https://artificialanalysis.ai/ - slopanalysis - comparison for all 5.6 models

GPT-5.6-model-selection-guide.md

GPT‑5.6 Sol, Terra, and Luna: a practical model-selection guide

Data snapshot: 11 July 2026
Primary source: Artificial Analysis GPT‑5.6 launch analysis and its live LLM leaderboard
Scope: all 15 combinations of Sol, Terra, and Luna at low, medium, high, xhigh, and max reasoning effort.

Short answer: Default to Luna for volume work and Sol when quality matters. Treat Terra as a dominated middle tier, not a default. Raise reasoning effort before moving to a larger model when the next effort remains on the better price/intelligence frontier. Use Sol max only where the last few quality points can change the outcome.

---

Executive summary

OpenAI has not created a simple “small / medium / large” ladder. It has created three overlapping ladders. A higher-effort Luna can meet or beat a lower-effort Terra; a higher-effort Terra can overlap lower-effort Sol; and, most importantly, Artificial Analysis finds that Terra is never on the Intelligence-vs-Cost Pareto frontier. At every Terra setting, Luna or Sol offers either more intelligence for no more money or similar intelligence for less.

That produces a much simpler practical policy than the 15 model/effort combinations suggest:

1. Luna is the economic default. Use it for triage, extraction, transformations, routine coding, subagents, and any workload run many times.
2. Sol is the quality default. Use it for difficult implementation, synthesis, consequential analysis, polished deliverables, and tasks where retries or mistakes are expensive.
3. Terra is a niche compatibility choice. Choose it only after your own workload test shows a specific latency, style, reliability, or provider constraint that the aggregate benchmark does not capture.
4. Reasoning effort is a real purchasing decision. It changes measured capability and per-task cost substantially. Do not set every call to max by habit.
5. Escalate by uncertainty and consequence, not by prompt length. Long but mechanical work may suit Luna; a short but ambiguous architectural decision may warrant Sol.

The data at a glance

Anchor measurements published by Artificial Analysis

Model	Reasoning	Intelligence Index	Cost per Intelligence Index task	Coding Agent Index	What it establishes
GPT‑5.6 Luna	max	51	$0.21	75	Strong low-cost ceiling; matches/exceeds GLM‑5.2 max and Gemini 3.5 Flash intelligence at lower cost
GPT‑5.6 Terra	max	55	$0.55	77	A 4-point gain over Luna max, but 2.6× the general benchmark cost
GPT‑5.6 Sol	max	59	$1.04	80	Near-frontier quality; one point behind Claude Fable 5 at about one-third of Fable’s cost

Exact ratios between the three max settings

Upgrade	Intelligence gain	Relative gain	Cost increase	Cost multiple	Marginal cost per extra Index point
Luna max → Terra max	+4	+7.8%	+$0.34	2.62×	$0.085 / point
Terra max → Sol max	+4	+7.3%	+$0.49	1.89×	$0.123 / point
Luna max → Sol max	+8	+15.7%	+$0.83	4.95×	$0.104 / point

The economic shape is stark: Luna max buys 86% of Sol max’s measured intelligence for about 20% of its task cost. Sol is not a value upgrade in the ordinary sense; it is a premium paid for the difficult final band of capability.

The complete all-effort dataset

The live leaderboard was filtered to exactly these 15 variants and no other models. These are the values displayed directly by Artificial Analysis—not estimates read from dot positions.

Model	Effort	Intelligence Index	Cost / Index task	Output tokens / task	Output speed	Decode time / task
Luna	low	33	$0.04	2k	204 tok/s	0.2 min
Luna	medium	38	$0.05	4k	195 tok/s	0.3 min
Luna	high	46	$0.09	8k	191 tok/s	0.7 min
Luna	xhigh	49	$0.14	12k	191 tok/s	1.0 min
Luna	max	51	$0.21	19k	206 tok/s	1.5 min
Terra	low	40	$0.10	2k	123 tok/s	0.3 min
Terra	medium	46	$0.13	4k	116 tok/s	0.5 min
Terra	high	49	$0.24	8k	120 tok/s	1.1 min
Terra	xhigh	52	$0.33	11k	136 tok/s	1.3 min
Terra	max	55	$0.55	19k	142 tok/s	2.2 min
Sol	low	49	$0.20	3k	64 tok/s	0.6 min
Sol	medium	54	$0.31	4k	61 tok/s	1.1 min
Sol	high	56	$0.45	7k	63 tok/s	1.8 min
Sol	xhigh	58	$0.68	10k	69 tok/s	2.4 min
Sol	max	59	$1.04	15k	74 tok/s	3.3 min

Output tokens / task and decode times are rounded as displayed by AA. Decode time excludes time to first token and other overhead. Output speed is the first-party API’s generation speed after streaming begins.

