expensive. There hasn't been a great budget option. Luna and Soul are OpenAI planting a flag in exactly that gap.

How Do Soul and Luna Score on the Frontier Code Benchmark?

The Frontier Code benchmark is where the cost-to-performance trade-off gets concrete. Watch the price column, not just the score.

• GPT-6 Astra — 53.3% at max effort, costing $4.59 per task.
• GPT-6 Soul — 49.3%, at $2.14 per task. That's less than half the cost of Astra for a drop of just four points.
• GPT-6 Luna — 42.4%, at 11 cents per task.

Read that Soul line again. You give up roughly four percentage points and you cut your cost by more than half. For a huge amount of real work, that's a trade you take every single time.

Now bring in the competition. Claude Opus 5.5 scored 54.4% on the same benchmark at $6.19 per task. Soul comes in at about a third of that cost for a five-point drop in score. Opus 5.5 is the stronger model on paper — nobody's arguing that — but "stronger" and "the right tool" are not the same thing when one costs three times as much.

And Soul scales down gracefully. Drop it from max effort to medium and you're still at 46% for 80 cents per task. That's the flexibility Soul and Luna give you that the premium tier doesn't: you can dial effort to match the job instead of paying top dollar for everything.

How Much Cheaper Is GPT-6 Compared to GPT-5.6?

This is where the release gets genuinely impressive. Compared to the previous 5.6 generation, GPT-6 delivers a small performance bump on Soul but a significant jump in cost efficiency. On Luna it's even more dramatic.

Look at Luna across generations. Performance is basically flat from GPT-5.6 Luna to GPT-6 Luna. But the cost? 15 cents per task on GPT-6 versus $2.57 per task on GPT-5.6. Same output, a fraction of the price.

Across the board, moving from the 5.6 series to the 6 series brings a 50% reduction in both input and output costs. For Luna specifically, input runs just 10 cents and output 50 cents. Those numbers are low enough that entire categories of high-volume, low-complexity work suddenly pencil out. In some cases GPT-6 even edges past Claude Fable 5.1 and beats low-effort GPT-6 Astra — a cheaper model outrunning a more expensive one at the low end.

Did GPT-6 Fix the Accuracy Problem?

Cheaper models usually come with a catch: they hallucinate more. GPT-6 pushes hard on factuality, and the improvements are real across the family.

Luna at max effort used to return an answer with a factual error 12% of the time on 5.6. On GPT-6 that's down to 7.6%. Push Luna to low effort — where you're squeezing for maximum cheapness — and the old 5.6 error rate was a rough 36%. GPT-6 brings that down to 27%. Still not something you'd trust unsupervised, but a meaningful gain for the price tier.

Soul is the headline here. Its factual error rate dropped from 8.5% on 5.6 to 4.6% on GPT-6. For reference, GPT-6 Astra — the flagship — sits at 3.9%. Soul is now within striking distance of the best model in the family on accuracy, at a fraction of the cost. When a mid-tier model gets that close to flagship-level factuality, the case for defaulting to the expensive option gets a lot weaker.

What Changed With Prompt Caching?

Benchmarks get the attention, but caching is where a lot of real-world cost lives, and GPT-6 made it noticeably better.

Here's how caching works: if you've got a warm cache — meaning you've been going back and forth with the model and haven't walked away for an hour — your input gets a 90% discount. Walk away for a day, come back to the same conversation, and the model re-sends the entire conversation and charges you the full input cost. On a long thread, that adds up fast. The same dynamic applies on the Anthropic side.

The big fix: adjusting your reasoning effort no longer resets the cache. Previously, bumping from low reasoning to high would blow away your warm cache and hit you with a large charge. Now you can change reasoning effort mid-conversation and stay warm — less usage, less money, especially if you're building on the API.

OpenAI also added a prompt caching dashboard. If you're on the API and want to get into the weeds, you can optimize cache breakpoints directly. That's a nice quality-of-life win for anyone running these models at volume.

Should You Use Soul, Luna, or Something Else?

The short answer: match the model to the difficulty of the task. The longer answer is that most work doesn't need a flagship.

• Reach for Astra or Opus 5.5 when the problem is genuinely hard — complex reasoning, tricky code, anything where a wrong answer is expensive.
• Reach for Soul when you want most of the flagship's capability without the flagship's bill. At ~half of Astra's cost and within a point of its accuracy, it's the sensible default for a lot of production work.
• Reach for Luna for high-volume, lower-complexity tasks where cost per call is the constraint. At 11 cents per task, it changes the math on what's worth automating at all.

GPT-6 Soul and Luna are available essentially everywhere already. The open question is whether OpenAI ships a GPT-6 Terra to round out the family — worth watching.

On paper, this release might look like a step below what Anthropic did with Opus 5.5. But that framing misses the point. Luna and Soul aren't trying to solve Fable or Opus problems. They're built for the tier a step below — which, frankly, is where most people actually operate — and they handle it far more efficiently. You can't ever discount how much cost matters to these equations. A model that's 90% as good for a third of the price wins a lot more real-world decisions than the leaderboard suggests.

Frequently Asked Questions

What's the difference between GPT-6 Soul and GPT-6 Luna?

Soul is the mid-tier model — close to flagship Astra performance at roughly half the cost, scoring 49.3% on Frontier Code at $2.14 per task. Luna is the cheapest in the family at 42.4% and just 11 cents per task, built for high-volume, lower-complexity work where cost per call matters most.

Are GPT-6 Soul and Luna better than Claude Opus 5.5?

No, and they're not trying to be. Opus 5.5 scored 54.4% on Frontier Code versus Soul's 49.3%, but Opus costs $6.19 per task against Soul's $2.14 — about three times more. They solve different problems: Opus 5.5 for the hardest work, Soul and Luna for everything a tier below.

How much cheaper is GPT-6 than GPT-5.6?

Moving from the 5.6 series to the 6 series brings a roughly 50% reduction in both input and output costs. The starkest example is Luna: nearly identical performance to GPT-5.6 Luna, but 15 cents per task instead of $2.57.

Did GPT-6 improve accuracy over 5.6?

Yes. Soul's factual error rate dropped from 8.5% to 4.6% — close to flagship Astra's 3.9%. Luna at max effort improved from 12% to 7.6%. The cheaper models got meaningfully more reliable, not just cheaper.

What changed with prompt caching in GPT-6?

Adjusting your reasoning effort no longer resets your cache, so you can move between low and high reasoning mid-conversation without losing the 90% warm-cache discount. OpenAI also added a prompt caching dashboard for optimizing cache breakpoints on the API.

---

If you want to go deeper into choosing the right AI models for your workflow, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.