Compare AI models
| Specifications | ||
ParametersThe learned values that shape how a model responds. B means billion and T means trillion. For mixture-of-experts models, the total count is larger than the number active for each token. Parameter count alone does not measure quality. | 16B | — |
Context windowHow much information the model can work with in one request, measured in tokens — pieces of words or other data. This includes your question, attachments, conversation and reply. Limits can vary by model variant and provider. | 8k | 1.05M |
| API pricingUSD per 1M tokens · lower wins | ||
Input priceWhat you pay for everything you send the model — your question, plus any documents or earlier conversation you include with it. | — | $20.00Daybreak Red approval required |
Output priceWhat you pay for the text the model writes back. It is normally the dearer half: producing an answer costs more than reading one. | — | $100.00 |
Cached input priceA reduced rate for text you send over and over. If every request starts with the same instructions or the same document, the provider keeps a copy ready and charges less to read it again. | — | $2.00 |
| — | $9,000OpenAI list price | |
| These models have no shared benchmark scores. | ||
LiveCodeBenchCompetitive coding — Coding problems published so recently the AI can't have seen them in training — a contamination-free test of raw programming skill. Higher is better. | 14.3% | — |
HumanEvalFunction synthesis — 164 small Python problems: the AI is given a function's description and has to write the working function. This was the coding benchmark of the GPT-3.5 and GPT-4 era, before the field moved to fixing real bugs in real repositories. Higher is better. | 48.1% | — |
MATHReasoning — accuracy on the original competition mathematics test set | 45.3% | — |
WinoGrandeReasoning — resolving ambiguous pronouns in context | 74.6% | — |
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better. | 35.2% | — |
SimpleQAKnowledge — accuracy on short, fact-seeking questions in the original SimpleQA set | 5.6% | — |
MMLU-ProKnowledge — accuracy on the ten-option MMLU-Pro benchmark | 42.4% | — |
MMLUGeneral knowledge — A 57-subject multiple-choice exam — history, law, medicine, maths — that was the standard measure of how much a model knows from 2020 until roughly 2024, when frontier scores crowded into the high 80s and labs moved on to harder tests. The scores here were published years apart under different testing setups, so read them as a historical record rather than a like-for-like ranking. Higher is better. | 70% | — |
GSM8KGrade-school math — Grade-school maths word problems that take a few steps of arithmetic to work through. It separated the models of 2022 and 2023 sharply, then saturated. One caveat on the historical numbers: OpenAI included part of the GSM8K training set in GPT-4's pre-training mix, so GPT-4's score is not a clean few-shot result. Higher is better. | 77.4% | — |
ExploitGym (remote target)Cybersecurity — Can the model exploit a vulnerability it was told about, on a remote target, and show that it used that vulnerability? OpenAI reported this as a percentage at the GPT-6 Cyber launch. It is a different measurement from the timed ExploitGym pass rates and from the count of solved instances. Higher is better. | — | 48.5% |
ExploitGym HardenedCybersecurity — The same kind of intended-exploit test as ExploitGym on a remote target, with stronger defenses turned on, such as address randomization, stack canaries, the V8 sandbox, and stricter kernel settings. Higher is better. | — | 36.94% |
RemoteExploitBenchCybersecurity — Can the model finish an exploit against a known vulnerability in a real application — such as GitLab, PostgreSQL, or Jenkins — running as a remote target in a controlled lab? The score is the share of tasks completed. Higher is better. | — | 95% |
ExploitGym Honeypot · without deployment promptScope adherence — The model is given a hard exploitation task and an easy out-of-scope service it could attack instead. This score is how often it leaves that service alone when no deployment safety prompt is added. Higher is better. | — | 99.93% |
ExploitGym Honeypot · with deployment promptScope adherence — The same honeypot test, with the deployment safety prompt included. The score is how often the model leaves the out-of-scope service alone. Higher is better. | — | 100% |
| Overview | ||
| Company | Moonshot AI | OpenAI |
| Release date | Feb 22 2025 | Nov 5 2026 |
| Access | Open Weight | Closed |
| Model details | View model | View model |
Other comparisons
MoonlightvsClaude Fable 5.2GPT-6 CybervsClaude Fable 5.2MoonlightvsGemini 4 HydrogenGPT-6 CybervsGemini 4 HydrogenMoonlightvsMuse EmberGPT-6 CybervsMuse EmberMoonlightvsGrok 4.7GPT-6 CybervsGrok 4.7MoonlightvsDeepSeek-V4.1-Flash-1017GPT-6 CybervsDeepSeek-V4.1-Flash-1017MoonlightvsMistral Large 4GPT-6 CybervsMistral Large 4Frequently asked questions
Moonlight and GPT-6 Cyber don't publish scores on any of the same benchmarks, so there's no direct head-to-head comparison. Only GPT-6 Cyber has a verified first-party price for an available API: $20.00 per million input tokens and $100.00 per million output tokens. No currently available pay-as-you-go API rate is tracked for Moonlight. The listed GPT-6 Cyber rates use the “Daybreak Red approval required” pricing scope. Moonlight shipped 621 days before GPT-6 Cyber, so benchmark comparisons should account for the intervening progress.
Context windows are 8k (Moonlight) vs 1.05M (GPT-6 Cyber). Moonlight is open weight, while GPT-6 Cyber is closed.
Direct benchmark comparisons are unavailable — Moonlight and GPT-6 Cyber don't publish scores on any of the same benchmarks.