GPT-5.6 Sol, Terra, and Luna Are Out. Here Is How They Benchmark Against Opus 4.8 and Fable 5, and When to Use Each.

TL;DR

As of this week, GPT-5.6 is fully out. The three-model family, Sol, Terra, and Luna, is generally available after a month behind a government review gate. The benchmark picture is not a clean sweep for anyone. GPT-5.6 tops Terminal-Bench, the agentic terminal-coding board OpenAI leans on, with Sol Ultra at 91.9%. Flip to SWE-bench and it inverts: Anthropic's Fable 5 and Opus 4.8 lead the repo-patch boards, and OpenAI did not publish a SWE-bench number for 5.6 at all. Add price and context and it becomes a routing decision: Luna is roughly 10 times cheaper than Fable, GPT-5.6 carries a context near 1.5M tokens, and the right model depends entirely on the job. Here are the numbers, and a simple way to pick.

The lineup, briefly

GPT-5.6 ships as three tiers so you buy your point on the cost-capability curve instead of paying flagship rates for everything. Sol is the flagship, with the strongest agentic coding and both "max" and "ultra" reasoning modes, and it is the one that spent June gated. Terra is the balanced everyday model at roughly half Sol's price. Luna is the fast, cheap tier for high-volume work. Output prices run $30, $15, and $6 per million tokens in that order, input $5, $2.50, and $1, and the context window sits near 1.5M tokens. If you want the full story of why Sol was locked up for a month, we covered the government review that gated it separately. This piece is about what the models actually do now that anyone can call them.

The benchmark story depends on which board you read

Start with the board OpenAI puts front and center. Terminal-Bench measures whether a model can drive a real terminal, run bash, and finish multi-step agentic tasks end to end, and GPT-5.6 owns it. Sol Ultra hits 91.9%, plain Sol 88.8%, and even Terra and Luna land around 84%, ahead of Anthropic's best. This is a genuine strength, not a rounding error, and it matches the product: GPT-5.6 was tuned hard for agentic, tool-driving coding.

Now flip to SWE-bench, which grades whether a model can resolve real GitHub issues by patching a real repository, and the winner changes. Opus 4.8 posts 88.6% on SWE-bench Verified and 69.2% on the harder, less-saturated SWE-bench Pro. Fable 5 goes further on the tougher board, around 80.3% on SWE-bench Pro, and tops the coding leaderboards elsewhere, including 89.78% on LiveCodeBench. GPT-5.6's number on either SWE-bench? Not published. For a model this heavily marketed on coding, a missing SWE-bench score is not an oversight, it is a choice, and it tells you which board OpenAI would rather you look at.

The lesson is not that one lab is lying. It is that benchmark choice is marketing, the way a restaurant quotes the one review category it wins. Terminal-Bench and SWE-bench measure different jobs. If your work is an agent driving a shell, GPT-5.6 has the edge. If it is autonomous multi-file patches against a large codebase, the Claude models are still the ones to beat.

Reasoning and the rest of the stack

Up at the frontier, the pure-reasoning boards are close and largely saturated. Opus 4.8 scores 93.6% on GPQA Diamond, a graduate-level science test that is now so near its ceiling that a point or two means little. Fable 5 leads the harder frontier probes, 59.0% on Humanity's Last Exam without tools and 29.3% on FrontierCode Diamond where GPT-5.5 managed only 5.7%. GPT-5.6 is expected to push FrontierMath past 40% from GPT-5.5's 35.4%, though OpenAI has not posted a confirmed figure. Translation: at the top, everyone is excellent and the leaderboards no longer separate them cleanly. Your own eval on your own tasks will tell you more than any of these decimals.

Price and context, the quiet deciders

Capability is half the decision. The other half is what it costs and how much it can hold. Here GPT-5.6 is aggressive. Luna runs $1 in and $6 out per million tokens, roughly a tenth of Fable 5's $10 and $50, and Terra sits in the middle at $2.50 and $15. Opus 4.8 is priced at $5 and $25 with an optional Fast mode at $10 and $50 for about 2.5 times the speed. On context, GPT-5.6's roughly 1.5M-token window edges the Claude 5 family's 1M. So for high-volume classification, long-document work, or anything where you are paying by the million tokens, GPT-5.6's economics are hard to argue with. For the gnarliest autonomous repo work, Fable's premium can still earn its keep.

