GPT-5.6 Sol '50% Off' Is a Gateway Promo, Not an OpenAI List-Price Cut

GPTLive1 is an independent site. We are not affiliated with, endorsed by, or connected to OpenAI, OpenRouter, or Vercel. This is an explainer, not an official source.

A Hacker News thread titled "GPT-5.6 Sol Pricing Cut by 50%" linked to OpenRouter's model page. The title is easy to misread as OpenAI permanently halving the flagship. That is not what OpenAI's published API price list shows.

According to OpenAI's GPT-5.6 Sol model card and the July 30, 2026 price-performance announcement as summarized in the developer forum, Sol's standard list price stayed $5.00 input / $30.00 output per million tokens. Terra and Luna were the models that got permanent list-price cuts (20% and 80% in that announcement). Sol got Fast mode: up to 2.5× speed at twice the standard price, not half.

The 50% figure that hit the HN title is a reseller or gateway promotion, not a change to OpenAI's rate card. Treat the headline as a routing event.

Three price sheets, three different sentences

Keep these apart or the spreadsheet lies:

1. OpenAI list price (direct API). Sol standard: $5 / $30 per 1M tokens. Cached input is billed lower; cache writes are billed higher. Prompts above the published long-context threshold are billed at a multiplier for the whole request. Fast mode is 2× those standard rates for speed. Batch is a discount tier in the other direction. None of that is "Sol is 50% off."

2. OpenAI family cuts in July 2026. Luna and Terra list prices dropped. Sol's headline rate did not. If a dashboard still shows June 2026 Terra/Luna numbers, the ladder is stale. The spread between Luna and Sol widened, which makes tier choice more important than it was at launch.

3. Gateway promotions. Vercel announced 50% off GPT-5.6 Sol on AI Gateway through September 18, 2026, on traffic billed through the gateway's OpenAI provider — not BYOK. Their table lists default, flex, and priority/fast-mode rates at half the usual gateway list. The HN post pointed at OpenRouter; whatever discount that page showed on August 18 is a marketplace price, not OpenAI's. Marketplace prices move without a model card update.

If you pay OpenAI directly, a gateway coupon does not apply. If you pay a gateway, read the expiry, the BYOK exclusion, and whether cached tokens and long-context multipliers are included. Vercel said the AI Gateway discount covers those add-ons on gateway-billed traffic. Do not assume another reseller copied the same terms.

What to do before you "switch everything to Sol"

A 50% off sticker on a flagship is catnip for agents, voice cascades, and anyone who has been running a cheaper tier "until Sol is affordable." Run the arithmetic on your mix, not on the title.

• Expiry. A promotion that ends September 18 is a calendar item. Putting Sol in a default routing table because it is cheap this week is how next month's bill surprises you.
• BYOK. Keys on your own OpenAI account bill at OpenAI's list (or your enterprise contract). Gateway 50% off does not follow those keys.
• Fast mode vs discount. Fast mode makes Sol more expensive per token for latency. A 50% gateway promo on Fast mode can land near standard list, which is a different deal than "flagship at half." Read which service tier the coupon touches.
• Output tokens dominate Sol. At $30 / 1M output list, a chatty agent is not a $5 problem. Voice stacks that dump long transcripts into a flagship model feel this first.
• Terra/Luna already moved. If the workload fits Terra or Luna, the July list-price cuts are the durable story. Sol-at-a-discount is the temporary story.

None of this is investment, procurement, or legal advice. Provider pages change. Recheck the model card and the gateway changelog before a production cutover.

Why this showed up next to a voice site

GPTLive1 is an independent explainer for OpenAI's GPT-Live-1 full-duplex voice model, plus a small in-browser TTS demo. It is not affiliated with OpenAI, OpenRouter, or Vercel. The Sol headline still matters to the same readers because text-model prices leak into voice products.

A typical cascade is STT → LLM → TTS. The LLM hop is often where flagship spend hides. A month of discounted Sol on a gateway can make a team put Sol in the middle of a voice agent that was running a cheaper tier. When the promo ends, the voice feature does not get slower; it gets more expensive, or someone silently routes back to Luna and QA notices tone changes.

Full-duplex Live-1 is a different SKU family from GPT-5.6 Sol. Do not convert a Sol coupon into "Live-1 is 50% off." If you are comparing conversation models, start from the GPT-Live-1 explainer, not from a text-model rate card.

TTS-only demos, including the one on GPTLive1, barely touch Sol. Script length still drives audio cost and listening time; that is a different budget. The voiceover word-count guide is the right tool for that side.

A short verification checklist

When the next "50% off" card hits HN:

1. Open the model card, not only the aggregator page. Does the provider's own list price change, or only a reseller?
2. Note which SKU moved. Sol, Terra, and Luna are not interchangeable, and Live-1 is not Sol.
3. Note who bills you. Direct, gateway, or BYOK through a gateway.
4. Note the end date and whether Fast, Flex, Batch, cache, and long context are in or out.
5. Re-run one golden task (a voice script, a tool-using agent, a long extractive prompt) at the new route before making it the default.

The HN title was good clickbait and a bad specification. OpenAI did not halve Sol. Someone in the routing layer ran a sale. Price the sale as a sale.