이미지: X — 프론티어랩 화면 갈무리
Summary
- Real estate startup Hypha AI maintained 98% document extraction accuracy with GPT-5.6 Luna while cutting costs by 18x
- Financial research startup Rogo AI cut input tokens by 21% through programmatic tool calling while maintaining evaluation quality
- GPT-5.6 Sol raised its ARC-AGI-3 score from 13.3% to 38.3% while cutting output tokens by roughly sixfold
- 하이파AI 정확도
- GPT-5.5 대비 98% 유지 (GPT-5.6 Luna)
- 하이파AI 비용
- 기존 대비 18분의 1
- 로고AI 토큰 절감
- 입력 토큰 21% 감소, 평가 품질 동일
- GPT-5.6 Sol ARC-AGI-3 점수
- 13.3% → 38.3%
- GPT-5.6 Sol 출력 토큰
- 약 6분의 1로 감소
- GPT-5.6 Luna 가격(배경)
- 100만 토큰당 입력 0.20달러 · 출력 1.20달러 (80% 인하)
- GPT-5.6 Sol 가격(배경)
- 100만 토큰당 입력 5달러 · 출력 30달러
Document processing costs cut by 18x
According to cases published on an OpenAI developer account, real estate startup Hypha AI applied GPT-5.6 Luna to the task of extracting information from documents such as contracts and disclosure materials. The result: compared to the previous model, GPT-5.5, accuracy stayed at 98% while costs dropped to roughly one-eighteenth. Document extraction is a task where even a slight drop in accuracy can cascade into errors downstream, which makes it notable that costs fell sharply while accuracy remained nearly unchanged.
Financial research: 21% fewer tokens, same quality
Financial research startup Rogo AI applied programmatic tool calling — a method where the AI directly invokes needed tools via code and retrieves the results — to the task of pulling and analyzing disclosure materials. This approach cut input token usage by 21% while keeping evaluation quality unchanged. It's a case of lowering costs by refining how tools are called, without switching models.
GPT-5.6 Sol: compressing without sacrificing reasoning
OpenAI confirmed similar results in its own benchmarks. ARC-AGI-3 is a test that measures how well a model solves problems it has never seen before, rather than repeating patterns it has learned — an area where LLMs are notably weak. GPT-5.6 Sol applied "retained reasoning" and "compaction" techniques to raise its score on this test from 13.3% to 38.3%. According to OpenAI, while the score nearly tripled, output token usage actually dropped to about one-sixth. Typically, raising scores requires using more tokens, so this result moved in the opposite direction.
Where cost-to-performance stands now
OpenAI also presented a coding index chart from independent analysis firm Artificial Analysis. Artificial Analysis is a grading organization that measures and ranks models by performance, speed, and price, but does not build models itself. The upper tier visible on the chart can be summarized as follows (approximate values based on chart position):
| Model | Est. API Cost | Est. Index Score |
|---|---|---|
| GPT-5.6 Sol | ~$2,400 | ~80.1 |
| Claude Fable 5 | ~$3,700 | ~77.2 |
| GPT-5.6 Luna | ~$550 | ~77.5 |
| GPT-5.5 | ~$1,700 | ~76.5 |
| GPT-5.6 Terra | ~$550 | ~73 |
| Claude Opus 4.8 | ~$2,400 | ~72.7 |
| Gemini 3.1 Pro Preview | ~$650 | ~42.7 |
At the same cost tier (~$2,400), GPT-5.6 Sol scored higher than Claude Opus 4.8, and GPT-5.6 Luna posted a similar or higher score than Terra at a much lower cost. This means the performance a team can extract from the same budget varies significantly depending on which of the three tiers — Luna, Terra, or Sol — they choose.
An announcement set against a price war
This case study is an extension of the price competition that has unfolded over the past several days. According to a report from August 14, OpenAI cut GPT-5.6 Luna's pricing by 80% for both input and output tokens, bringing it down to $0.20 per million input tokens and $1.20 per million output tokens. Anthropic has also released Opus 5 at half the price of its top-tier model, Fable 5, as the trend of price cuts continues in response to competition from Chinese models such as Moonshot and DeepSeek. This announcement looks more like a demonstration of how those price cuts actually play out in practice. Beyond simply lowering model prices, what matters more for the actual bill is choosing the right model tier for each task and designing tool calls efficiently — a common thread across these cases.
Editor's take
What's really worth paying attention to in this announcement isn't the numbers themselves, but how OpenAI chose its case studies. Promoting "how to do the same work more cheaply" rather than "a smarter model" signals that the axis of competition has shifted from intelligence scores to unit cost. This picture becomes clearer when placed alongside the August 12 report that Grok 4.6 matched GPT-5.6 Sol's intelligence score without changing its price, and the news that DeepSeek 4 Pro priced itself even lower. Right now, the fight isn't over who delivers the best performance — it's over who delivers the same performance more cheaply.
The shift feels bigger when compared generation over generation. Until last year, the default approach to deploying agents was to "pick the single smartest model and use it for everything." Now it has become standard practice to use a cheap tier for document extraction and an expensive tier for complex reasoning, and even within the same tier, to cut tokens simply by redesigning tool calls. The cost gap between teams that just plug in one model and call it done, and teams that segment tasks by tier and refine tool calls, is now measured not in multiples but in orders of magnitude.
For Korean startups, there's one thing to check right now — whether the agents currently in use are running the same model tier across every task. If a top-tier model is still being used for simple extraction or classification tasks, dropping to a Luna-level model will likely cost only a few percentage points of accuracy. Conversely, forcing a cheap model onto tasks that require complex reasoning is also a loss. Experimenting with tiering can start immediately, and the cost savings from rearranging model tiers tend to be far larger than what comes from tweaking a single prompt.
Over the coming weeks, these "cost case study" disclosures won't remain unique to OpenAI. Anthropic and xAI are likely to continue the price competition by touting similar customer cases, and the next round of the battle will likely be fought not over price sheets, but over who documents tiering strategy and tool-call optimization know-how first.
