What are the main differences between GPT-6 Luna and GPT-6 Sol?

Both are OpenAI reasoning models released on Sep 22, 2026 with the same 1.05M context window and 128K output. Luna costs $0.10/$0.50 per MTok — 20x cheaper than Sol at $2/$10. Sol leads all benchmarks, scoring 48 vs 38 on the Intelligence Index at max effort. Luna is roughly 1.7x faster for output tokens. Luna is the high-volume budget model; Sol is the quality model.

Is GPT-6 Luna 20x cheaper than GPT-6 Sol?

Yes. Input is $0.10 vs $2.00 (20x), output is $0.50 vs $10.00 (20x), cache reads are $0.01 vs $0.20 (20x), and batch pricing is $0.05/$0.25 vs $1/$5 (20x). Every pricing line is exactly 20x cheaper. Per task the ratio is roughly 14.9x on the Artificial Analysis benchmark because Luna uses fewer reasoning tokens at the same effort level.

Is GPT-6 Sol worth 20x the price of GPT-6 Luna?

It depends on the task. Sol leads the Intelligence Index by 26% (48 vs 38) and is significantly stronger on reasoning (BenchLM 79.6 vs 55.6) and engineering benchmarks. But on DeepSWE, Luna nearly matches Sol (66.6% vs 68.8%) at a fraction of the cost. For routine coding and high-volume classification, Luna's 14.9x lower per-task cost usually wins. For complex reasoning and agentic workflows, Sol's quality premium pays for itself.

How close is GPT-6 Luna to GPT-6 Sol on coding benchmarks?

Very close on DeepSWE v1.1: Luna scores 66.6% vs Sol's 68.8% at max effort — only 2.2 percentage points apart. But FrontierCode 1.1 shows a larger gap (42.4% vs 49.3%), and BenchLM coding scores diverge further (50.2 vs 61.9). Luna is competitive on straightforward code tasks but falls behind on harder agentic coding.

Is GPT-6 Luna faster than GPT-6 Sol?

Yes. Luna outputs 127 tokens per second compared to Sol's 74 tokens per second — roughly 1.7x faster. Combined with its 20x lower price, Luna is the clear choice for latency-sensitive, high-throughput workloads where benchmark quality above the 38-point threshold is not required.

Does GPT-6 Luna work well at low effort?

Barely. On DeepSWE, Luna at low effort scores just 2.4% — nearly unusable for agentic tasks. On AutomationBench, low effort scores 1.2%. Luna needs at least medium effort (DeepSWE 44.5%) to produce useful results. Sol at low effort is much more capable (DeepSWE 37.2%, AutomationBench 21.2%).

Can GPT-6 Luna do function calling?

Only with reasoning_effort set to none. At all other effort levels (low through max), function calling is not supported. Sol supports function calling at every effort level. If your workflow requires tool use with reasoning, Sol is the only option between the two.

Is there a GPT-6.1 Luna?

No. OpenAI released GPT-6.1 Sol on Sep 29, 2026, but there is no GPT-6.1 Luna. Luna has no announced successor. If you need an upgrade path, you would move from Luna to Sol or GPT-6.1 Sol.

When should I use GPT-6 Luna instead of GPT-6 Sol?

Use Luna for high-volume, cost-sensitive workloads: classification, summarization, routine code generation, data extraction, and any task where the 38-point Intelligence Index score is sufficient. Use Sol for complex reasoning, agentic coding, multi-step workflows, and tasks where accuracy matters more than per-token cost.

Do GPT-6 Luna and GPT-6 Sol have the same context window?

Yes, both have a 1.05M token context window with a 922K max input and 128K max output. Both also share the same long-context surcharge: past 272K input tokens, pricing doubles for input and cache, and output goes up 1.5x. The surcharge structure is identical; only the base rates differ by 20x.

GPT-6 Luna vs GPT-6 Sol

Same family, 20x price gap — Luna costs $0.10/$0.50 per MTok vs Sol's $2/$10, but Sol leads every benchmark by a wide margin.

GPT-6 Luna and GPT-6 Sol both launched on September 22, 2026 as OpenAI's reasoning model duo. Luna is the lightweight tier at $0.10/$0.50 per MTok; Sol is the quality tier at $2/$10. This GPT-6 Luna vs GPT-6 Sol comparison covers their 20x pricing gap, benchmark results across 13 evaluations, effort-ladder scaling, output speed, and when each model is the right pick.

Last updated 2026-10-01 · Vendor data sourced from OpenAI Luna and OpenAI Sol

At a Glance: GPT-6 Luna or GPT-6 Sol?

Choose GPT-6 Luna when…

Volume and cost matter most — classification, summarization, data extraction and routine code generation at 20x lower cost and ~1.7x faster output. Quality threshold below ~38 on the Intelligence Index.