Do not miss the overlap:

• Luna high equals Terra medium at 46, for $0.09 instead of $0.13.
• Luna xhigh equals Terra high and Sol low at 49, for $0.14 instead of $0.24 or $0.20.
• Luna max reaches 51 for $0.21; Terra xhigh buys one more point for another $0.12.
• Sol medium scores 54 at $0.31, beating Terra xhigh at 52/$0.33.
• Sol high scores 56 at $0.45, beating Terra max at 55/$0.55.

Those comparisons explain why Luna and Sol form the frontier while Terra does not.

The displayed cost/intelligence frontier

Using the displayed integer Index scores and cent-rounded task costs, a point is dominated when another point is at least as intelligent and no more expensive.

Dominated setting	Better displayed alternative	Result
Terra low — 40 / $0.10	Luna high — 46 / $0.09	+6 points, save $0.01
Terra medium — 46 / $0.13	Luna high — 46 / $0.09	Same score, save 31%
Terra high — 49 / $0.24	Luna xhigh — 49 / $0.14	Same score, save 42%
Terra xhigh — 52 / $0.33	Sol medium — 54 / $0.31	+2 points, save $0.02
Terra max — 55 / $0.55	Sol high — 56 / $0.45	+1 point, save 18%
Sol low — 49 / $0.20	Luna xhigh — 49 / $0.14	Same displayed score, save 30%

The resulting displayed frontier is:

Luna low → Luna medium → Luna high → Luna xhigh → Luna max → Sol medium → Sol high → Sol xhigh → Sol max

Scores and costs are rounded for display, so equal-looking points may differ beneath the shown precision. That is a reason to A/B test close crossovers—not a reason to ignore a large visible price gap.

What each effort step buys

Step	Index gain	Task-cost increase	Marginal cost / displayed point
Luna low → medium	+5	+$0.01	$0.002
Luna medium → high	+8	+$0.04	$0.005
Luna high → xhigh	+3	+$0.05	$0.017
Luna xhigh → max	+2	+$0.07	$0.035
Terra low → medium	+6	+$0.03	$0.005
Terra medium → high	+3	+$0.11	$0.037
Terra high → xhigh	+3	+$0.09	$0.030
Terra xhigh → max	+3	+$0.22	$0.073
Sol low → medium	+5	+$0.11	$0.022
Sol medium → high	+2	+$0.14	$0.070
Sol high → xhigh	+2	+$0.23	$0.115
Sol xhigh → max	+1	+$0.36	$0.360

This is why max should not be the reflexive setting. Diminishing returns are especially severe at the top of Sol.

---

The breakpoints that actually matter

1. Luna max vs Sol low: the first confusing crossover

This initially looked like the important crossover, but the direct values resolve it: Luna xhigh and Sol low both display 49, while Luna costs $0.14/task against Sol’s $0.20. Luna is also about 3× faster in output generation (191 vs 64 tok/s), although it uses far more output tokens on the benchmark (12k vs 3k). Luna max scores 51 at $0.21, only one cent above Sol low.

Default from the aggregate numbers: Luna xhigh if 49 is sufficient; Luna max if it is not.
Test Sol low anyway when: concise outputs, lower token use, or stronger-base-model behaviour matters more than streamed speed and displayed cost efficiency. At displayed AA precision, Sol low is not on the cost/intelligence frontier.

2. Terra max vs Sol high: Terra’s clearest warning sign

Terra max scores 55 at $0.55/task. Sol high scores 56 at $0.45/task. Aggregate data therefore gives no economic reason to pick Terra max over Sol high.

Default decision: Sol high.
Exception: a measured workload-specific Terra advantage.

3. Luna max vs Terra max: modest quality, large price jump

Moving from Luna max to Terra max buys four Index points while multiplying cost by 2.62. That only makes sense if the workload has a threshold between those capability levels—e.g. Luna often fails a class of tasks that Terra reliably completes. If both still require review, the upgrade may buy little operational value.

4. Terra max vs Sol max: equal Index-point gain, steeper marginal price

The next four points cost another $0.49/task. Sol max is justified when those points affect acceptance rate, prevent expensive retries, or unlock tasks beneath other models. It is difficult to justify for ordinary throughput work.

5. Sol xhigh vs Sol max: the premium tail

There is exactly one displayed Index point between Sol xhigh and max, while task cost rises from $0.68 to $1.04: 53% more cost for a 1.7% Index increase. Decode time also rises from 2.4 to 3.3 minutes. This makes max a deliberate specialist setting, not a sensible universal default.

