GPT-6 Sol & Luna: Cost-Speed Tier vs Astra

Last updated：2026-09-24· 16 min read

🚀 Quick access

ChatGPT Domestic：Open entry↗
Mirror site：Open mirror↗
Official ChatGPT：chatgpt.com ↗

Updated 2026-09-24. GPT-6 Sol and Luna began rolling out on 2026-09-22. Treat model pickers, API strings, quotas, and prices as live data from ChatGPT, the OpenAI developer announcement, and the Platform models page.

Overview

Searches for “GPT-6 Sol”, “GPT-6 Luna”, or gpt-6-sol usually need three answers first: which tier they are relative to GPT-6 Astra, who can see them in ChatGPT Work / Codex / Free·Go desktop, and how to migrate from GPT-5.6 Sol / Luna by task instead of flipping everything. This page covers identity checks, selection tables, and a shadow-eval checklist. Before production, re-copy that day’s model list.

What this page solves

Position Sol / Luna as the GPT-6 cost-speed daily tier—carrying Astra’s advances without being another flagship label
Separate product names from API ids gpt-6-sol and gpt-6-luna
Decide when to use Sol, Luna, Astra, or stay on 5.6
Run a bypass comparison with a one-click rollback path

What GPT-6 Sol and Luna are

GPT-6 Sol and GPT-6 Luna joined the GPT-6 family around 2026-09-22. Official messaging frames them as faster, more affordable models that bring much of Astra’s strength to work at scale. Platform docs (confirm that day) roughly say:

Model	Public positioning (summary)	Common API id	How to remember it
GPT-6 Sol	Balanced for complex coding and agentic workflows	`gpt-6-sol`	Default “new Sol” for most coding and office work
GPT-6 Luna	Efficient for focused, high-volume tasks	`gpt-6-luna`	Short jobs, batch extraction, lower unit cost
GPT-6 Astra (reference)	Flagship	`gpt-6-astra`	Still the first stop for long agents / computer use / hard deliverables

Procurement notes must record entry + exact model / UI label + date + plan. UI short names ≠ billing API strings.

Relation to Astra and GPT-5.6

Model	Role	Good for	Upgrade note
GPT-5.6 Sol / Terra / Luna	Prior stable baseline	Proven production routes	Keep as controls; do not delete just because “6 exists”
GPT-6 Luna	Light / throughput tier	Summaries, extraction, short rewrites	Free/Go desktop may see Luna first
GPT-6 Sol	Daily balanced tier	Coding help, multi-step office, most agent orchestration	Primary candidate when migrating production defaults
GPT-6 Astra	Flagship	Computer use, long agents, polished artifacts	Higher unit cost and interrupt risk—see the Astra guide

Public materials in this wave did not list a “GPT-6 Terra” beside Sol/Luna. For deeper reasoning, raise Sol’s reasoning.effort or shadow-test Astra—do not invent a missing Terra id.

Deeper 5.6 routing: GPT-5.6 Parameters: Sol / Terra / Luna.

Who can see what

ChatGPT Work / Codex: Officially rolling out to Plus, Pro, Business, Enterprise, and Edu; Enterprise may need an admin enablement.
Free / Go: Officially, GPT-6 Luna on the desktop app; do not assume web Chat mirrors desktop.
OpenAI API: Use documented gpt-6-sol / gpt-6-luna (or that day’s snapshot). ChatGPT plan quotas usually are not API quotas.
GitHub Copilot and other hosts: May ship the same labels on their own billing—validate separately from direct OpenAI.
Staged rollout: “Someone else has it” ≠ “your workspace does.” Screenshots are not contracts.

Minimal runnable request (placeholder)

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
# Confirm the string on Platform that day
model = "gpt-6-sol"  # or "gpt-6-luna"

resp = client.responses.create(
    model=model,
    input="In three bullets: why moving from GPT-5.6 Sol to GPT-6 Sol requires logging the full model id, entry, and cost per accepted answer.",
)
print(resp)

If you still use Chat Completions, field names and reasoning / reasoning_effort shapes may differ—copy from that day’s API Reference. Key handling: OpenAI API developer guide. Never ship keys in the browser.

Reasoning effort

Platform lists multiple reasoning.effort values for Sol-class models (often none / low / medium default / high / xhigh / max). Practical starts:

Goal	Starting point	What to log
Smoke test	medium or low	Auth, billing, format health
Hard refactor / multi-constraint plans	high or above	First-pass rate, tokens, latency p95
High-throughput extraction / short rewrites	Luna + lower effort	Cost per accepted answer, not list price alone

Higher effort is usually slower and pricier. Put “task → default model → default effort” in the team wiki.

Pricing language (do not hard-code)

The developer announcement said API prices for Sol and Luna are about 50% lower versus GPT-5.6 promotional pricing, reflecting efficiency gains. Platform tables still shift with context length, Batch/Flex/Fast, and regional premiums.

