What is the difference between GPT-6 Astra, Sol and Luna?

All three are OpenAI GPT-6 models with a 1,050,000-token context window, 128,000 max output, text and image input, and reasoning effort from none to max. Astra, released September 3, 2026, is the flagship at $10/$50 per million tokens and OpenAI's best model across the board. Sol, released September 22 at $2/$10, is built for complex coding and agentic workflows. Luna, also September 22 at $0.10/$0.50, is the most efficient model for focused, high-volume tasks. Astra has an April 30, 2026 knowledge cutoff, Sol April 20 and Luna May 18.

Is GPT-6 Sol good enough instead of GPT-6 Astra?

For business-workflow automation, often yes: OpenAI reports GPT-6 Sol at xhigh scoring 33.2% on AutomationBench versus 30.3% for Astra at low effort, at about a quarter of Astra's cost per task. For hard reasoning, no: Artificial Analysis scores Sol at 48 on its Intelligence Index versus 53 for Astra, and OpenAI says Astra remains its best model for computer use and the most demanding projects. OpenAI's own guidance is to choose Astra when you want the best results and an uncompromising experience.

How cheap is GPT-6 Luna compared to other models?

At $0.10 per million input tokens and $0.50 per million output, GPT-6 Luna is one of the cheapest models OpenAI has ever released, beaten only by GPT-4.1 Nano and GPT-5 Nano from 2025. It is one tenth the price of Claude Haiku 4.5 ($1/$5), about 7x cheaper on output than Gemini 3.8 Flash ($0.75/$3.75 intro), cheaper than DeepSeek V4 Flash off-peak ($0.22/$0.66), and half the price of GPT-5.6 Luna. Cached input is $0.01 per million tokens.

Do GPT-6 models cost more for long prompts?

Yes. On GPT-6 Astra, Sol and Luna, any request with more than 272,000 input tokens is billed at 2x the input and cache rates and 1.5x the output rate for the entire request, not just the tokens above the threshold. A 300K-token prompt on GPT-6 Sol therefore costs $4 per million input and $15 per million output. Claude Opus 5.5 and Fable 5.1 have no such threshold on their 1M windows, which matters for whole-repository prompts.

Quick Answer

GPT-6 Sol vs Luna vs Astra: Which GPT-6 to Use (Sep 2026)

Published: September 23, 2026

The short answer

Route by task, not by loyalty: GPT-6 Luna ($0.10/$0.50) for high-volume focused work, GPT-6 Sol ($2/$10) for coding and business agents, GPT-6 Astra ($10/$50) only where you can measure that its extra depth pays. As of September 22, 2026 the three models share a context window, tool set and API surface, so switching is a one-line model-ID change and a reasoning-effort setting. OpenAI’s own numbers show Sol beating low-effort Astra on AutomationBench at a quarter of the cost, and independent scores put Astra five index points ahead of Sol, so the middle model is the default for most production agents.

Side by side

	GPT-6 Astra	GPT-6 Sol	GPT-6 Luna
Released	September 3, 2026	September 22, 2026	September 22, 2026
Price (in / cached / out per MTok)	$10 / $1.00 / $50	$2 / $0.20 / $10	$0.10 / $0.01 / $0.50
Cache write	$12.50	$2.50	$0.125
Context / max input / max output	1,050,000 / 922,000 / 128,000	Same	Same
Knowledge cutoff	April 30, 2026	April 20, 2026	May 18, 2026
Effort levels	low → max	none → max	none → max
Modalities	Text + image in, text out	Same	Same
AA Intelligence Index	53	48	Level with GPT-5.6 Luna per AA
Output tokens per AA task (max)	~27K	Not yet published	Not yet published
>272K prompt repricing	2x in / 1.5x out	Same	Same
Batch, Flex / fast mode	50% / 2x	Same	Same
ChatGPT	Paid plans (staged rollout since Sep 3)	Work + Codex, paid plans	Work + Codex; Free/Go desktop
Tier-5 TPM	—	40M	180M

