Key Takeaway
- 🪙 Two ships in one week: GPT-6 Sol and GPT-6 Luna landed together on September 22 via OpenAI’s API changelog — text+vision in, text out, 1,050,000-token context and 128,000-token output on both.
- 💰 The price fork is the story: Sol runs $2 in / $10 out per million tokens; Luna runs $0.10 in / $0.50 — a 20x gap between two models with the same giant context window.
- 🧠 Different cutoffs matter: Sol’s knowledge cutoff is April 20, 2026; Luna’s is May 18 — the cheaper model knows more-recent events.
- 🎯 Pick by workload: Sol for complex coding and agentic chains; Luna for focused, high-volume tasks — classification, extraction, drafting, summarizing at scale.
- 🇵🇭 For Filipino builders: the $0.10 tier makes always-on automation viable for the first time — a full day of Luna batch work costs what three Sol prompts used to.
Table of Contents
The Frontier Just Split in Two
For three years, using OpenAI meant watching one flagship at a time. GPT-4, then its Turbo variants, then GPT-5’s ladder of dated snapshots — one name, one price, one decision. On September 22, 2026, that mental model broke.
OpenAI’s API changelog shipped GPT-6 Sol and GPT-6 Luna in the same window, and the question “which model do you use?” suddenly has a wallet-dependent answer.
The numbers first, because they are unusual. Sol — positioned for complex coding and agentic workflows — costs $2 per million input tokens and $10 per million output tokens. Luna, which OpenAI describes as its most efficient model for focused, high-volume tasks, costs $0.10 in and $0.50 out. Same 1,050,000-token context window.
Same 128,000-token output ceiling. Twenty times apart on price.
What Actually Shipped on September 22
Both models accept text and image input and return text — no audio, no video generation, no new modality tricks. The interesting asymmetry hides in the knowledge cutoffs: Sol’s training data ends April 20, 2026, while Luna’s runs to May 18.
Counterintuitive at first — the budget model is the more current one — but sensible once you see the role split. Luna is built to be embedded: in pipelines, cron jobs, support triage, document flows. Recency in a volume model prevents the silent-failure mode where a cheap model confidently asserts stale facts.
Sol’s April cutoff pairs with a model built for deep work where the reasoning matters more than yesterday’s headlines.
The changelog release pattern — specs and prices published, positioned against each other, no launch event — matches how OpenAI has moved all year, as the official API changelog shows in its own history.
Astra’s September 17 launch — specs and prices published, positioned against each other, no launch event — matches how OpenAI has moved all year. Astra’s September 17 launch carried the $10/$50 per-million top tier and the 230-million-URL index for legal work. Sol sits at $2/$10 — the previous flagship tier, one-fifth of Astra’s input price.
Luna at $0.10/$0.50 undercuts Gemini 3.8 Flash’s introductory $0.75/$3.75 by a factor of 7.5 on input.
The Three-Tier GPT-6 Family
Stack them up and OpenAI’s strategy reads plainly:
- GPT-6 Astra — $10 in / $50 out · 1.05M context · the frontier for the hardest end-to-end work: computer use, browsing, software engineering, the Astra for Law legal bundle
- GPT-6 Sol — $2 in / $10 out · 1.05M context · complex coding and agentic workflows, April 20 cutoff
- GPT-6 Luna — $0.10 in / $0.50 out · 1.05M context · focused high-volume tasks, May 18 cutoff
That is a 100x spread from top to bottom on input tokens, inside one model family, with context width held constant across all three.
Google’s Flash line and Anthropic’s Fable line segment by capability tier; OpenAI has now segmented by workload intensity — and left the context window constant so the cheap model can still read your entire codebase.
What the 20x Price Gap Buys You
Run the arithmetic on a real workload. A support-desk triage agent reading 200 tickets a day, each ~1,500 tokens with history, answering in ~300 tokens: roughly 300,000 input tokens and 60,000 output tokens daily. On Sol, that is $0.60 + $0.60 = $1.20 a day, about $36 a month.
On Luna, $0.03 + $0.03 = $0.06 a day — under $2 a month. Same context window, same API surface. The classification quality gap will matter in edge cases, but for routing, tagging, and first-draft answers, twentyfold is a category difference, not a rounding error.
For a Filipino freelance developer building client automations, this is the arrival of the always-on tier. A cron job watching a Google Sheet for invoices, a Viber-bot FAQ responder, a nightly summarizer for five podcasts — workloads that failed the budget test at $2/$10 pricing now clear at Luna rates with room to spare.
Keep Sol for the weekly heavy build; let Luna run the pipes.
How It Compares Down-Market
Luna’s $0.10/$0.50 lands below every mainstream Flash-tier competitor: Gemini 3.8 Flash at $0.75/$3.75 through December 31 (then $1.50/$7.50), Meta’s Muse Spark 1.3 at $1.25/$4.25, and DeepSeek’s V4.1 Flash — the September 10 release that briefly owned the value tier.
The competitive read: OpenAI used Astra to defend the top and Luna to reclaim the bottom in one move, compressing everyone’s mid-market, a pattern trackers like Artificial Analysis now quantify per model.
The one dimension Luna loses on is output: 128K tokens max, same as Sol and Astra, but per-task depth on hard reasoning still favors the expensive tiers — Artificial Analysis cost-per-task indexes exist precisely because cheap-per-token does not always mean cheap-per-task.
Where the Cutoffs Bite
Knowledge cutoffs are the quiet trap in a two-model family. May 18 vs April 20 is four weeks of world events, library releases, and API deprecations.
