GPT-6.1 Sol 출시: 대체하는 모델이 나온 지 일주일 만에, 5분의 1 가격으로 Near-Astra에 근접

Open​AI released GPT-6.1 Sol on September 29, 2026, at its DevDay 2026 keynote — a refresh of the GPT-6 Sol tier that shipped exactly one week earlier, repriced nowhere, and aimed squarely at a model costing five times as much. The headline numbers are unchanged from the model it replaces: $2.00 per million input tokens and $10.00 per million output tokens. The number that did move is the one most agent workloads actually pay: cached input fell from $0.20 to $0.10 per million tokens, half of GPT-6 Sol's rate and a tenth of GPT-6 Astra's. Open​AI's own framing is that GPT-6.1 Sol "nearly matches" GPT-6 Astra's intelligence on agentic coding, computer use and professional work at one-fifth of Astra's standard token prices, and that is a vendor claim, not a measured one — as of September 30 no independent evaluator has published a single score for the model. What follows separates the two.

무엇이 배포되었고, 기록이 어디에 있는지

The release is unusually well documented for a DevDay announcement. Within hours there was a model page at developers.openai.com, a pricing block, a system-card addendum at deploymentsafety.openai.com, and a commit in the vendor's own public API schema titled "Add gpt-6.1-sol model to model enums," timestamped 17:16 UTC on September 29. Open​AI's developer changelog, by contrast, still ends at the September 22 entry that announced GPT-6 Sol and GPT-6 Luna — the 6.1 entry has not been written up there, so the model page and the launch post are the primary record. The resolution matters because there is no second-generation snapshot to point at: the page's snapshot list contains one identifier, gpt-6.1-sol, with no dated variant, so "GPT-6.1 Sol" and the deployment you call today are the same thing.

그것의 형태를, 그것이 대체하는 것과 나란히 놓고 보면:

• Identifier — gpt-6.1-sol, a single snapshot vs gpt-6-sol with no dated variant either
• Context — 1,050,000 tokens and 128,000 max output tokens for both; the 6.1 page states both figures explicitly, where the 6.0 page documents a 922,000-token maximum input instead
• Knowledge cutoff — April 30, 2026 vs April 20, 2026
• Reasoning effort — low, medium (default), high, xhigh, max, with none and minimal unsupported vs the same ladder plus none
• Pricing — $2.00 in / $0.10 cached / $2.50 cache write / $10.00 out per million tokens vs $2.00 / $0.20 / $2.50 / $10.00
• Long prompts — above 272K input tokens the whole request is repriced at 2× input and cache rates and 1.5× output, identical on both models
• Availability — Plus, Pro, Business, Enterprise and Edu in ChatGPT Work and Codex, plus the API; explicitly not in the consumer Chat product
• Residency — US and EU data residency supported, with fast mode unavailable under EU residency
• Tools — web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search, all supported through the Responses API
• Fine-tuning — not supported on either

마이그레이션하기 전에 짚어 둘 만한 API 형태가 하나 있습니다: GPT-6.1 Sol은 Chat Completions를 지원하지만, 도구 호출에는 Responses API가 필요합니다. GPT-6 Sol에서는 제약이 거의 반대에 가까웠습니다 — Chat Completions의 함수 호출은 reasoning_effort: "none"일 때만 작동했습니다. 스택이 도구 호출을 /v1/chat/completions 엔드포인트를 통해 수행한다면, 그것은 모델 교체가 아니라 코드 변경입니다.

주장으로 표시된 벤치마크의 주장

Every number in this section is Open​AI's, published in the launch post, and none of them has been reproduced by anyone outside Open​AI yet. They are worth reading in full because the pattern across them is consistent — the gains over GPT-6 Sol are real but the framing that travels is the Astra comparison, and the Astra comparison is always a cost comparison.

• DeepSWE v1.1, complex software-engineering tasks in real codebases — Open​AI reports GPT-6.1 Sol matching GPT-6 Astra at roughly one-fifth of the cost, and beating GPT-6 Sol's best score by 6.4 percentage points at a lower reasoning effort and lower cost.
• GDP.pdf, professional questions over complex PDF documents — Open​AI reports GPT-6.1 Sol scoring above Claude Opus 5.5 with fallbacks at less than half the cost per task across tested reasoning settings, and approaching Astra's state-of-the-art result at roughly one-fifth the cost per task.
• AutomationBench 1.0.6, multi-step business workflows across 47 tools — Open​AI reports 2.2 percentage points above Claude Opus 5.5 at medium reasoning effort at roughly a third of the cost, and 4.8 points above GPT-6 Sol at the same setting.
• OSWorld 2.0 offline set, long-horizon computer use — Open​AI reports a seven-point improvement over GPT-6 Sol at maximum reasoning effort at less than half the cost, and a result within 2.1 points of Astra at roughly one-seventh the cost per task.
• Terminal-Bench Science 0.1, scientific workflows — Open​AI reports GPT-6.1 Sol more than doubling GPT-6 Sol's score at maximum effort at less than half the cost per task, at $5.47 per task on average against $23.21 for Claude Opus 5.5 and $23.80 for Astra. Open​AI also states that Astra still holds the top score among models it tested at 68.1% and should be used for the hardest scientific work — a usefully un-marketing sentence.
• Factuality — on de-identified conversations where users had flagged an earlier model's error, Open​AI reports the share of responses containing a factual error falling from 11.4% to 7.7% at low reasoning effort, a reduction of about 32%, with the error rate staying within 1.9 points of Astra's across tested settings at less than one-fifth the cost per task.

