GPT-5.6 (Luna, Terra, Sol): Pricing, Benchmarks & Local Alternatives

Quick answer

If you came here asking "which GPT-5.6 should I use?" — Sol for agentic coding and complex reasoning, Terra if you were happy with GPT-5.5 and want the same class of output for half the money, Luna for high-volume cheap-and-fast calls. If you came here asking "can I download it?" — no, and no amount of quantization changes that. Skip to the local section for what you can actually run.

One release detail worth knowing: a June 2026 US executive order requires powerful models to be submitted for government review 30 days before release. The Commerce Department's Center for AI Standards and Innovation cleared GPT-5.6 in about two weeks, and government-approved partners had a limited preview from June 26, 2026 via the API and Codex before the public launch.

Luna vs Terra vs Sol

In ChatGPT itself (per secondary reporting): no GPT-5.6 access on Free or Go plans; Plus gets Medium/High reasoning levels; Pro, Business, and Enterprise get Medium/High/Extra High/Pro. GPT-5.5 Instant remains the everyday default layer.

Pricing

Every figure below is input / output in US dollars per 1M tokens, taken from OpenAI's published rate card. The one number most write-ups miss: pricing splits at 272K context, so a long-context agent run does not cost the sticker rate.

Three things fall out of that table. First, the new flagship is cheaper than the old one: Sol at $4 / $20 undercuts GPT-5.5 at $5 / $30, which is not how flagship launches usually go — OpenAI describes Sol's rate as promotional and available at least through November 21, 2026, so treat it as a floor that may lift. Second, Luna is not merely cheaper but a different order of magnitude: at $0.20 / $1.20 it costs about 4% of GPT-5.5's input rate, which changes what is economic to automate rather than just trimming a bill. Third, crossing 272K context roughly doubles your rate, so a long-context agent loop can quietly cost double what a spreadsheet built on sticker prices predicted.

Context window is reported (secondary sources) at ~1,050,000 tokens with 128K max output. Batch trades latency for half price and suits overnight jobs; Fast mode roughly doubles the standard rate.

Benchmarks (reported)

We have not run GPT-5.6 through our own harness, so everything here is labeled by source. The headline third-party number: on Artificial Analysis's Coding Agent Index, Sol scores 80 — reported as 2.8 points above Claude Fable 5, while using under half the output tokens, under half the time, and about a third less cost. Terra reportedly lands slightly above Fable 5; Luna above Claude Opus 4.8.

OpenAI-cited GPT-5.6 vs GPT-5.5 deltas, as compiled by secondary sources:

Read those security-benchmark jumps (ExploitBench 2, SEC-Bench Pro) as vendor-cited until independently reproduced. The token-efficiency story is the more defensible takeaway: both OpenAI's 54% claim and Artificial Analysis's under-half-the-tokens observation point the same direction. For crowd-voted head-to-heads rather than vendor numbers, see where Sol ranks on the Arena (LMArena) leaderboard.

Can you run GPT-5.6 locally?

No. Luna, Terra, and Sol are all closed-weight, same as GPT-5.5. There is no download, no GGUF, no self-host tier. Any site telling you how to "install GPT-5.6 with Ollama" is describing a different model wearing the name.

What OpenAI does let you run locally — unchanged since August 2025 — is the gpt-oss pair, both Apache 2.0, on Hugging Face, and runnable via Ollama, vLLM, Transformers, or llama.cpp:

ollama run gpt-oss:20b   # 16GB machines
ollama run gpt-oss:120b  # single 80GB GPU

And no, there is no "gpt-oss 2" — that's rumor-level at best as of this writing.

API cost vs local hardware

Sol's $30 per million output tokens is where agent workloads bite: agent loops are output-heavy (tool results, intermediate reasoning), so the meter runs on the expensive side of the pricing table — and above 272K input the reported surcharge takes Sol to $10/$45. That per-token bill never stops. A local GPU is the opposite trade: one-time hardware cost, zero per-token fees, but you get gpt-oss-class or open-frontier quality, not Sol.

The honest framing: if you need Sol-level agentic coding, pay for Sol (or Terra, which cuts the rate in half for GPT-5.5-class output). If your workload fits what a 16GB machine or a single 80GB GPU can serve, local wins on cost over time. We walk through the break-even math in Ollama vs the ChatGPT API: the real cost math and the broader trade-offs in Local AI vs ChatGPT.

Open-weight alternatives

Nothing open-weight matches Sol today. But if closed weights are a dealbreaker, these are the models actually chasing this tier:

• →GLM-5.2 — MIT-licensed open-weight frontier MoE.
• →DeepSeek V4 — the other MIT-licensed open-weight frontier line.
• →Kimi K3 — Moonshot's open-weight line.
• →Inkling (Thinking Machines Lab, July 15-16, 2026) — 975B-param Apache 2.0 MoE, 1M context. Its makers claim it's competitive with DeepSeek V4, GLM-5.2, and Kimi K2.6 while trailing the closed frontier. Needs ~2TB GPU memory at BF16 (~8× B300 or 16× H200; roughly half with NVFP4); supports vLLM, SGLang, and llama.cpp. Datacenter-class only.

Bottom line

Sol is the reported flagship of GPT-5.6: 80 on the Coding Agent Index — 2.8 points above Claude Fable 5 at about a third less cost — with token efficiency as the most defensible takeaway. Terra delivers GPT-5.5-class performance (per OpenAI) at 40% of the price, and Luna, at $0.20/$1.20 per Mtok, reportedly still scores above Claude Opus 4.8.

For local-first readers the verdict is unchanged from GPT-5.5: all three models are closed-weight, so if you need Sol-level agentic coding, pay for Sol or Terra. If closed weights are a dealbreaker, nothing open matches Sol today — GLM-5.2, DeepSeek V4, and Inkling chase this tier from behind, and on modest hardware the gpt-oss pair remains the realistic local pick.

FAQ