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(s'obre en una finestra nova). 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(s'obre en una finestra nova)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(s'obre en una finestra nova).
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.
Les salvaguardes també han de continuar sent eficaces quan els atacants adapten les seves tàctiques. Una protecció que només funciona amb un conjunt fix d’atacs coneguts no és prou robusta per a un model d’avantguarda.
Per això estem aplicant més intel·ligència i capacitat de càlcul que mai a la seguretat, fent servir els nostres propis models per trobar febleses i millorar les salvaguardes més ràpidament. Hem dedicat més de 700.000 hores de GPU equivalents a A100 a l’equip vermell automatitzat per trobar evasions de les proteccions de seguretat universals: atacs que poden funcionar en moltes indicacions o contextos, no només en un entorn estret. Centrar-nos en aquests atacs més difícils i generals ens ha permès posar a prova les salvaguardes més enllà d’un conjunt fix d’errors coneguts. També ens permet explorar molts més patrons d’atac dels que podrien cobrir només les proves humanes, identificar abans els patrons d’error i escurçar el camí entre trobar una feblesa i corregir-la.
A més de l’equip vermell automatitzat, hem treballat amb verificadors externs per dur a terme un equip vermell humà expert extens, que continuarà durant el període de vista prèvia. L’equip vermell humà complementa el treball automatitzat posant a prova les salvaguardes davant d’experts creatius que intenten fer un mal ús del model de maneres que els nostres sistemes potser no anticiparien.
Cap avaluació pot representar totes les configuracions de producte, tots els atacs de diversos passos ni tots els fluxos de treball del món real. Per això mantenim un procés de resposta ràpida per reproduir, avaluar, prioritzar i corregir les evasions de les proteccions de seguretat descobertes recentment, i després afegir-los a les nostres avaluacions contínues per poder provar errors similars en el futur.
Durant la vista prèvia, els models GPT‑5.6 estaran disponibles inicialment a través de l’API i Codex per a un grup seleccionat de socis i organitzacions de confiança. Tenim previst posar-los aviat a disposició de manera més àmplia per a les persones que utilitzen ChatGPT, Codex i l’API.
En aquest nou sistema de noms introduït amb GPT‑5.6, el número identifica la generació d’un model, mentre que Sol, Terra i Luna identifiquen nivells de capacitat duradors que poden avançar al seu propi ritme. En conjunt, la família ofereix a persones i desenvolupadors opcions més clares en intel·ligència, velocitat i cost.
GPT‑5.6 té un preu per cada milió de segments en tres mides de model: Sol costa 5 $ d’entrada / 30 $ de sortida; Terra, 2,50 $ d’entrada / 15 $ de sortida; i Luna, 1 $ d’entrada / 6 $ de sortida. GPT‑5.6 també introdueix una memòria cau d’indicacions més predictible, amb compatibilitat per a punts de tall de memòria cau explícits i una vida mínima de la memòria cau de 30 minuts. Per a GPT‑5.6 i models posteriors, les escriptures a la memòria cau es facturen a 1,25 vegades la tarifa d’entrada sense memòria cau del model, mentre que les lectures de memòria cau continuen rebent el descompte del 90% d’entrada en memòria cau.
També llancem GPT‑5.6 Sol a Cerebras al juliol, amb fins a 750 segments per segon, per portar intel·ligència d’avantguarda als clients a una velocitat sense precedents. Inicialment, l’accés estarà limitat a clients seleccionats mentre ampliem la capacitat.
Ens entusiasma continuar aprenent d’aquest període de vista prèvia i portar aviat GPT‑5.6 Sol, Terra i Luna a més persones.
1. Estimem la latència i el cost de l’API observant el comportament en producció dels nostres models i simulant-ho fora de línia. Aquestes estimacions tenen en compte els detalls de les crides a eines, els segments mostrejats i els segments d’entrada. Els resultats reals poden variar substancialment i depenen de molts factors que la nostra simulació no recull. Simulem la latència a velocitats ràpides d’API i el cost amb els preus habituals de l’API.
2. Tots els models s’avaluen amb l’entorn de proves de l’API d’ExploitBench, amb 5 llavors i continuïtat de raonament.
3. Vam executar ExploitGym a la nostra API alfa, que genera respostes més de pressa que la nostra API pública, i després ho vam reescalar perquè coincidís amb la nostra API pública. En reescalar les latències a les velocitats esperades per a la nostra API pública, algunes latències estimades superen els límits de temps de 2 h i 6 h, tot i que es van respectar correctament en l’execució de l’avaluació. Per obtenir més velocitat en treballs sensibles al temps, oferim processament prioritari a l’API i mode ràpid a Codex.
4. Els models sense segments de sortida, latència o cost comunicats es representen com a línies de punts horitzontals.