Two of those deserve translation. First, the factuality evaluation runs on prompts deliberately chosen because a previous model got them wrong; Open​AI says so in the post, and it means the 11.4% and 7.7% describe a hostile slice of traffic, not your traffic. Second, the paragraph that reads most strongly for the model is the one about search-tool transparency in the safety addendum: on tasks built to elicit failures at maximum reasoning effort, GPT-6.1 Sol fails to tell the user that its search tool is broken in 2.1% of cases, against 4.9% for GPT-6 Sol, 1.5% for GPT-6 Astra and 28.7% for GPT-6 Luna. That is the kind of number that decides whether an agent is deployable, and it is also, again, the vendor's own.

독립적인 기록은 비어 있고, 그것이 바로 이야기의 나머지 절반이다

Artificial Analysis — 이 블로그가 중립적 기준으로 삼는 제3자 게시판 — 에는 GPT-6.1 Sol 항목이 없다. 6.1 Sol 슬러그에 대한 모델 URL은 404를 반환하며, 해당 문자열은 라이브 리더보드 HTML에 나타나지 않는다. 그 보드에 실제로 있는 것은 최대 추론 노력에서의 GPT-6 Sol에 대한 전체 평가다. Intelligence Index 48, 인덱스 작업당 비용 $1.06, 인덱스를 실행하는 동안 생성된 7,700만 출력 토큰(보드 중앙값 8,800만 대비), 그리고 작업당 $2.99의 Coding Agent Index 57이다. 이 수치들은 GPT-6.1 Sol이 대체하는 모델에 대한 가장 가까운 독립적 기준점이며, 모든 6.1 주장이 결국 견주어져야 할 기준이다.

Until that entry appears, anyone quoting a 6.1 Sol score against another model is quoting Open​AI against Open​AI. That is not a reason to distrust the launch post; it is a reason to run your own evaluation set before you move a production path. The vendor gives an unusually concrete instruction on this itself: Open​AI tells developers already using GPT-6 Sol to review the migration guidance before switching, and the migration guidance tells them to compare configurations on representative tasks rather than assuming the highest effort is the best trade-off.

캐시된 입력 할인이 실제로 청구서를 바꾸는 부분이다

가격표는 그대로인데 캐시 요율이 절반으로 줄어드는 변화는 그냥 지나치기 쉽지만, 에이전트를 돌리는 사람이라면 누구에게나 이번 릴리스에서 가장 중대한 항목이다. 백만 토큰당 $0.10인 캐시 입력은 비캐시 입력 요율의 5%, GPT-6 Sol의 캐시 요율의 절반이며 — GPT-6 Astra의 캐시 입력 $1.00과 비교하면 — 플래그십의 10분의 1이다. 안정적인 프리픽스를 수천 번 다시 보내는 에이전트 루프는 거의 전적으로 그 항목에 의존한다. GPT-6 Sol에서 캐시 입력으로 대략 $40이 청구될 2억 토큰 규모의 동일한 에이전트 월 사용량은 GPT-6.1 Sol에서 약 $20이 청구된다; GPT-6 Astra에서는 동일한 캐시 트래픽이 $200이다.

The arithmetic never gets that clean in production, because cache reads only bill at that rate when the prefix actually hits, and Open​AI notes that changing reasoning effort or the available tool set no longer invalidates a cache entry the way it once did — which is the quiet fix that makes the cache line usable in an agent loop at all. The other threshold to watch is unchanged: cross 272,000 input tokens and the entire request reprices at twice the input and cache rates and 1.5× the output rate. Batch and Flex run 50% below standard, fast mode runs at 2× standard, and regional processing adds a 10% premium where it is available.

아직 출시되지 않은 것은 무엇이며, 그것이 라우팅에 의미하는 바

Open​AI says GPT-6.1 Sol Ultrafast arrives "in the coming days" in Codex at up to 8× faster token generation, with no price attached. That is a promise, not a product, and it is worth naming as such: there is no Ultrafast rate card for the 6.1 tier today, and Open​AI has not said whether the 8× figure is measured against standard speed or against fast mode, which is itself priced at 2× standard. Treat the announcement as a scheduling signal and price it when a number appears.

On our side, honestly: GPT-6.1 Sol is not on OrcaRouter's routes yet. The public catalogue endpoint returns "model not found" for openai/gpt-6.1-sol today, and our Open​AI slate carries GPT-6 Astra, GPT-6 Sol and GPT-6 Luna. What is callable on OrcaRouter right now is GPT-6 Sol at Open​AI's own list price with zero markup — the $2.00 / $0.20 / $10.00 rate card, and the 272K repricing rule, passed through exactly as the vendor lists it. That matters more than usual this week for a specific reason: the same pass-through means if Open​AI cuts the 6.1 tier's price, or attaches a number to Ultrafast, the change is live on our side the same day it is live on Open​AI's, with no second contract to renegotiate. When 6.1 Sol lands on the routes it will appear in the catalogue at the vendor's price like everything else, and until then the honest answer is that the model it replaces is the one you can call.

결정의 실용적인 버전: 이미 GPT-6 Sol을 사용 중이고 워크로드가 캐시 집약적이거나 에이전트형이거나 장기 호라이즌이라면, 6.1 리프레시는 당신이 비용을 지불하는 바로 그 차원들을 정확히 개선하며, 마이그레이션은 모델 문자열 변경에, 도구를 사용한다면 Chat-Completions에서 Responses로의 이동을 더한 것뿐입니다. 워크로드가 재사용이 거의 없는 짧은 프롬프트 생성이라면, 이번 릴리스는 당신에게 거의 아무것도 바꾸지 않습니다 — 동일한 입력 가격, 동일한 출력 가격, 동일한 윈도우. 그리고 도입을 결정하기 전에 제3자 점수를 기다리고 있었다면, 그 기다림은 여전히 진행 중입니다.