Choose GPT-6 Sol when…

Accuracy and reasoning depth are critical — agentic coding, multi-step reasoning, legal analysis and engineering tasks where Sol's 26% higher Intelligence Index score justifies the 20x price premium.

GPT-6 Luna vs GPT-6 Sol: Key Specifications

API identifiers, pricing, context limits and reasoning behavior — sourced from OpenAI's official documentation.

Spec	GPT-6 Luna	GPT-6 Sol
API model ID	gpt-6-luna	gpt-6-sol
Released	Sep 22, 2026	Sep 22, 2026
Tagline	OpenAI's lightweight reasoning model — 20x cheaper than GPT-6 Sol	OpenAI's efficient reasoning model (superseded by GPT-6.1 Sol)
Latency	Very fast	Fast
Input price (per MTok)	$0.10	$2
Output price (per MTok)	$0.50	$10
Context window	1.05M tokens	1.05M tokens
Max output	128K tokens	128K tokens
Thinking	Reasoning model, effort configurable	Reasoning model, effort configurable
Default effort	medium	medium
Knowledge cutoff	May 18, 2026	Apr 2026
Input modalities	Text + images	Text + images
Retirement	At least 6 months notice per OpenAI policy	Superseded by GPT-6.1 Sol (Sep 29, 2026)

GPT-6 Luna Pricing: 20x Cheaper Than GPT-6 Sol

Every pricing line is exactly 20x cheaper on Luna: input $0.10 vs $2.00, output $0.50 vs $10.00, cache reads $0.01 vs $0.20, cache writes $0.125 vs $2.50, and batch pricing $0.05/$0.25 vs $1/$5. A 100K-input, 20K-output request costs $0.02 on Luna vs $0.40 on Sol.

Pricing Tier	GPT-6 Luna	GPT-6 Sol
Standard API (per MTok)	$0.10 / $0.50	$2 / $10
Batch API (50% off)	$0.05 / $0.25	$1 / $5
Cache write (5 min)	$0.13	$2.50
Cache write (1 hour)	$0.13	$2.50
Cache read	$0.01 (0.09999999999999999x base)	$0.20 (0.1x base)

Per-task cost narrows the gap somewhat. On the Artificial Analysis benchmark, Luna averages $0.07 per task vs Sol's $1.04 — a 14.9x ratio rather than 20x — because Luna generates fewer reasoning tokens at equivalent effort.

Long-context surcharge (both models): prompts exceeding 272K input tokens trigger 2x input/cache pricing and 1.5x output pricing. Luna's surcharge is still 20x cheaper than Sol's in absolute terms.

Fast mode doubles both models' base price: Luna fast at $0.20/$1.00 vs Sol fast at $4/$20 per MTok — still 20x apart.

Output Speed: GPT-6 Luna Is 1.7x Faster Than GPT-6 Sol

Luna outputs 127 tokens per second compared to Sol's 74 tokens per second — roughly 1.7x faster. For latency-sensitive applications like chatbots, real-time extraction and streaming UIs, Luna's speed advantage compounds on top of its 20x cost advantage. Sol is the slower, more deliberate model.

Artificial Analysis Benchmarks: Sol Leads All 7 Indices

Domain-specific indices from Artificial Analysis. Both models at max effort.

Benchmark	GPT-6 Luna	GPT-6 Sol	Lead
Intelligence Index (max)	38	48	Sol +10
Finance & Accounting	40	49	Sol +9
Strategy & Ops	46	52	Sol +6
Legal	40	51	Sol +11
Healthcare & Medical	37	44	Sol +7
Engineering	37	49	Sol +12
Economics	45	54	Sol +9

Sol leads every domain index. The gaps range from +6 (Strategy & Ops) to +12 (Engineering). The overall Intelligence Index gap is +10 points (48 vs 38), a 26% lead. Luna's closest showing is Strategy & Ops (46 vs 52).

Vendor Benchmarks: DeepSWE Is Surprisingly Close

Benchmarks from OpenAI vendor charts via Kingy. Both models at their best effort level.

Benchmark	GPT-6 Luna	GPT-6 Sol	Lead
DeepSWE v1.1	66.6% (max)	68.8% (max)	Sol +2.2 pp
Agents' Last Exam V1	50.9% (max)	56.4% (max)	Sol +5.5 pp
FrontierCode 1.1	42.4% (max)	49.3% (max)	Sol +6.9 pp
OSWorld 2.0 (offline)	52.7% (max)	64.4% (max)	Sol +11.7 pp
AutomationBench 1.0.6	20.7% (max)	33.2% (xhigh)	Sol +12.5 pp
Factual error rate	7.6% (max)	4.6% (max)	Sol 1.7x better

The headline: Luna scores 66.6% on DeepSWE vs Sol's 68.8% — only 2.2 pp apart — at 20x lower cost. This makes Luna a strong value pick for straightforward coding tasks. But wider gaps on OSWorld (+11.7 pp) and AutomationBench (+12.5 pp) show Sol's advantage on complex agentic work.