Use max for genuinely hard, long-horizon, consequential work. Use xhigh for most “best model” production calls unless your own evals show a meaningful max advantage.

---

Recommended routing policy

Workload	First choice	Escalate to	Why
Classification, routing, tagging	Luna low	Luna medium	Cheap; extra depth usually has low value
Extraction, reformatting, deterministic transformations	Luna low/medium	Luna high	Scale matters more than frontier reasoning
Summaries and straightforward research synthesis	Luna medium	Luna high	Good balance; reserve max for genuinely tangled sources
Routine code edits with clear acceptance tests	Luna high	Luna max or Sol medium	Verification catches errors; skip the visibly dominated Sol-low point by default
Repository exploration / subagents	Luna medium/high	Luna max or Sol medium	Many calls make Luna economics compelling
Difficult debugging	Sol medium	Sol high	Strong base model and additional reasoning both matter
Architecture and cross-system changes	Sol high	Sol xhigh/max	Errors propagate and review is expensive
High-stakes analysis or irreversible decisions	Sol xhigh	Sol max + human review	Pay for the tail, but do not mistake benchmark strength for certainty
Polished reports, slides, spreadsheets	Sol high/xhigh	Sol max	Sol max has the highest Presentation Elo measured in AA‑Briefcase
Long-horizon knowledge work	Sol xhigh/max	—	Sol is second to Fable 5 in AA‑Briefcase, though Fable still leads rubric and analytical quality
Coding-agent benchmark chasing	Sol max in Codex	—	Leads the AA Coding Agent Index at 80
Very high-volume mixed workload	Luna medium/high with escalation	Sol medium/high	Two-stage routing captures most value

A KISS three-tier setup

If 15 options are too many, expose only these internally:

• Fast: Luna medium
• Standard: Luna high
• Deep: Sol medium or high
• Exceptional override: Sol max

This removes Terra and most effort settings while preserving the useful economic frontier.

A safer agentic escalation loop

1. Start with the cheapest tier appropriate to the consequence of failure.
2. Require the model to identify uncertainty, missing evidence, and failed checks.
3. Verify with tools/tests rather than asking the same model whether it is correct.
4. Escalate once when evidence indicates a capability failure—not merely because the answer is stylistically imperfect.
5. Send the stronger model the original evidence and failure state, not only the weaker model’s summary.
6. Measure accepted outcomes, retries, latency, and total task cost.

A cheap call followed by an expensive retry can cost more than beginning with Sol. Conversely, using Sol max on thousands of easily verified subtasks wastes money without improving the final result.

---

Coding: benchmark result vs real routing

Artificial Analysis reports:

Model (max, Codex harness)	Coding Agent Index	Relative cost statement
Sol	80	Baseline; leads all three included coding evaluations
Terra	77	~60% lower cost/task than Sol
Luna	75	~80% lower cost/task than Sol

Sol max leads DeepSWE, Terminal‑Bench v2, and SWE‑Atlas‑QnA (tying Grok 4.5/Grok Build on the latter). Yet the practical difference is only five Index points from Luna max while Luna costs roughly one-fifth as much per coding task.

That suggests:

• Use Luna for parallel discovery, mechanical edits, test generation, migrations with strong checks, and bounded issues.
• Use Sol for ambiguous bugs, planning that spans systems, difficult review, security-sensitive changes, and rescue after a weaker attempt fails.
• Evaluate whole-task economics, not token price: accepted PR rate, number of retries, human review time, and regressions matter more than the nominal call cost.
• Keep the harness fixed when comparing models. The Coding Agent Index evaluates model-plus-harness combinations; it is not a pure model score.

---

Pricing and token economics

Published API prices per million tokens:

Model	Input	Output	Cache read	Cache write
Luna	$1.00	$6.00	$0.10	$1.25
Terra	$2.50	$15.00	$0.25	$3.125
Sol	$5.00	$30.00	$0.50	$6.25

Cache reads retain a 90% discount. GPT‑5.6 introduces OpenAI cache-write charges at 1.25× input price. Therefore:

• Stable, repeatedly reused prefixes can make caching highly valuable.
• Frequently changing “cached” context may incur writes without enough subsequent hits.
• Output-heavy agent loops are expensive: output is 6× input for every family member.
• A model that solves the task with fewer reasoning/output tokens can beat a cheaper token rate.

Sol max uses about 15k output tokens per Intelligence Index task, versus 16k for GPT‑5.5, and defines AA’s best Intelligence-vs-Output-Tokens frontier. It also uses fewer tokens while scoring higher than Claude Opus 4.8 max, GLM‑5.2 max, and Gemini 3.5 Flash high. This is why per-token price alone is misleading.

