We're beginning a limited preview of the GPT‑5.6 series: Sol, our flagship model; Terra, a balanced model for everyday work; and Luna, a fast and affordable model. Terra has competitive performance to GPT‑5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost.
GPT‑5.6 Sol launches with our most robust safety stack to date. We strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, and spent multiple weeks finding weaknesses, pressure-testing our system, and hardening it against real-world attacks.
We believe in broad access, and we plan to make GPT‑5.6 Sol, Terra, and Luna generally available in the coming weeks. As part of our ongoing engagement with the U.S. government, we previewed our plans and the models’ capabilities ahead of today’s launch. At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly. During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability. We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks, while we work with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases.
GPT‑5.6 Sol is our strongest model yet. To give a preview of model performance, we share a set of evaluations highlighting improved agentic capabilities in coding, biology, and cybersecurity, with additional safety and preparedness evaluations available in our system card(si apre in una nuova finestra). We will share an expanded suite of evaluation results when we make the model broadly available.
With GPT‑5.6, we’re introducing a new max reasoning effort to give Sol the most time to reason deeply. Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
For coding workflows, GPT‑5.6 Sol sets a new state of the art on Terminal‑Bench 2.1, which tests command-line workflows requiring planning, iteration, and tool coordination.
GPT‑5.6 Sol also shows broad improvements in biology workflows. On GeneBench v1, which evaluates long-horizon genomics and quantitative-biology analyses, it achieves stronger results than GPT‑5.5 while using fewer tokens.
GPT‑5.6 Sol is our most capable model yet for cybersecurity. It shifts the performance-efficiency frontier for long-horizon security tasks including vulnerability research and exploitation. On ExploitBench², GPT‑5.6 Sol is competitive with Mythos Preview using only ~1/3 of the output tokens. On ExploitGym(si apre in una nuova finestra)3, a benchmark created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs, GPT‑5.6 Sol, Terra, and Luna models all demonstrate strong improvements in cyber capabilities as we increase reasoning.
We developed GPT‑5.6 Sol, Terra and Luna with our most robust safeguards to date, with configurations matched to each model’s capabilities. As the model becomes more capable, we design safeguards to increasingly hold up to real-world adversarial pressure while preserving access to legitimate work such as code review, vulnerability research, patch development, debugging, security education, and defensive testing. Our goal is to make prohibited offensive activity more difficult, uncertain, and detectable without unnecessarily limiting those beneficial uses. Based on our assessment of the model and safeguards, we expect substantial benefit for legitimate defensive work, while meaningfully constraining prohibited offensive use.
GPT‑5.6 Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks. As these capabilities continue to advance, our priority is to make sure they reach and benefit defenders, who can use these tools to find weaknesses, develop patches, and strengthen systems more broadly.
GPT‑5.6 Sol does not cross the Cyber Critical threshold under our Preparedness Framework. In evaluations involving Chromium and Firefox, it identified bugs and exploitation primitives—the building blocks of an exploit—but did not autonomously produce a functional full-chain exploit under the conditions tested. Still, benchmark thresholds cannot capture every way a model may be used or combined with other tools. That uncertainty, along with the model’s broader step change in capabilities, is why we are pairing the model’s increased capabilities with stronger safeguards and a phased release. We share more details about our safeguards in the GPT‑5.6 Preview system card(si apre in una nuova finestra).
No single safeguard is sufficient against determined or adaptive misuse. Across the GPT‑5.6 preview, we use layered safeguards, with exact configurations varying across models, and pressure-test them for real-world attacks. These include protections trained into the model, real-time checks during generation, account-level signals, differentiated access, monitoring, enforcement, and continued testing.
GPT‑5.6 is trained to refuse prohibited cyber assistance, including when users attempt to disguise their intent or jailbreak the model. These model-level safeguards establish the first boundary around what the model should and should not help with.
Real-time cyber and biology misuse classifiers provide another layer by evaluating output as it is generated. For higher risk cases, if they detect a potential violation, the generation may be paused while a larger reasoning model reviews the conversation and its context. If the output is assessed as disallowed, it is withheld before it reaches the user.
Flagged activity can also trigger account-level review across relevant conversations and risk signals, consistent with our terms and policies around content retention and review. Looking beyond a single conversation helps our systems distinguish persistent malicious behavior from legitimate dual-use security work, where similar technical concepts may appear in very different contexts.
Together, these layers make the overall approach more robust than any one safeguard on its own. Model behavior reduces the likelihood of harmful responses, real-time systems can intervene during generation, account-level review can identify broader patterns, and differentiated access preserves important defensive work without making the most sensitive capabilities broadly available by default.
Especially during the preview, users may encounter safeguards that block or refuse some requests. Other requests may take longer because generation is paused for additional review. Safeguards may occasionally intervene on legitimate work, particularly in dual-use areas where defensive and offensive activity can initially look similar.
