OpenAI Launches GPT-5.6 Family: Sol, Terra, and Luna Reach General Availability
OpenAI launched the GPT-5.6 family for general availability on July 9, 2026 — Sol (flagship), Terra (balanced), and Luna (fast and affordable) — following a two-week limited preview that began June 26 under a government safety review. The models post state-of-the-art results in coding, cybersecurity, and long-horizon knowledge work while using markedly fewer tokens than their predecessor, GPT-5.5, and OpenAI says the rollout will reach full availability across ChatGPT, Codex, and the API within 24 hours.
Intermediate
A new naming system: durable tiers, not just version bumps
GPT-5.6 introduces a naming scheme OpenAI plans to keep going forward: the number marks the generation, while Sol, Terra, and Luna are “durable capability tiers” that can each advance on their own cadence. Sol is the flagship, Terra is pitched as competitive with GPT-5.5 at roughly half the cost, and Luna is the fastest and cheapest of the three. Pricing per 1M tokens is $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. OpenAI also introduced explicit prompt-cache breakpoints with a 30-minute minimum cache life, though cache writes now cost 1.25x the uncached input rate.
The release also debuts two new ways to scale reasoning at inference time: a max effort setting that gives Sol more time to explore and revise its answer, and ultra, which coordinates four agents in parallel by default (configurable up to 16) to split up complex tasks. On evaluations like BrowseComp and Terminal-Bench 2.1, OpenAI’s charts show that adding parallel agents shifts the score-versus-latency curve up and to the left — better results, delivered faster — though at higher token cost.
Benchmark results
OpenAI’s own comparisons position Sol against Anthropic’s Claude Fable 5, Claude Opus 4.8, and Google’s Gemini 3.1 Pro. On Agents’ Last Exam, a test of long-running professional workflows across 55 fields, Sol scores 52.7% against Fable 5’s 40.5%. On the Artificial Analysis Coding Agent Index, Sol with max reasoning hits 80 (a new state of the art), edging out Fable 5’s 77.2 while using less than half the output tokens and about one-third less estimated cost. Sol also leads on Terminal-Bench 2.1 (88.8%, rising to 91.9% with ultra) and posts a new state of the art on BrowseComp agentic-browsing tasks at 90.4% (92.2% with ultra).
The cybersecurity gains are the sharpest jump in the release. On ExploitBench, which measures progress from a vulnerable code sample to arbitrary code execution, Sol scores 73.5% versus GPT-5.5’s 47.9% at a comparable token budget. On ExploitGym, a benchmark built with UC Berkeley researchers that asks agents to turn real vulnerabilities into working exploits, Sol nearly doubles GPT-5.5’s peak pass rate — from 15.1% to 24.9% under a two-hour cap, reaching 33.7% with six hours. OpenAI still says Sol does not cross the “Cyber Critical” threshold in its Preparedness Framework: in testing against Chromium and Firefox, it found bugs and exploitation primitives but did not autonomously chain them into a functional full exploit.
Safety review and the government preview window
Rather than launching directly to general availability, OpenAI first shipped GPT-5.6 to a small group of trusted partners on June 26, coordinated with a U.S. government safety review requested under a June executive order asking major AI developers to voluntarily submit frontier models for evaluation before release. OpenAI has said publicly it does not want this kind of government gating to become a permanent step, framing it as a short-term measure while it works with the administration on a repeatable review process for future launches. The review cleared faster than OpenAI initially expected, and general availability began July 9 with a rollout OpenAI says will reach full availability over 24 hours.
To back the higher-capability cyber release, OpenAI says it dedicated roughly 700,000 A100-equivalent GPU hours to automated red-teaming aimed at finding “universal jailbreaks” — attacks that generalize across prompts rather than working only in one narrow case — on top of continued human expert red-teaming. The company reports that its GPT-5.6 Sol cyber safeguards now block roughly ten times more potentially harmful activity than prior models, with a reasoning-based monitor reviewing flagged conversations in addition to real-time classifiers. Verified security professionals can still get expanded access to Sol’s defensive capabilities — vulnerability triage, malware analysis, patch validation — through OpenAI’s Trusted Access for Cyber program, though individual members must enable hardware-backed passkey security by September 1 to keep that access.
What this means
The government preview step is notable in its own right: it follows reporting that Anthropic’s Claude Mythos cybersecurity model drew government concern over dual-use risk, and OpenAI’s framing suggests this kind of pre-release check could become more common industry-wide even as OpenAI argues against making it a permanent default. On capability, the headline claim — matching or beating a competing frontier model while using a fraction of the tokens and cost — is consistent with the broader trend RITS has tracked this year: newer frontier releases increasingly compete on efficiency and cost-per-task rather than raw benchmark ceiling alone, following patterns seen with GLM-5.2 and other recent open and closed models.
Related Coverage
- OpenAI Releases GPT-5.5: Agentic Coding Ceiling Tops 14 Benchmarks — the predecessor model GPT-5.6 replaces.
- OpenAI Opens GPT-5.5-Cyber to Vetted Defenders via Trusted Access — background on the Trusted Access for Cyber program GPT-5.6 extends.
- Anthropic Launches Claude Fable 5, Its First Public Mythos-Class Model — the model OpenAI benchmarks GPT-5.6 Sol against throughout its release.
沪公网安备31011502017015号