Effort Ladder: How GPT-6 Luna and GPT-6 Sol Scale

DeepSWE v1.1 and AutomationBench 1.0.6 scores at each effort level (vendor charts via Kingy). Luna at low effort is barely usable for agentic tasks.

Effort	Luna DeepSWE	Sol DeepSWE	Luna AutoBench	Sol AutoBench
low	2.4%	37.2%	1.2%	21.2%
medium	44.5%	56.6%	9.4%	26.9%
high	59.3%	65.3%	14.5%	31.2%
xhigh	61.3%	66.6%	12.6%	33.2%
max	66.6%	68.8%	20.7%	32.0%

Luna at low effort is near-zero. DeepSWE 2.4% and AutomationBench 1.2% — effectively unusable. Sol at low effort is already at 37.2% and 21.2%. Luna needs medium effort as a minimum for any agentic work.

Luna converges at max effort. On DeepSWE, Luna climbs from 2.4% (low) to 66.6% (max), closing to within 2.2 pp of Sol. But on AutomationBench, the gap persists: 20.7% vs 33.2% even at max.

Sol peaks at xhigh on AutomationBench. Sol hits 33.2% at xhigh and drops to 32.0% at max — a sign of diminishing returns. Luna keeps climbing through max on both benchmarks.

Independent Evaluations: GPT-6 Luna vs GPT-6 Sol on BenchLM

BenchLM category averages (Sep 30, 2026).

Category	GPT-6 Luna	GPT-6 Sol
Overall	64.95	79.16
Agentic	53.0	59.1
Coding	50.2	61.9
Reasoning	55.6	79.6
Multimodal	72.3	83.8
Knowledge	63.6	77.3

Sol leads every BenchLM category. The widest gap is reasoning (+24.0 points), followed by knowledge (+13.7). The narrowest is agentic (+6.1). Overall: 79.16 vs 64.95 — Sol scores 22% higher.

Function Calling: GPT-6 Luna vs GPT-6 Sol

Luna supports function calling only at reasoning_effort=none. At all other effort levels (low through max), tool use is not available. This is a significant limitation for agentic workflows that need both reasoning and tool calls.

Sol supports function calling at every effort level. If your application requires reasoning with tool use, Sol is the only choice between these two models. Alternatively, consider GPT-6 Astra or Claude Opus 5.5 for full tool-use support with reasoning.

GPT-6 Luna vs GPT-6 Sol: Which to Choose by Use Case

Match your workload to the model that gives the best balance of quality and cost.

Scenario	Recommended	Why
Classification & extraction	GPT-6 Luna	20x cheaper, 1.7x faster; accuracy gap is narrow for structured tasks
Summarization at scale	GPT-6 Luna	$0.07 per task vs $1.04 — 14.9x savings with acceptable quality
Routine code generation	GPT-6 Luna	DeepSWE gap is only 2.2 pp (66.6% vs 68.8%) at 20x lower cost
Agentic coding workflows	GPT-6 Sol	AutomationBench +12.5 pp; OSWorld +11.7 pp; function calling at all effort levels
Complex reasoning	GPT-6 Sol	BenchLM reasoning 79.6 vs 55.6 — 43% higher score
Legal & compliance analysis	GPT-6 Sol	Legal index 51 vs 40; factual error rate 4.6% vs 7.6%
Real-time chat / streaming	GPT-6 Luna	127 t/s vs 74 t/s output speed; 20x lower cost per conversation

No GPT-6.1 Luna: Upgrade Paths

OpenAI released GPT-6.1 Sol on September 29, 2026 — but there is no GPT-6.1 Luna. Luna has no announced successor.

If you outgrow Luna's quality ceiling, your upgrade options within OpenAI are GPT-6 Sol ($2/$10, 20x more expensive) or GPT-6.1 Sol (same $2/$10 base, better benchmarks, cheaper cache reads at $0.10). There is no intermediate pricing tier between Luna's $0.10/$0.50 and Sol's $2/$10.

For cross-vendor alternatives, consider Claude Opus 5.5 ($4/$20) which leads Sol on most benchmarks, or Claude Sonnet 5.5 ($2/$10) as a same-priced Sol alternative from Anthropic.

