OpenAI launches GPT-5.6 models with new coding push

OpenAI has launched the GPT-5.6 family of artificial intelligence models for general availability, with three tiers: Sol, Terra and Luna.

The range is available across ChatGPT, Codex and the OpenAI API. OpenAI positions Sol as its flagship model, Terra as a lower-cost option for everyday use and Luna as its cheapest offering.

The launch also introduces an "ultra" setting for more demanding tasks. In this mode, four agents work in parallel by default, while a "max" setting gives the models more time to reason, run checks and revise their approach.

API pricing is set per one million tokens. Sol costs USD $5 for input and USD $30 for output, Terra costs USD $2.50 for input and USD $15 for output, and Luna costs USD $1 for input and USD $6 for output.

GPT-5.6 Sol scored 53.6 on Agents' Last Exam, which OpenAI described as an evaluation of long-running professional workflows across 55 fields. According to OpenAI, that put Sol 13.1 points ahead of Claude Fable 5 in adaptive reasoning, and 11.4 points ahead in medium reasoning at about one-quarter of the estimated cost.

On the Artificial Analysis Intelligence Index, OpenAI said GPT-5.6 Sol with maximum reasoning came within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost. It added that Terra and Luna outperformed Fable 5 at around one-sixteenth of the cost in some tests.

Coding focus

Software development is central to the launch. OpenAI said GPT-5.6 Sol scored 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time and costing about one-third less.

OpenAI also said the model set posted leading results on Terminal-Bench 2.1 and DeepSWE, which test command-line workflows and engineering work in codebases. Terra performed just above Fable 5 in the same coding index, while Luna outperformed Opus 4.8, according to OpenAI.

A new feature called Programmatic Tool Calling in the Responses API is intended to let the models write and run lightweight programs that coordinate tools, process intermediate results and choose the next action. OpenAI said this should reduce the need to pass every tool response back through the model.

Knowledge work

OpenAI also made a broader pitch for office and professional use, saying GPT-5.6 can turn material from documents and workplace tools such as Slack, Notion, Microsoft 365 and Google Drive into presentations, documents and spreadsheets.

Sol achieved 92.2% on BrowseComp and 62.6% on OSWorld 2.0, according to OpenAI, with the OSWorld result exceeding Opus 4.8 while using 85% fewer output tokens. OpenAI added that Luna nearly matched GPT-5.5's peak performance at less than half the estimated cost, while Terra surpassed it at a lower cost.

OpenAI said the new model can create editable presentations from scratch and infer a slide deck's design system, including layouts, typography, spacing and recurring content patterns. It also said GPT-5.6 improves work on documents and spreadsheets, including equations and financial models.

Cybersecurity and science

OpenAI described GPT-5.6 as its strongest cybersecurity model so far. On ExploitBench 2, the company said the model scored 73.5% against GPT-5.5's 47.9% at a comparable output-token budget.

On ExploitGym 3, the pass rate rose from 15.1% for GPT-5.5 to 24.9% under a two-hour cap, reaching 33.7% with six hours, OpenAI said. On SEC-Bench Pro, GPT-5.6 scored 71.2% against GPT-5.5's 45.8% with improved latency, according to the company.

Access to more sensitive cyber functions will remain restricted. OpenAI said qualified individuals and organisations in its Trusted Access for Cyber programme can obtain broader access for authorised defensive work, including vulnerability triage, malware analysis, detection engineering and patch validation.

Users in that programme will need Advanced Account Security with hardware-backed passkeys to retain access to its most cyber-focused frontier models, OpenAI said. The company also said it is taking additional steps to restrict access for high-risk entities and in high-risk jurisdictions.

In life sciences, OpenAI said GPT-5.6 showed gains over GPT-5.5 across biology, life-science research workflows and chemistry. It added that the model did not cross what it called the Critical threshold in either biology or cybersecurity.

Safety controls

OpenAI said GPT-5.6 launches with its most extensive safeguards so far, combining protections built into the model with real-time checks, monitoring and account-level enforcement. The company said it used about 700,000 A100e GPU hours on black-box automated red teaming before general availability.

It added that a reasoning monitor reviews conversations for potential harm, rather than relying only on classifier flags. OpenAI said its GPT-5.6 Sol cyber safeguards block roughly 10 times more potentially harmful activity than previous models, while allowing users to retry prompts on lower-capability models in ChatGPT and Codex if needed.

Internal use was another part of OpenAI's case for the new model. The company said researchers used GPT-5.6 across debugging, training system optimisation, experiment running and result interpretation, with average daily output tokens per active researcher more than double the highest level observed for GPT-5.5 during internal testing.

"There is no such thing as perfect security, and our work to secure increasingly capable models continues. New weaknesses will be discovered, as will new jailbreaks that circumvent existing safeguards. Each new generation of model will also create new avenues for attack and misuse," OpenAI said.