Safe wording:

In proposals, cite “Pricing / invoice as of date”
Accept based on dollars per accepted output + human fix minutes, not input list price alone
Re-check promo windows, residency uplifts, and long-context multipliers on a retest day

This page intentionally does not paste dollar figures into copy-paste configs—open that day’s pricing page when you need a quote.

When to use Sol, Luna, or Astra

Goal	Suggestion	Why
Move production defaults off 5.6 Sol onto GPT-6 daily tier	Shadow-test GPT-6 Sol first	Official balanced coding / agent line
High-frequency short work, batch extraction, lower unit cost	Shadow-test GPT-6 Luna	Efficiency tier; Free/Go desktop may expose it early
Computer use / long agents / high-stakes deliverables	Bypass with Astra—do not assume Sol equals flagship	See the Astra guide
Stable 5.6 Terra deep-reasoning flows	A/B: 5.6 Terra vs 6 Sol (high effort) vs Astra	Score fix minutes and severe errors
Enterprise admin still off	Get approval; do not mix customer data on personal accounts	Compliance first
Hard-coding model names for customers	Must roll back to GA 5.6 or a proven 6 tier	Staged rollout causes incidents

Copyable comparison prompt

You are an evaluation scribe. Run the same task on as many of these as you can open:
GPT-5.6 Sol / GPT-5.6 Luna / GPT-6 Sol / GPT-6 Luna / GPT-6 Astra
Fill:
| Dimension | 5.6 | 6 Sol | 6 Luna | Astra | Winner | Evidence snippet |
At least: instruction following, first usable draft, hallucination/overreach count, human fix minutes, latency feel, accepted-cost feel.
No style-only scoring; every claim must cite a concrete snippet.
Task: […]
Outputs: […]

Migration checklist

Config switch: remote config / env vars for model—no orphan hard-codes.
Route by task: short → Luna; default office/coding → Sol; flagship → Astra.
Traffic isolation: internal or 5%–10% shadow first.
Auto rollback: unavailable / 429 / timeout → documented 5.6 GA or last stable 6 id.
Billing alerts: Platform budgets; watch effort increases that erase list-price wins.
Log fields: UI label, model, effort, request id, entry (Work / Codex / API / desktop).
Least privilege: split read vs write agent actions.

Reproducible eval checklist

Header: date, entry, full model string, effort, plan
≥8 daily office/coding tasks (5.6 Sol vs 6 Sol)
≥8 high-throughput short tasks (5.6 Luna vs 6 Luna)
≥3 “wanted Astra” hard tasks to see if Sol already suffices
2–3 runs per task; log latency and format crash rate
Cost per accepted answer (tokens × that-day price + human minutes)
Verdict: default / shadow only / do not switch + untested gaps
Rollback model configured; customers not stuck on irreversible ids

Quick access

Official chat: ChatGPT
Regional convenience entry: ChatGPT regional chat (third-party ≠ official entitlement; Sol/Luna visibility follows that day’s list)
Mirror handbook: Mirror sites guide
API / projects: OpenAI Platform
Model pages: GPT-6 Sol · GPT-6 Luna

Third-party page titles are not version proof. Read terms before uploading sensitive code or customer data; Enterprise boundaries follow contract and admin settings.

FAQ

Will Sol / Luna replace GPT-5.6 immediately?

Not necessarily. Coexistence windows are common. Pin and log versions for critical flows.

Can Free users access them?

Officially, Free/Go can try Luna on desktop. Sol and full Work pickers follow what your account shows after login.

Is there a GPT-6 Terra?

Not listed alongside Sol/Luna in this wave’s public materials. Raise Sol effort or shadow Astra, and document the id you actually call.

Is `gpt-6-sol` the same as the UI label “Sol”?

Same product line, but display names and API ids need not match character-for-character, and billing often splits. Log entry, UI label, docs id, and check date together.

Are they really ~50% cheaper?

The announcement referenced GPT-5.6 promotional pricing, and accepted cost still depends on tokens and effort. Trust that day’s Pricing page plus your own same-task invoices.

Do regional mirrors match the official models?

Never assume they do. Read the technical notes and re-run fixed tasks.

Should we flip the whole API fleet today?

No. Offline eval → shadow traffic → small canary → expand only with quality, latency, cost, and rollback proven.

Official resources

Summary

GPT-6 Sol and Luna exist to test whether the daily line is already good enough versus Astra’s flagship cost, and whether accepted cost truly drops versus 5.6—not to become a blind production default. Confirm entitlements and admin switches → align UI labels with gpt-6-sol / gpt-6-luna → route by task → price accepted answers on that day’s rates → set budgets and rollback → change defaults only after evidence. Treat the staged window as a data-gathering upgrade period, not a screenshot-driven cutover.