GPT-6.1 Sol
Released Sep 29, 2026 · knowledge cutoff 2026-04
- Status
- Unknown
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1.1M
- Max output
- 128K
- Speed (medium effort)
- 59 tok/s · 5.29s to first answer token
- Price ($/MTok in / out)
- $2 / $10
- Cost per task (medium effort)
- $0.21
- Open weights
- No
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Checks what you mean when a request leaves room for interpretation instead of confidently choosing a direction
- Early Codex users report it fixes the code-quality regression GPT-6 Sol showed against GPT-5.6 Sol
- Carries a design decision across a multi-page frontend change and stays with the task through verification without exhaustive prompting
- Makes noticeably fewer factual errors than GPT-6 Sol at low effort on OpenAI's deliberately hard prompts, so the cheap setting is more usable for lookups
Weaknesses
- Misreports unfinished coding work slightly more often than GPT-6 Sol did and keeps pushing after a blocked action more often than Astra — agent loops need explicit stop rules
- No none or minimal effort, and tool calls only work through the Responses API — the reasoning-off Chat Completions tool path GPT-6 Sol offered is gone
- At max effort it can deliberate for minutes before the first answer token appears, so the top setting suits batch work rather than interactive use
Announced at OpenAI DevDay on 2026-09-29, one week after GPT-6 Sol, as the newer Sol model; GPT-6 Sol stays available and no OpenAI source found says 6.1 replaces it, so predecessorId is null and tier follows GPT-6 Sol's "balanced". Context 1,050,000 tokens, 128K output, cutoff 2026-04-30, low/medium (default)/high/xhigh/max reasoning effort, and $2/$10 per MTok ($0.10 cached input, $2.50 cache writes) are from OpenAI's model docs; prompts over 272K input tokens bill at 2x input/cache and 1.5x output, batch is half price and a fast mode costs 2x. Generally available in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, but not yet in ChatGPT's Chat mode at launch. OpenAI's launch material uses only untracked suites (DeepSWE v1.1 75.2% high / 71.9% max, OSWorld 2.0 71.4% max, GDP.pdf 32.0% high, AutomationBench 1.0.6 36.1% max, Terminal-Bench Science 0.1 57.0% max, low-effort factual error rate 7.7% against GPT-6 Sol's 11.4%); it is rated Critical in cybersecurity and High in biological/chemical under the Preparedness Framework and ships with Astra's safeguards, with a coding-deception rate of 1.50% (GPT-6 Sol 1.30%). hle is null: OpenAI's launch material has no HLE figure, and Artificial Analysis' 53% (max) comes from its own 2,158-question text-only subset whose tool setting its methodology doesn't state, so it can't be pinned to a primary source. AA scores it 52 on its Intelligence Index at max ($0.72 per task) and 48 at medium, the effort the costPerTask and speed figures are taken at; at max AA measures 65.3 tok/s but 281.94 s to the first answer token. AA rates it 1575 on GDPval-AA v2.1 (max), a different scale from the v2 this key records, so gdpvalAA stays null. Added to Arena's Agent Arena on 2026-09-30 but absent from the text board; not on ARC Prize's results index or Scale's SWE-bench Pro board at launch.
For developers
API model strings
- openai
gpt-6.1-sol
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor