Maanta waxaan sii deynaynaa GPT‑5.4 mini iyo nano, kuwaas oo ah noocyadeenna yaryar ee ugu awoodda badan ilaa hadda. Waxay keenayaan xoogag badan oo GPT‑5.4 ah noocyo ka dheereeya oo hufan oo loogu talagalay culaysyo shaqo oo mug badan.
GPT‑5.4 mini si weyn ayay uga wanaagsanaatay GPT‑5 mini dhinacyada koodh qorista, caqliyeynta, fahamka multimodal, iyo adeegsiga qalabka, iyadoo socota in ka badan 2x si ka dheereeya. Waxay sidoo kale ku dhowdahay waxqabadka nooca weyn ee GPT‑5.4 dhowr qiimeyn, oo ay ku jiraan SWE-Bench Pro iyo OSWorld-Verified.
GPT‑5.4 nano waa nooca GPT‑5.4 ee ugu yar uguna jaban hawlaha ay xawaaraha iyo kharashku ugu muhiimsan yihiin. Sidoo kale waa horumar weyn marka loo eego GPT‑5 nano. Waxaan kugula talineynaa kala-soocid, soo saarista xogta, kala-hormarin, iyo subagents-ka koodh qorista ee qabta hawlo taageero oo fudud.
Noocyadan waxaa loo dhisay noocyada culaysyada shaqo ee daahitaanku si toos ah u qaabeeyo khibradda badeecadda: kaaliyeyaasha koodh qorista ee u baahan inay dareemaan jawaab-degdeg, subagents si dhaqso leh u dhammaystira hawlo taageero, nidaamyada adeegsiga kombiyuutarka ee qabta oo fasira shaashad-qabashooyin, iyo codsiyada multimodal ee ka caqliyeyn kara sawirrada waqtiga-dhabta ah. Goobahan, nooca ugu fiican badanaa ma aha kan ugu weyn—waa kan si degdeg ah uga jawaabi kara, si lagu kalsoonaan karo u adeegsan kara qalab, isla markaana si fiican uga shaqayn kara hawlo xirfadeed oo adag.
| GPT-5.4 (xhigh) | GPT-5.4 mini (xhigh) | GPT-5.4 nano (xhigh) | GPT-5 mini (high¹) | |
|---|---|---|---|---|
| SWE-Bench Pro (Public) | 57.7% | 54.4% | 52.4% | 45.7% |
| Terminal-Bench 2.0 | 75.1% | 60.0% | 46.3% | 38.2% |
| Toolathlon | 54.6% | 42.9% | 35.5% | 26.9% |
| GPQA Diamond | 93.0% | 88.0% | 82.8% | 81.6% |
| OSWorld-Verified | 75.0% | 72.1% | 39.0% | 42.0% |
1 reasoning_effort-ka ugu sarreeya ee GPT‑5 mini waxaa laga heli karaa waa 'high'.
Waa tan waxa macaamiisheennu ka fikireen kadib markii ay ku tijaabiyeen GPT‑5.4 mini iyo nano socod-hawleedkooda:
“GPT-5.4 mini waxay bixisaa waxqabad dhammaad-ilaa-dhammaad ah oo xooggan nooc ku jira fasalkan. Qiimeynadeenna waxay la mid noqotay ama ka sarreysay noocyada tartanka dhowr hawlood oo output ah iyo dib-u-xasuusashada citation-ka iyadoo kharashku aad uga hooseeyo. Waxay sidoo kale gaadhay heerar gudbid dhammaad-ilaa-dhammaad ah oo sarreeya iyo u-yeelid ilaha oo ka xooggan nooca weyn ee GPT-5.4.”
GPT‑5.4 mini iyo nano waxay si gaar ah waxtar ugu leeyihiin socod-hawleedyo koodh qorista ah oo ka faa’iidaysta ku-celcelin degdeg ah. Noocyadu waxay qabtaan wax-ka-beddello bartilmaameed leh, dhex-socodka codebase-ka, abuurista front-end, iyo wareegyada khalad-saarista iyagoo leh daahitaan hoose, taas oo ka dhigaysa kuwo aad ugu habboon hawlaha koodh qorista ee u baahan in lagu dhammaystiro xawaare sare iyo kharash hoose.
Heerarka cabbirka, GPT‑5.4 mini si joogto ah ayay uga fiicnaataa GPT‑5‑mini iyadoo leh daahitaanno la mid ah, waxayna ku dhowdahay heerarka gudbinta GPT‑5.4 iyadoo si aad ah uga dheereysa, taas oo keenta mid ka mid ah isu-dheellitirrada ugu xooggan ee waxqabadka-marka-loo-eego-daahitaanka ee socod-hawleedyo koodh qorista.
Waxaan qiyaasnaa daahitaanka annagoo eegayna hab-dhaqanka wax-soo-saar ee noocyadeenna, kuna dayaneyna tan si offline ah. Qiyaasta daahitaanku waxay tixgelisaa muddada tool call-ka (wakhtiga fulinta koodhka), tokens la muunadeeyay, iyo input tokens. Daahitaanka dunida dhabta ahi si weyn ayuu u kala duwanaan karaa, wuxuuna ku xiran yahay arrimo badan oo aan lagu qaban dayashadeenna. Sidoo kale, kharashaadka waxaa lagu qiyaasay iyadoo lagu salaynayo qiimaha API-ga ee noocyadan xilliga qoraalka. Kharashaadku way is beddeli karaan mustaqbalka. Heerarka caqliyeynta waxaa laga kala qaaday low ilaa xhigh.
GPT‑5.4 mini sidoo kale aad ayay ugu habboon tahay nidaamyada isku dara noocyo cabbirro kala duwan leh. Tusaale ahaan Codex, nooc weyn sida GPT‑5.4 ayaa maamuli kara qorshaynta, isku-dubbaridka, iyo go’aanka ugu dambeeya, halka uu u xilsaari karo subagents-ka GPT‑5.4 mini hawlo-hoosaadyo cidhiidhi ah oo si barbar socda loo qabto—sida raadinta codebase, dib-u-eegista fayl weyn, ama farsamaynta dukumiintiyo taageero ah. Baro sida subagents u shaqeeyaan Codex gudaha docs(ku furmaa daaqad cusub).
Qaabkani wuxuu sii faa’iido badanayaa marka noocyada yaryari ay noqdaan kuwo ka dheereeya oo awood badan. Halkii hal nooc wax walba loogu adeegsan lahaa, horumariyayaashu waxay samayn karaan nidaamyo ay noocyada waaweyni go’aamiyaan waxa la qabanayo, halka noocyada yaryari ay si degdeg ah ugu fuliyaan baaxad weyn. GPT‑5.4 mini waa nooca mini ee noogu xooggan ilaa hadda ee qaabkan socod-hawleed.
GPT‑5.4 mini sidoo kale waa ku xooggan tahay hawlaha multimodal, gaar ahaan kuwa la xiriira adeegsiga kombiyuutarka. Noocku si degdeg ah ayuu u fasiri karaa shaashad-qabashooyinka is-dhexgal isticmaale ee ciriiriga ah si uu hawlaha adeegsiga kombiyuutarka ugu dhammaystiro xawaare. OSWorld-Verified, GPT‑5.4 mini waxay ku dhowdahay GPT‑5.4 iyadoo si weyn uga sarreysa GPT‑5 mini.
GPT‑5.4 mini maanta waxaa laga heli karaa API-ga, Codex, iyo ChatGPT.
Gudaha API-ga, GPT‑5.4 mini waxay taageertaa gelinta qoraal iyo sawir, adeegsiga qalabka, xusida function-ka, raadinta webka, raadinta faylka, adeegsiga kombiyuutarka, iyo skills. Waxay leedahay daaqad context ah oo 400k ah waxayna ku kacaysaa $0.75 halkii 1M input tokens iyo $4.50 halkii 1M output tokens.
Gudaha Codex, GPT‑5.4 mini waxaa laga heli karaa app-ka Codex, CLI, kordhinta IDE iyo webka. Waxay isticmaashaa keliya 30% kootooyinka GPT‑5.4, taas oo u oggolaanaysa horumariyeyaasha inay si degdeg ah ugu qabtaan hawlo koodh qorid oo fudud gudaha Codex qiyaastii saddex-meelood meel kharashka. Codex sidoo kale waxay u xilsaari kartaa subagents-ka GPT‑5.4 mini si shaqada aan u baahnayn caqliyeyn badan ay ugu socoto nooca jaban.
Gudaha ChatGPT, GPT‑5.4 mini waxaa heli kara isticmaaleyaasha Free iyo Go iyada oo loo marayo astaanta “Thinking” ee ku jirta menu-ga +. Dhammaan isticmaaleyaasha kale, GPT‑5.4 mini waxaa loo heli karaa sidii fallback xadka rate-ka ee GPT‑5.4 Thinking.
GPT‑5.4 nano waxaa laga heli karaa oo keliya API-ga waxayna ku kacaysaa $0.20 halkii 1M input tokens iyo $1.25 halkii 1M output tokens.
Macluumaad dheeraad ah oo ku saabsan gaashaannada badbaadada ee noocyadan, fadlan eeg lifaaqa System Card ee ku yaal Deployment Safety Hub(ku furmaa daaqad cusub).
Coding
| GPT-5.4 (xhigh) | GPT-5.4 mini (xhigh) | GPT-5.4 nano (xhigh) | GPT-5 mini (high¹) | |
|---|---|---|---|---|
| SWE-bench Pro (Public) | 57.7% | 54.4% | 52.4% | 45.7% |
| Terminal-Bench 2.0 | 75.1% | 60.0% | 46.3% | 38.2% |
Tool-calling
| GPT-5.4 (xhigh) | GPT-5.4 mini (xhigh) | GPT-5.4 nano (xhigh) | GPT-5 mini (high¹) | |
|---|---|---|---|---|
| MCP Atlas | 67.2% | 57.7% | 56.1% | 47.6% |
| Toolathlon | 54.6% | 42.9% | 35.5% | 26.9% |
| τ2-bench (telecom) | 98.9% | 93.4% | 92.5% | 74.1% |
Intelligence
| GPT-5.4 (xhigh) | GPT-5.4 mini (xhigh) | GPT-5.4 nano (xhigh) | GPT-5 mini (high¹) | |
|---|---|---|---|---|
| GPQA Diamond | 93.0% | 88.0% | 82.8% | 81.6% |
| HLE w/ tool | 52.1% | 41.5% | 37.7% | 31.6% |
| HLE w/o tools | 39.8% | 28.2% | 24.3% | 18.3% |
MM / Vision / CUA
| GPT-5.4 (xhigh) | GPT-5.4 mini (xhigh) | GPT-5.4 nano (xhigh) | GPT-5 mini (high¹) | |
|---|---|---|---|---|
| OSWorld-Verified | 75.0% | 72.1% | 39.0% | 42.0% |
| MMMUPro w/ Python | 81.5% | 78.0% | 69.5% | 74.1% |
| MMMUPro | 81.2% | 76.6% | 66.1% | 67.5% |
| OmniDocBench 1.5 (no tools)² — lower is better | 0.109 | 0.1263 | 0.2419 | 0.1791 |
Long context
| GPT-5.4 (xhigh) | GPT-5.4 mini (xhigh) | GPT-5.4 nano (xhigh) | GPT-5 mini (high¹) | |
|---|---|---|---|---|
| OpenAI MRCR v2 8-needle 64K–128K | 86.0% | 47.7% | 44.2% | 35.1% |
| OpenAI MRCR v2 8-needle 128K–256K | 79.3% | 33.6% | 33.1% | 19.4% |
| Graphwalks BFS 0K–128K | 93.1% | 76.3% | 73.4% | 73.4% |
| Graphwalks parents 0–128K (accuracy) | 89.8% | 71.5% | 50.8% | 64.3% |
1 reasoning_effort-ka ugu sarreeya ee GPT‑5 mini waxaa laga heli karaa waa 'high'.
2 Overall Edit Distance. OmniDocBench waxaa lagu socodsiiyay reasoning_effort oo loo dejiyay 'none' si loo muujiyo waxqabadka kharash-hoose iyo daahitaan-hoose.