What OpenAI’s benchmarks say

Evaluation	Astra	Sol	Luna
AutomationBench 1.0.6	30.3% (low), 41.4% (best, per Anthropic’s table)	33.2% (xhigh), $0.27/task	+5.4 pts vs GPT-5.6 Luna at 58% lower cost
Agents’ Last Exam V1	—	56.4% (max)	—
DeepSWE v1.1	—	68.8% (max)	66.6% (max)
FrontierCode 1.1 Main	53.3%	“matches Fable 5.1 xhigh”	—
OSWorld 2.0 offline	”World’s best” per OpenAI	60.5% (xhigh)	Beats GPT-5.6 Sol (medium) at 1/10 cost
Terminal-Bench 4.0	57.9% (high)	—	—
Terminal-Bench-Science	64.6%	—	—

The DeepSWE numbers are the most telling: Luna at max (66.6%) sits only 2.2 points below Sol at max (68.8%) at a twentieth of the price, and OpenAI says Luna at max matches Claude Opus 5 and Fable 5 at medium while costing 93–96% less per task. On repository-level coding, the cheapest GPT-6 is close to the middle one.

Cost per task: the 20x and 5x gaps

A representative agent turn with 30K input (80% cached) and 5K output:

Model	Input cost	Output cost	Per turn	1,000 turns
Astra	$0.084	$0.250	$0.334	$334
Sol	$0.0168	$0.050	$0.067	$67
Luna	$0.00084	$0.0025	$0.0033	$3.34

Astra is 5x Sol and 100x Luna per token. Astra partly claws that back with token efficiency (about 27K output tokens per Artificial Analysis task at max, versus 78K–119K for the Claude flagships), but Sol and Luna were trained with the same methods and OpenAI’s cost-per-task charts show them undercutting Astra at equal scores on automation and coding.

The 272K rule

Every GPT-6 model doubles input and cache rates and multiplies output by 1.5x for the whole request once input exceeds 272,000 tokens. Consequences:

A 300K-token repository prompt on Sol bills at $4/$15, on Astra at $20/$75.
Chunk large inputs, summarise, or use file_search rather than stuffing the window.
Claude Opus 5.5 ($4/$20) and Fable 5.1 have no threshold, so above 272K the Claude models can be cheaper than Sol. See GPT-6 Astra long-context pricing.

Routing table

Workload	Model	Effort	Why
Classification, extraction, routing, summarisation at volume	Luna	none–low	$0.10/$0.50; 180M TPM at tier 5
Customer-facing chat	Luna	low–medium	Factuality matches GPT-5.6 Sol at 1/100th cost
SQL, small app generation, scripts	Luna	medium–high	Simon Willison moved a production agent demo to Luna
Codex daily driver, PR-ready changes	Sol	medium–high	FrontierCode ≈ Fable 5.1 xhigh; DeepSWE 68.8%
Multi-app business workflows (Zapier-style)	Sol	xhigh	33.2% AutomationBench at $0.27/task, beats Astra low
Long-horizon professional tasks	Sol, escalate to Astra	max	Agents’ Last Exam 56.4%
Computer use, browser agents	Astra	high	OpenAI: still the world’s best; Sol 60.5% OSWorld
Scientific / research agents	Astra	high–max	Terminal-Bench-Science 64.6%
Hard single-shot reasoning, math	Astra	xhigh–max	Index 53 vs 48
Anything over 272K input tokens	Reconsider	—	Whole-request repricing; compare Opus 5.5

A common pattern: Luna as the router and first responder, Sol as the worker, Astra as the escalation path triggered by failed tests or low confidence. Because all three share the new GPT-6 caching system, you can change effort mid-conversation without invalidating the cache, which makes escalation cheap.

Versus the competition

At $2/$10, Sol is now the same input price as Grok 4.7 ($2/$6) and half of Claude Opus 5.5 ($4/$20), a model that scores ten index points higher; see Opus 5.5 vs GPT-6 Sol. At $0.10/$0.50, Luna undercuts every mainstream flash-tier model, including Gemini 3.8 Flash’s introductory $0.75/$3.75 and Xiaomi’s new MiMo-V2.6-Flash at $0.14/$0.28 on input. Anthropic has promised Haiku 5.5 “in the coming weeks”; Haiku 4.5 at $1/$5 is currently 10x Luna.

Last verified: September 23, 2026.