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(dibuka dalam tetingkap baru). 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(dibuka dalam tetingkap baru)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(dibuka dalam tetingkap baru).
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.
Langkah perlindungan juga perlu kekal berkesan apabila penyerang menyesuaikan taktik mereka. Perlindungan yang hanya berfungsi pada set serangan diketahui yang tetap tidak cukup teguh untuk model garis hadapan.
Sebab itu kami menggunakan lebih banyak kecerdasan dan komputasi berbanding sebelum ini untuk keselamatan, menggunakan model kami sendiri bagi mencari kelemahan dan memperbaiki perlindungan dengan lebih pantas. Kami memperuntukkan lebih 700,000 jam GPU setara A100 untuk red teaming automatik yang bertujuan mencari jailbreak sejagat: serangan yang boleh berfungsi merentas banyak prom atau konteks, bukan hanya satu tetapan yang sempit. Dengan menumpukan pada serangan yang lebih sukar dan lebih umum ini, kami dapat menguji perlindungan melampaui set kegagalan diketahui yang tetap. Ini juga membolehkan kami meneroka jauh lebih banyak corak serangan berbanding ujian manusia semata-mata, mengenal pasti corak kegagalan lebih awal, dan memendekkan laluan daripada menemui kelemahan kepada menanganinya.
Selain proses penyerang automatik, kami bekerjasama dengan penguji pihak ketiga untuk menjalankan proses penyerang pakar manusia secara meluas, yang akan diteruskan sepanjang tempoh pratonton. Proses penyerang manusia melengkapi kerja automatik dengan menguji perlindungan terhadap pakar kreatif yang cuba menyalahgunakan model dengan cara yang mungkin tidak dijangka oleh sistem kami.
Tiada penilaian yang dapat mewakili setiap konfigurasi produk, serangan berbilang langkah, atau aliran kerja dunia sebenar. Oleh itu, kami mengekalkan proses respons pantas untuk menghasilkan semula, menilai, mengutamakan dan membaiki jailbreak yang baru ditemui, kemudian menambahkannya pada penilaian berterusan kami supaya kami boleh menguji kegagalan serupa pada masa depan.
Sepanjang pratonton, model GPT‑5.6 pada mulanya akan tersedia melalui API dan Codex kepada kumpulan terpilih rakan kongsi dan organisasi yang dipercayai. Kami merancang untuk menyediakannya dengan lebih meluas kepada pengguna ChatGPT, Codex dan API tidak lama lagi.
Dalam sistem penamaan baharu yang diperkenalkan bersama GPT‑5.6 ini, nombor mengenal pasti generasi model, manakala Sol, Terra dan Luna mengenal pasti tahap keupayaan tahan lama yang boleh maju mengikut rentak masing-masing. Secara bersama, keluarga ini memberi orang ramai dan pembangun pilihan yang lebih jelas merentas kecerdasan, kelajuan dan kos.
GPT‑5.6 dihargakan bagi setiap 1 juta token merentas tiga saiz model: Sol ialah $5 input / $30 output; Terra ialah $2.50 input / $15 output; dan Luna ialah $1 input / $6 output. GPT‑5.6 juga memperkenalkan caching prom yang lebih boleh dijangka, termasuk sokongan untuk titik henti cache eksplisit dan hayat cache minimum 30 minit. Untuk GPT‑5.6 dan model selepasnya, penulisan cache dibilkan pada 1.25x kadar input tanpa cache model, manakala bacaan cache terus menerima diskaun input bercache 90%.
Kami juga melancarkan GPT‑5.6 Sol di Cerebras pada kelajuan sehingga 750 token sesaat pada bulan Julai, membawa kecerdasan garis hadapan kepada pelanggan pada kelajuan yang belum pernah dicapai. Akses pada mulanya akan dihadkan kepada pelanggan terpilih sementara kami mengembangkan kapasiti.
Kami teruja untuk terus belajar daripada tempoh pratonton ini, dan membawa GPT‑5.6 Sol, Terra dan Luna kepada lebih ramai orang tidak lama lagi.
1. Kami menganggarkan kependaman dan kos API dengan melihat tingkah laku produksi model kami, lalu membuat simulasi luar talian. Anggaran ini mengambil kira butiran panggilan alat, token yang disampel dan token input. Keputusan dunia sebenar mungkin berbeza dengan ketara, dan bergantung pada banyak faktor yang tidak dirangkum dalam simulasi kami. Kami mensimulasikan kependaman pada kelajuan API pantas, dan kos pada harga API biasa.
2. Semua model dinilai menggunakan harness API ExploitBench dengan 5 seed dan kesinambungan penaakulan.
3. Kami menjalankan ExploitGym pada API alfa kami, yang mengeluarkan respons lebih pantas daripada API awam kami, kemudian menskalakan semula agar sepadan dengan API awam kami. Apabila kependaman diskalakan semula kepada kelajuan yang dijangka untuk API awam kami, ini menyebabkan sesetengah anggaran kependaman melebihi had masa 2 jam dan 6 jam, walaupun dipatuhi dengan betul dalam larian penilaian. Untuk mendapatkan kelajuan lebih pantas bagi kerja sensitif masa, kami menawarkan pemprosesan keutamaan dalam API dan mod pantas dalam Codex.
4. Model tanpa token output, kependaman atau kos yang dilaporkan diplot sebagai garisan bertitik mendatar.