The Bottom Line

GPT-6 Luna and GPT-6 Sol are the budget and quality tiers of the same GPT-6 reasoning family. Luna is 20x cheaper per token, 14.9x cheaper per task, and 1.7x faster — but Sol leads every benchmark, often by double digits. Luna nearly matches Sol on DeepSWE (66.6% vs 68.8%), making it a genuine value pick for straightforward coding. But Sol pulls ahead sharply on reasoning (BenchLM 79.6 vs 55.6), agentic workflows and function calling. Use Luna for volume; use Sol for quality.

GPT-6 Luna vs GPT-6 Sol: FAQ

What are the main differences between GPT-6 Luna and GPT-6 Sol?
Both are OpenAI reasoning models released on Sep 22, 2026 with the same 1.05M context window and 128K output. Luna costs $0.10/$0.50 per MTok — 20x cheaper than Sol at $2/$10. Sol leads all benchmarks, scoring 48 vs 38 on the Intelligence Index at max effort. Luna is roughly 1.7x faster for output tokens. Luna is the high-volume budget model; Sol is the quality model.
Is GPT-6 Luna 20x cheaper than GPT-6 Sol?
Yes. Input is $0.10 vs $2.00 (20x), output is $0.50 vs $10.00 (20x), cache reads are $0.01 vs $0.20 (20x), and batch pricing is $0.05/$0.25 vs $1/$5 (20x). Every pricing line is exactly 20x cheaper. Per task the ratio is roughly 14.9x on the Artificial Analysis benchmark because Luna uses fewer reasoning tokens at the same effort level.
Is GPT-6 Sol worth 20x the price of GPT-6 Luna?
It depends on the task. Sol leads the Intelligence Index by 26% (48 vs 38) and is significantly stronger on reasoning (BenchLM 79.6 vs 55.6) and engineering benchmarks. But on DeepSWE, Luna nearly matches Sol (66.6% vs 68.8%) at a fraction of the cost. For routine coding and high-volume classification, Luna's 14.9x lower per-task cost usually wins. For complex reasoning and agentic workflows, Sol's quality premium pays for itself.
How close is GPT-6 Luna to GPT-6 Sol on coding benchmarks?
Very close on DeepSWE v1.1: Luna scores 66.6% vs Sol's 68.8% at max effort — only 2.2 percentage points apart. But FrontierCode 1.1 shows a larger gap (42.4% vs 49.3%), and BenchLM coding scores diverge further (50.2 vs 61.9). Luna is competitive on straightforward code tasks but falls behind on harder agentic coding.
Is GPT-6 Luna faster than GPT-6 Sol?
Yes. Luna outputs 127 tokens per second compared to Sol's 74 tokens per second — roughly 1.7x faster. Combined with its 20x lower price, Luna is the clear choice for latency-sensitive, high-throughput workloads where benchmark quality above the 38-point threshold is not required.
Does GPT-6 Luna work well at low effort?
Barely. On DeepSWE, Luna at low effort scores just 2.4% — nearly unusable for agentic tasks. On AutomationBench, low effort scores 1.2%. Luna needs at least medium effort (DeepSWE 44.5%) to produce useful results. Sol at low effort is much more capable (DeepSWE 37.2%, AutomationBench 21.2%).
Can GPT-6 Luna do function calling?
Only with reasoning_effort set to none. At all other effort levels (low through max), function calling is not supported. Sol supports function calling at every effort level. If your workflow requires tool use with reasoning, Sol is the only option between the two.
Is there a GPT-6.1 Luna?
No. OpenAI released GPT-6.1 Sol on Sep 29, 2026, but there is no GPT-6.1 Luna. Luna has no announced successor. If you need an upgrade path, you would move from Luna to Sol or GPT-6.1 Sol.
When should I use GPT-6 Luna instead of GPT-6 Sol?
Use Luna for high-volume, cost-sensitive workloads: classification, summarization, routine code generation, data extraction, and any task where the 38-point Intelligence Index score is sufficient. Use Sol for complex reasoning, agentic coding, multi-step workflows, and tasks where accuracy matters more than per-token cost.
Do GPT-6 Luna and GPT-6 Sol have the same context window?
Yes, both have a 1.05M token context window with a 922K max input and 128K max output. Both also share the same long-context surcharge: past 272K input tokens, pricing doubles for input and cache, and output goes up 1.5x. The surcharge structure is identical; only the base rates differ by 20x.

Related Comparisons

Claude Opus 5.5 vs GPT-6 Sol — Anthropic's flagship vs Sol
Claude Opus 5.5 vs GPT-6 Astra — Opus vs OpenAI's flagship ($10/$50)
Claude Opus 5.5 vs Claude Sonnet 5.5 — Anthropic's same-vendor choice
GPT-6 Sol vs GPT-6.1 Sol — Sol's seven-day successor