So which do you actually use

Forget the leaderboards for a second and use three dials. The model is the brain. The effort is how long it is allowed to think. The speed setting, Fast, just pays extra credits to get the same brain's answer sooner, it does not make it smarter. Set them independently for the job.

On the model dial, OpenAI's Codex guidance lines up cleanly with the tiers. Reach for Sol as the default for hard or important work: architecture, complex bugs, multi-file changes, research, security, anything you want polished. If you are unsure, start there. Drop to Terra for the everyday workhorse jobs, normal implementation, code review, codebase scans, where it is cheaper and faster with strong tool use. Use Luna for clear, repeatable tasks: extraction, summaries, simple edits, classification, the "do this exact thing" work. The older 5.5 is still a strong previous frontier model if you have a workflow tuned to it, 5.4 suits pinned or legacy setups, and 5.4 Mini is cheap and fast for small chores and subagents but not for gnarly debugging.

On the effort dial, keep Medium as your default, which is also what OpenAI's own guidance recommends as the balance point. Go Low for quick, scoped, low-risk asks. Push to High or Extra High for multi-step work, debugging, edge cases, reviews, and migrations where you cannot afford to miss anything. Save Max for the hardest single-agent reasoning where depth beats latency, and Ultra for big divisible work where subagents help, like broad audits or parallel implementation. On the speed dial, Fast mode is worth it when you are sitting there waiting on an interactive back-and-forth, and wasteful on long autonomous runs where nobody is watching the cursor.

Practical defaults that cover most days: serious task, Sol plus Medium or High. Normal coding, Terra plus Medium. Quick cheap task, Luna, or 5.4 Mini on Low. Important hard audit or debug, Sol plus Extra High or Max. Need it now, add Fast and expect the credits to burn faster.

The honest caveats

Three things to hold onto. First, every number here is vendor-run or aggregator-collected, and the labs pick the boards they win. Sol leading the benchmark OpenAI optimized for, while 5.6 ships no SWE-bench score at all, is the clearest tell in the set. Second, the aggregators disagree at the edges, with Fable 5's SWE-bench Verified reported anywhere from the high 80s to the mid 90s depending on who ran it, so treat any single decimal with suspicion and trust your own eval on your own codebase. Third, "available" has been a moving target all summer: GPT-5.6 was gated then cleared, and Fable 5 was itself suspended in mid-June and relaunched in early July after the same kind of government review. The models are out now, but this has been the season of frontier launches that arrive, vanish, and come back.

Key Takeaways

• GPT-5.6 (Sol, Terra, Luna) is generally available, and it leads Terminal-Bench 2.1 for agentic terminal and bash coding, with Sol Ultra at 91.9% and Terra and Luna near 84%.
• On SWE-bench, the story flips: Opus 4.8 hits 88.6% Verified and 69.2% Pro, Fable 5 leads the harder Pro board near 80.3% and the coding leaderboards, and OpenAI published no SWE-bench number for 5.6.
• Pure-reasoning boards are saturated (Opus 4.8 at 93.6% GPQA), so real separation now lives in agentic and long-horizon coding, where your own eval beats any leaderboard.
• Price and context favor GPT-5.6 for volume: Luna at $1/$6 is about 10x cheaper than Fable 5's $10/$50, and GPT-5.6's ~1.5M context edges the Claude 5 family's 1M. Opus 4.8 sits at $5/$25.
• Pick with three dials: model is the brain (Sol for hard work, Terra everyday, Luna for cheap repeatable tasks), effort is think time (Medium default, High or Max when it matters), and Fast just buys speed, not smarts.

Sources: OpenAI (GPT-5.6), OpenAI Codex model guide and speed guide (model, effort, and Fast recommendations), ExplainX (GPT-5.6 vs Fable 5 benchmarks), LLM-Stats (Opus 4.8 SWE-bench Verified and GPQA), Morph (SWE-bench Pro leaderboard). Some GPT-5.6 figures are vendor-reported or estimated pending full benchmark publication.