That is part of what the preview is designed to test. We want to understand not only whether the safeguards constrain misuse, but whether legitimate users can still complete normal work reliably and efficiently. Feedback during the preview will help us reduce unnecessary blocks and delays, improve how the safeguards interpret context, and create a smoother experience before wider release.
We are also working with enterprise customers on longer-term approaches—including privacy-preserving detection, customer-operated safety controls, and access calibrated to the risk of a customer, user, or workload—to advance safety while supporting enterprise privacy requirements.
Anche le misure di sicurezza devono restare efficaci quando gli attaccanti adattano le proprie tattiche. Una protezione che funziona solo su un insieme fisso di attacchi noti non è abbastanza robusta per un modello di frontiera.
Per questo stiamo applicando alla sicurezza più intelligenza e capacità di calcolo che mai, usando i nostri modelli per individuare debolezze e migliorare più rapidamente le misure di protezione. Abbiamo dedicato oltre 700.000 ore GPU equivalenti ad A100 al red teaming automatizzato mirato a trovare jailbreak universali: attacchi capaci di funzionare in molti prompt o contesti, non solo in uno scenario ristretto. Concentrarci su questi attacchi più difficili e generali ci ha permesso di testare le misure di sicurezza oltre un insieme fisso di fallimenti noti. Ci consente inoltre di esplorare molti più schemi di attacco di quanti ne potrebbe coprire il solo testing umano, individuare prima i pattern di fallimento e accorciare il percorso tra la scoperta di una debolezza e la sua correzione.
Oltre al red teaming automatizzato, abbiamo collaborato con tester di terze parti per condurre un ampio red teaming umano con esperti, che proseguirà durante il periodo di anteprima. Il red teaming umano integra il lavoro automatizzato testando le misure di sicurezza contro esperti creativi che cercano di usare impropriamente il modello in modi che i nostri sistemi potrebbero non prevedere.
Nessuna valutazione può rappresentare ogni configurazione di prodotto, attacco multi-step o workflow reale. Per questo manteniamo un processo di risposta rapida per riprodurre, valutare, dare priorità e correggere i jailbreak appena scoperti, quindi aggiungerli alle nostre valutazioni continue così da poter testare in futuro fallimenti simili.
Durante l’anteprima, i modelli GPT‑5.6 saranno inizialmente disponibili tramite API e Codex per un gruppo selezionato di partner e organizzazioni fidati. Prevediamo di renderli presto disponibili in modo più ampio a chi usa ChatGPT, Codex e l’API.
Nel nuovo sistema di denominazione introdotto con GPT‑5.6, il numero identifica la generazione di un modello, mentre Sol, Terra e Luna indicano livelli di capacità durevoli che possono progredire con una propria cadenza. Nel complesso, la famiglia offre a persone e sviluppatori scelte più chiare tra intelligenza, velocità e costo.
GPT‑5.6 ha un prezzo per 1 milione di token in tre dimensioni di modello: Sol costa 5 USD in input / 30 USD in output; Terra 2,50 USD in input / 15 USD in output; e Luna 1 USD in input / 6 USD in output. GPT‑5.6 introduce anche una cache dei prompt più prevedibile, con supporto per breakpoint di cache espliciti e una durata minima della cache di 30 minuti. Per GPT‑5.6 e i modelli successivi, le scritture in cache vengono addebitate a 1,25 volte la tariffa di input non in cache del modello, mentre le letture dalla cache continuano a ricevere lo sconto del 90% sull’input in cache.
A luglio lanceremo anche GPT‑5.6 Sol su Cerebras fino a 750 token al secondo, portando ai clienti intelligenza di frontiera a una velocità senza precedenti. L’accesso sarà inizialmente limitato a clienti selezionati mentre ampliamo la capacità.
Siamo entusiasti di continuare a imparare da questo periodo di anteprima e di portare presto GPT‑5.6 Sol, Terra e Luna a più persone.
1. Stimiamo latenza e costo API osservando il comportamento in produzione dei nostri modelli e simulandolo offline. Queste stime tengono conto dei dettagli delle chiamate agli strumenti, dei token campionati e dei token di input. I risultati reali possono variare in modo sostanziale e dipendono da molti fattori non considerati nella nostra simulazione. Simuliamo la latenza a velocità API elevate e il costo ai prezzi API standard.
2. Tutti i modelli sono valutati usando l’harness API di ExploitBench con 5 seed e continuità del ragionamento.
3. Abbiamo eseguito ExploitGym sulla nostra API alpha, che produce risposte più rapidamente della nostra API pubblica, e poi abbiamo riscalato i risultati per allinearli alla nostra API pubblica. Quando le latenze vengono riscalate alle velocità previste per la nostra API pubblica, alcune latenze stimate superano i limiti di tempo di 2 e 6 ore, pur essendo state rispettate correttamente nella valutazione. Per ottenere velocità maggiori nei lavori sensibili al tempo, offriamo elaborazione prioritaria nell’API e modalità veloce in Codex.
4. I modelli senza token di output, latenza o costo riportati sono tracciati come linee tratteggiate orizzontali.
