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(magbubukas sa bagong window). 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(magbubukas sa bagong window)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(magbubukas sa bagong window).
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.
Kailangan ding manatiling epektibo ang mga safeguard kapag binabago ng mga attacker ang kanilang mga taktika. Hindi sapat ang tibay ng proteksiyong gumagana lamang sa isang nakapirming hanay ng mga kilalang pag-atake para sa isang frontier na modelo.
Kaya mas maraming intelligence at compute kaysa dati ang inilalapat namin sa safety, gamit ang sarili naming mga modelo para mas mabilis na makahanap ng kahinaan at mapahusay ang mga safeguard. Naglaan kami ng mahigit 700,000 A100-equivalent GPU hours sa automated red teaming na naglalayong makahanap ng mga universal jailbreak: mga pag-atakeng maaaring gumana sa maraming prompt o konteksto, hindi lang sa isang makitid na setting. Sa pagtutok sa mas mahihirap at mas pangkalahatang pag-atakeng ito, nasubok namin ang mga safeguard nang lampas sa nakapirming hanay ng mga kilalang failure. Nagbibigay-daan din ito sa amin na tuklasin ang mas maraming pattern ng pag-atake kaysa kayang saklawin ng human testing lamang, mas maagang matukoy ang mga pattern ng failure, at paikliin ang proseso mula sa paghahanap ng kahinaan hanggang sa pagtugon dito.
Bukod sa automated red-teaming, nakipagtulungan kami sa mga third-party tester para magsagawa ng malawakang human expert red teaming, na magpapatuloy sa panahon ng preview. Kinukumpleto ng human red-teaming ang automated na gawain sa pamamagitan ng pagsubok sa mga safeguard laban sa malikhaing mga eksperto na sumusubok abusuhin ang modelo sa mga paraang maaaring hindi mahulaan ng aming mga system.
Walang evaluation ang makakatawan sa bawat configuration ng produkto, multi-step na pag-atake, o real-world workflow. Kaya nagpapanatili kami ng mabilisang proseso para i-reproduce, tasahin, unahin, at ayusin ang mga bagong natuklasang jailbreak, pagkatapos ay idagdag ang mga ito sa aming patuloy na evaluations upang masubok namin ang katulad na mga failure sa hinaharap.
Sa panahon ng preview, ang mga modelo ng GPT‑5.6 ay unang magiging available sa pamamagitan ng API at Codex sa piling grupo ng mga pinagkakatiwalaang partner at organisasyon. Plano naming gawing mas malawak na available ang mga ito sa mga taong gumagamit ng ChatGPT, Codex, at API sa lalong madaling panahon.
Sa bagong naming system na ito na ipinakilala sa GPT‑5.6, tinutukoy ng numero ang henerasyon ng modelo, habang tinutukoy ng Sol, Terra, at Luna ang matitibay na tier ng capability na maaaring umusad sa sarili nilang bilis. Bilang isang pamilya, nagbibigay ito sa mga tao at developer ng mas malinaw na pagpipilian sa intelligence, bilis, at gastos.
Ang GPT‑5.6 ay pinipresyuhan kada 1M token sa tatlong laki ng modelo: Sol ay $5 input / $30 output; Terra ay $2.50 input / $15 output; at Luna ay $1 input / $6 output. Ipinapakilala rin ng GPT‑5.6 ang mas predictable na prompt caching, kabilang ang suporta para sa explicit cache breakpoints at 30-minutong minimum cache life. Para sa GPT‑5.6 at mga susunod na modelo, sinisingil ang cache writes sa 1.25x ng uncached input rate ng modelo, habang patuloy na tumatanggap ang cache reads ng 90% cached-input discount.
Ilulunsad din namin ang GPT‑5.6 Sol sa Cerebras sa bilis na hanggang 750 token kada segundo sa Hulyo, na nagdadala ng frontier intelligence sa mga customer sa walang kapantay na bilis. Sa simula, magiging limitado ang access sa piling customer habang pinapalawak namin ang kapasidad.
Nasasabik kaming patuloy na matuto mula sa panahon ng preview na ito, at dalhin ang GPT‑5.6 Sol, Terra at Luna sa mas maraming tao sa lalong madaling panahon.
1. Tinatantiya namin ang latency at gastos sa API sa pamamagitan ng pagtingin sa production behavior ng aming mga modelo, at pag-simulate offline. Isinasaalang-alang ng mga tantiyang ito ang mga detalye ng tool call, sampled tokens, at input tokens. Maaaring malaki ang ibahin ng real-world results, at nakadepende ang mga ito sa maraming salik na hindi nakuha sa aming simulation. Sine-simulate namin ang latency sa mabilis na API speeds, at ang gastos sa regular na API pricing.
2. Sinusuri ang lahat ng modelo gamit ang ExploitBench API harness na may 5 seed at reasoning continuity.
3. Pinatakbo namin ang ExploitGym sa aming alpha API, na naglalabas ng mga tugon nang mas mabilis kaysa sa aming public API, at pagkatapos ay ni-rescale upang tumugma sa aming public API. Kapag ni-rescale ang mga latency sa bilis na inaasahan para sa aming public API, nagdudulot ito na lumampas ang ilang tinantiyang latency sa 2h at 6h hour time limits, kahit tama itong nasunod sa evaluation run. Para makakuha ng mas mabilis na bilis para sa time-sensitive na trabaho, nag-aalok kami ng priority processing sa API at fast mode sa Codex.
4. Ang mga modelong walang iniulat na output token, latency, o gastos ay ipinapakita bilang mga horizontal dotted line.
