OpenAI cut GPT-6 Sol and GPT-6 Luna list prices in half versus the GPT-5.6 promo on 22 Sep 2026 — and freelancers still ask the same three questions. Which tier for coding versus clerical volume? Is Astra still worth it on the hardest client work? And why do Work and Codex show the new models while Chat mode does not?

This is a tier chooser, not Agents API sandbox math and not a full Claude-vs-OpenAI ladder. Harness meters (tokens + tools + containers) stay on our Agents API freelancer budget sheet. Cross-vendor $/MTok stays on the API cost ladder. Here the only job is Sol vs Luna vs Astra vs stay-on-Chat after the cut — with IDs gpt-6-sol and gpt-6-luna, not the older gpt-5.6-sol strings blogs still paste.

Affiliate disclosure: some tool or platform links may later be affiliate or referral. Prices below are from OpenAI’s live model docs and ChatGPT rate card as of 28 Sep 2026 (SAST) — re-check before you quote a client. Not legal advice; not a guarantee of clients, savings, or income.

What shipped 22 Sep 2026 — Sol + Luna in the GPT-6 family

On 22 Sep 2026, OpenAI staff announced GPT-6 Sol and GPT-6 Luna in the API, Codex, and ChatGPT. The post frames both as building on GPT-6 Astra, with caching and inference efficiency passed through as 50% lower API prices versus GPT-5.6 promotional pricing. Treat the “50%” line as OpenAI’s claim against that promo baseline — not a guarantee about your last invoice.

Model docs: GPT-6 Sol and GPT-6 Luna. API IDs are gpt-6-sol and gpt-6-luna. Do not confuse them with gpt-5.6-sol / gpt-5.6-luna. Both sit in the GPT-6 family under Astra; Sol is the complex coding / agentic seat, Luna the high-volume efficient seat.

Rollout language on the announcement: Work + Codex for Plus, Pro, Business, Enterprise, and Edu; both models on the API; Free and Go can try Luna on the desktop app. Chat availability is a separate gate — covered below.

Sol vs Luna in plain English

Sol is positioned for complex coding and agentic workflows with reasoning controls (none through max; default medium on the model page). Use it when the job needs multi-file judgment, tool loops, or reasoned edits you will actually ship. Sol is the seat you A/B when last month’s GPT-5.6 Sol (or a mid Claude) felt close on quality but the new list price changes the quote math.

Luna is positioned as the most efficient model for focused, high-volume tasks — summaries, extract, light batch transforms, clerical volume where “good enough + cheap” beats frontier polish. Official blurbs matter more than secondary “Sol is the smart one” shorthand. If the deliverable is fifty similar rows and a human spot-check, Luna is usually the first meter — not a moral judgment about “cheap models.”

Both share a large context window on the model pages (1,050,000). Prompts over 272K input tokens hit long-context multipliers on the full request — see the price section. Knowledge cut-offs differ on the docs (Sol Apr 2026; Luna May 2026 on the pages re-pulled for this draft); realtime facts still need tools when the brief needs them.

A practical freelance split: Luna for intake and volume; Sol for the coding or reasoning pass you bill as craft; Astra only when the stakes or the harness already demand it. That split is a sheet habit, not a vendor slogan.

The price sheet freelancers actually need

USD list rates from the official model docs (Standard):

- Sol: $2.00 input / $0.20 cached input / $10.00 output per 1M tokens. Cache writes $2.50. - Luna: $0.10 input / $0.01 cached input / $0.50 output per 1M. Cache writes $0.125.

That Sol row is the post-cut number freelancers care about: half the GPT-5.6 Sol promo list OpenAI cites in the announcement, on OpenAI’s own framing. Luna’s $0.10 / $0.50 is the volume seat — still quote tokens, not “basically free.”

Prompts with more than 272K input tokens are priced at 2× input and cache rates and 1.5× output for the full request. Dumping a giant repo “because context is big” can flip the whole request onto the long row. Regional processing adds a 10% premium where available. Batch and Flex are 50% of Standard when the job can wait. Fast mode on the API is priced at 2× the applicable rates on the model pages.

The ChatGPT rate card lists the same Sol/Luna $/MTok rows for supported Work and Codex activity, and lists Fast mode at 2.5× Standard for GPT-6 Astra / Sol / Luna on Work and Codex. Prefer the multiplier for the surface you are on — API docs say 2×; Work/Codex rate card says 2.5× for those GPT-6 tiers. Do not invent monthly averages or ChatGPT Plus/Pro dollar prices from secondary blogs. Rand clients still see USD API and token invoices — verify card and billing country; do not invent FX markups.

ChatGPT product gates that blogs blur

Help Center language is blunt: GPT-6 Sol and GPT-6 Luna are models for ChatGPT Work and Codex. They are not available in Chat.

That single sentence explains most “I don’t see Sol in my picker” threads. Switch to Work or open Codex on an eligible plan. Free/Go Luna-on-desktop is a try path, not a reason to invent Free-tier Sol access. Do not paste MacRumors plan dollars into a client quote — quote what your account screen and the rate card show.

If the client deliverable is ordinary chat drafting, your Chat model lineup may still be the older seat. If the deliverable is Work/Codex coding or delegated work, Sol/Luna are the new rows. Community threads complaining that Sol is missing from Chat are describing the product gate, not a broken account — until OpenAI changes that Help Center line.

Included plan usage, five-hour windows, and legacy credit rates are separate commercial shapes from token-based Work/Codex metering. Read the rate card footnotes for your workspace type before you promise a client “unlimited Sol.” Chat crown comparisons stay on ChatGPT vs Claude vs Gemini.

When Astra still wins

GPT-6 Astra remains the top-end seat. The rate card lists Astra at $10.00 input / $1.00 cached / $50.00 output per 1M for supported Work/Codex activity — cite that card as context, not as a reason to burn Astra on every overnight job. Use Astra when the brief is highest-stakes agentic work, the client pays for frontier review quality, or your Agents harness samples already assume gpt-6-astra.

Skip Astra as a vanity default after the Sol cut. Sol at $2 / $10 exists so everyday coding does not need the Astra row. If Sol already clears acceptance criteria on a logged job, moving “up” to Astra without a failed Sol row is spend, not craft.

If the deliverable is a hosted agent with sandbox meters, read the Agents API budget sheet — Sol/Luna choice is only the model meter; tools and containers are separate. For everyday coding after the price cut, Sol is usually the first A/B against Luna, not an automatic Astra default. Long API or Work sessions still need a stable uplink; for mobile-data packaging see fibre vs mobile data for AI work.

Buy/skip matrix (sheet)

Get the 1-page Sol / Luna / Astra chooser

Optional printable of this free guide. Soft link until checkout. Not a “paid summary.” Not a guarantee of clients, savings, or income.

This week’s action — one recurring job, Luna once and Sol once

Pick one recurring job — for example a weekly content batch versus a multi-file bugfix. Run it once on Luna and once on Sol with a hard spend cap. Same brief, non-secret materials only. Note which surface you used (API vs Work/Codex) and whether Fast mode was on. Fill one sheet row: Luna for volume / Sol for everyday coding / Astra for top-end only / stay Chat if Work-Codex isn’t your surface / API when the client owns the key.

If Sol wins on quality but Luna wins on meter for clerical rows, keep both seats — that is the point of a tier chooser after a price cut. Do not change a production default mid-client week on a community announcement alone. Verify billing country and card on platform.openai.com at signup; do not invent ZAR.

Sources: gpt-6-sol, gpt-6-luna, ChatGPT rate card, community announcement — re-check IDs and Fast multipliers at paste time.