---

What the benchmarks do—and do not—prove

The Artificial Analysis Intelligence Index v4.1 combines nine evaluations: GDPval‑AA v2, τ³‑Banking, Terminal‑Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA‑Omniscience, and AA‑LCR.

Important qualifications:

• An Index point is an aggregate, not a guaranteed percentage improvement on your task.
• Close scores may not be operationally distinguishable without repeated workload-specific tests.
• Cost per task depends on AA’s prompts, token use, cache assumptions, and benchmark weights.
• Reasoning settings can change latency and output volume as well as score.
• Agent results depend on the harness, tools, prompt, context management, and stopping policy.
• Sol max’s AA‑Briefcase presentation strength does not imply the best analytical content. Claude Fable 5 still leads overall, including 56% vs 42% rubric score and 1764 vs 1592 Analytical Quality Elo.
• AA reports a small GPT‑5.6 Sol uplift over GPT‑5.5 in AA‑Omniscience accompanied by an increase in hallucination rate. More capable does not mean self-verifying.

---

How to choose for your own workload

Run a small blind evaluation before standardising. Sample 20–50 real tasks, preserve the original task distribution, and compare only plausible frontier settings:

1. Luna medium
2. Luna high
3. Luna max
4. Sol low
5. Sol medium
6. Sol high
7. Sol xhigh or max for the hardest slice

Do not spend evaluation budget exhaustively testing Terra unless there is a concrete reason.

Record:

• pass/fail against an external rubric or tests;
• human correction minutes;
• number of retries and escalations;
• end-to-end latency;
• input, cache-write, cache-hit, reasoning, and answer tokens;
• actual cost per accepted result;
• severity of failures, not only average score.

Then pick the cheapest setting that clears the required reliability threshold. The key metric is:

Total cost per accepted outcome = model cost + retry cost + reviewer time + expected failure cost.

A tiny benchmark gain can be worth a lot in a costly workflow and essentially nothing in a reversible one.

---

Bottom line

• Best economic default: Luna high.
• Best family-transition pair: Luna max vs Sol medium.
• Best serious-work default: Sol high.
• Best near-frontier default: Sol xhigh.
• Best absolute GPT‑5.6 result: Sol max—but reserve it for tasks where the premium tail matters.
• Hardest tier to justify from aggregate data: every Terra setting.

The family becomes much less confusing once treated as two useful curves rather than three product tiers: Luna buys throughput; Sol buys capability; reasoning effort chooses how far up either curve to travel.

---

Sources and reproducibility notes

• Artificial Analysis, “GPT‑5.6 benchmarks across Intelligence, Speed and Cost,” 9 July 2026.
• Artificial Analysis live LLM leaderboard, inspected 11 July 2026.
• All 15 variants were explicitly selected in the live model picker. The all-effort table transcribes AA’s directly displayed values; token and time values retain AA’s display rounding.
• Prices exclude platform subscriptions, human labour, and failure impact.

Nice comparison, thanks! What are your thoughts after the new pricing with 80% off for Luna and 20% off for Terra? Is there any logical reason to keep the Terra-High as Main Agent (Orchestrator), or is Luna-Max as Main Agent + Sol-High as Reviewer still the best combo?

Thanks for this!

Nice comparison, thanks! What are your thoughts after the new pricing with 80% off for Luna and 20% off for Terra? Is there any logical reason to keep the Terra-High as Main Agent (Orchestrator), or is Luna-Max as Main Agent + Sol-High as Reviewer still the best combo?

@PVLPM apologies for the delay :D Currently my setup is:

• sol high for main agent for actual things I care about
• luna high for managing PCs/quick exploratory work
• luna high for explorer and reviewer subagents
• sol mid (or high cant remember) for manual review loops.

If you can't use Sol as main agent because of quota/cost constraints I would advise Terra, yes. It's just vibes/logic, not actual evals/benchmarks, but a bigger model will always be better for this due to its inherent bigger out of the box knowledge. Luna can rival bigger models because of reasoning, but I do have a funny feeling it will not be as good of a main agent as terra.

Thanks!

What's most convenient way to herd bunch of Luna's highs with Sol high? Is there any industry standard? AFAIK Codex subagents can be only the same model as invoker?

@peterpro opencode and preferably ohmyopencode-slim which has common subagents i.e planner, librarian, explorer, fixer, orchestrator.

i do the following:

orchestrator: sol high/medium
oracle: sol high/xhigh
fixer: luna high/max
explorer: luna medium/high
librarian: luna medium

i have 2 presets: high and medium, depending on the difficulty of a task.