GPT-5.6 Sol, Terra, & Luna

GPT-5.6 is a three-tier model line OpenAI is previewing as Sol (flagship), Terra (the workhorse), and Luna (the cheap one). The number is the generation; Sol, Terra, and Luna are durable tiers that advance on their own cadence. It’s OpenAI’s strongest release to date, it beats Claude Mythos 5 on the one coding benchmark OpenAI chose to publish, and it shipped under a US-government access gate that OpenAI itself calls “unsustainable.”

What shipped

Model: GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna, in the API as gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. Sol also runs in two heavier modes, “max” (deeper single-model reasoning) and “ultra” (parallel sub-agents), with “Sol Ultra” reported as a separate benchmark line.

Model type: Text + image input, text output, tuned hard for agentic coding, cybersecurity, and biology.

Pricing: Per million tokens, Sol is $5 input / $30 output, Terra is $2.50 / $15, and Luna is $1 / $6. Sol holds the same rate card as GPT-5.5 ($5 / $30), and Terra is pitched as GPT-5.5-class capability at half the price.

Available on: During the preview, the OpenAI API and Codex only, restricted to roughly 20 organizations whose participation was cleared with the US government. Broad access through ChatGPT, Codex, and the API is gated behind that same government review, with no firm date.

Headline benchmarks: On Terminal-Bench 2.1 (command-line agentic coding), Sol Ultra scores 91.9% and Sol scores 88.8%, against Claude Mythos 5 at 88.0%, Terra and Claude Fable 5 tied at 84.3%, and GPT-5.5 at 83.4%. On ExploitBench (offensive cyber), Sol is competitive with Claude Mythos while spending roughly a third of the output tokens, though Mythos 5 still leads at around 80% and Sol stops short of full autonomous exploit generation in real browsers. On health, HealthBench Professional is 60.5 (+8.7 over GPT-5.5), HealthBench 57.0, HealthBench Hard 33.1.

Other info: OpenAI’s model pages list a 1.05M-token context window for all three (922K input, 128K output) and a February 16, 2026 knowledge cutoff. On OpenAI’s Preparedness Framework, all three models are rated High for Cybersecurity and High for Biological & Chemical, and Below High for AI Self-Improvement. The system card reports Sol can find vulnerabilities and exploit pieces but “were unable to carry out autonomous, end-to-end attacks against hardened targets,” and flags that Sol shows “a greater tendency than GPT-5.5 to go beyond the user’s intent” in internal coding traffic. The gate traces to a Trump-administration executive order asking frontier labs to submit advanced models for review up to 30 days before release, the same mechanism that reportedly forced Anthropic to pull Mythos-class access weeks earlier.

More details: Previewing GPT-5.6 Sol, GPT-5.6 Preview System Card

OpenAI previewed a three-tier generation instead of a single flagship, and that’s the structural news as much as the capability is. Sol is the model for the hardest work (complex coding, security research, deep reasoning), Terra is the high-volume business tier (support, internal tools, document analysis) priced at half of Sol, and Luna is the fast, cheap everyday tier for summarizing, drafting, and routine automation. The naming decouples the generation number from the capability tier on purpose: 5.6 is the family, and Sol/Terra/Luna are positions OpenAI says it will keep advancing independently, the same product split Anthropic implies with Opus/Sonnet/Haiku. Sol also ships with two extra gears, a “max” mode for deeper single-model reasoning and an “ultra” mode that fans a task out to sub-agents running in parallel (which is where the 91.9% Terminal-Bench number comes from).

The benchmarks

OpenAI’s numbers

Every score is read from the chart data in OpenAI’s preview post. Rows show GPT-5.6 Sol, with Terra and Luna listed beside it, and GPT-5.5 at its highest setting.

• Terminal-Bench 2.1 (Sol at max): 88.8%. +4.5 over GPT-5.6 Terra at max (84.3%), +6.3 over GPT-5.6 Luna at max (82.5%), +3.2 over GPT-5.5 at xhigh (85.6%). Sol Ultra, which splits the task across sub-agents, scores 91.9%.

• DeepSWE v1.1 (Sol at max): 72.7%. +3.1 over GPT-5.6 Terra at max (69.6%), +5.5 over GPT-5.6 Luna at max (67.2%), +5.7 over GPT-5.5 at xhigh (67.0%), +2.8 over Claude Fable 5 at xhigh (69.9%), +13.7 over Claude Opus 4.8 at max (59.0%).

• BrowseComp (Sol at max): 90.4%. +2.9 over GPT-5.6 Terra at max (87.5%), +6.4 over GPT-5.6 Luna at xhigh (84.0%), +6.0 over GPT-5.5 at xhigh (84.4%). Sol Ultra scores 92.2%.

• SEC-Bench Pro (Sol at max): 71.2%. +13.5 over GPT-5.6 Terra at max (57.7%), +22.3 over GPT-5.6 Luna at max (48.9%), +25.4 over GPT-5.5 at xhigh (45.8%). Sol Ultra scores 74.3%.