If your Luna pipeline summarizes news, prices, or anything time-sensitive, that extra month is real value; if your Sol workflow reasons over a frozen codebase, the older cutoff costs nothing. Check the model page before assuming either — OpenAI updates cutoffs with refreshed snapshots, and the changelog, not memory, is the source of truth.
What Filipino Teams Should Do This Week
- Audit your token spend for the last 30 days — anything that is classification, extraction, formatting, or short summarization is a Luna candidate; move it and watch the bill drop by an order of magnitude.
- Keep agentic chains on Sol or Astra. The 20x saving evaporates if the cheap model makes tool-calling mistakes that need human repair — cost-per-task, not cost-per-token, decides.
- Test Luna on your actual data with a 50-case sample against Sol — a half-day of evaluation buys you the pricing certainty that marketing pages will never give you.
- Watch the deprecation clock. Two ships in one week means older GPT-6 snapshots will rotate out; check the changelog subscription before your cron jobs break silently.
For teams weighing the model against Google’s stack, our Gemini 3.8 Flash breakdown covers the intro-price expiry trap in detail; and Anthropic’s cache-read repricing in Fable 5.1 shifts the same cost-per-task math. The one-model era ended not with a bang but with a price tag.
Sol is the frontier workhorse most teams actually need; Luna is the volume engine that turns AI from a budgeted line item into infrastructure. The teams that map their workloads to the fork this week will spend October’s token budget on capability, not habit.
The Freelancer’s Ledger: One Month on the Fork
Concrete plan for a small Filipino dev shop billing retainers: keep one Sol key for client builds — migration scripts, new features, agent prototypes — and route every operational task to Luna.
The monthly invoice for a shop running five client automations drops from roughly $85 to under $12, and the work product clients notice (the code) never touches the cheap tier.
Price your maintenance retainers on Luna rates and your build days on Sol rates, and the same man-hours carry a visibly better margin — margin you can reinvest in the Astra tier for the one or two flagship problems each quarter that genuinely need the frontier.
Migration Notes: Moving Workloads Down-Tier
Three practical notes from teams making the Sol-Luna switch. First, temperature and system prompts port unchanged — the API surface is identical, so downgrading a pipeline is a model-name swap, not a rewrite.
Second, watch tool-calling strictness: volume workloads that lean on function calls should keep schema descriptions tight, because efficient models spend fewer tokens interpreting loose schemas.
Third, batch endpoints still apply their usual discounts on top — Luna’s batch rate lands near pennies per thousand calls, which is why the volume tier is where the always-on automations live now.
| Workload | Right model | Why |
|---|---|---|
| Support triage, tagging, routing | Luna | High volume, bounded answers, 20x cost gap dominates |
| Client feature builds, migrations | Sol | Reasoning depth pays for itself per task |
| News and price summaries | Luna | May 18 cutoff beats Sol by four weeks |
| Long agentic chains with tool calls | Sol or Astra | Error-repair cost on cheap tiers exceeds savings |
| Legal-grade document analysis | Astra | 230M-URL index and frontier reasoning stay top-tier |
What the 20x Split Means for Working Developers
Most API workloads are not clever. They are thousands of small requests — classification, extraction, formatting, routing — that never needed a frontier brain in the first place. The Sol/Luna fork is OpenAI finally pricing that reality into the catalog.
A team spending $2,000 a month on mixed workloads can route 70-80% of its calls to Luna at $0.10/$0.50 per million tokens, keep Sol for the agentic workflows that justify it, and land near $700-900 for the same output quality on the paths that matter.
The savings do not come from a discount; they come from matching the model to the job.
The catch is routing discipline. Teams that skip the matching step and default everything to Sol get the same bill as before with extra steps.
The teams that benefit treat the fork as an architecture decision: audit last month’s traffic, bucket by task complexity, measure Luna’s accuracy against Sol on the easy bucket, and only pay the 20x premium where the quality delta is real. That measurement, not the pricing page, is what converts the split into savings.
The Timing, Read Closely
September 2026 now reads as the week the market’s structure changed three times over. Zhipu’s espionage playbook surfaced Monday. The standards body formed with Cohere objecting Wednesday. GPT-6’s two-model debut closed Friday.
Each event is a story alone; together they sketch the next twelve months — capability concentrating in fewer labs, pricing fragmenting into tiers, and governance arguing with itself about who sets the rules. Builders do not need to pick a side in the governance fight. They do need to pick their tier.
For the OFW freelancer in Riyadh invoicing a US client by the hour, the practical takeaway is smaller and sharper: the tools that make your work competitive got cheaper this week, and the premium tier got more powerful.
The gap between “uses AI” and “orchestrates AI well” is now the actual wage premium — and it widened on Friday.
Frequently Asked Questions
When were GPT-6 Sol and Luna released?
September 22, 2026, per OpenAI’s API changelog — both models in the same release window.
How much does GPT-6 Luna cost?
$0.10 per million input tokens and $0.50 per million output tokens — 20x cheaper than GPT-6 Sol ($2/$10) and 100x cheaper than GPT-6 Astra ($10/$50) on input.
Do Sol and Luna have the same context window?
Yes — both list 1,050,000-token context with up to 128,000-token output, and both accept text and image input.
Which should I use for coding?
Complex, multi-file, agentic coding on Sol; bulk code review, test generation, and doc extraction on Luna. Run both on a real sample before committing.
What is the knowledge cutoff for each model?
Sol: April 20, 2026. Luna: May 18, 2026 — the cheaper model carries the more recent training data.
Is GPT-6 Astra still available?
Yes — Astra remains the top tier at $10/$50 per million tokens for the hardest work, including the Astra for Law bundle.
