OpenAI’s flagship model is on sale. From August 17 through September 18, 2026, Vercel is running a 50% discount on GPT-5.6 Sol for every request that runs through its AI Gateway. Input drops from $5.00 to $2.50 per million tokens. Output drops from $30.00 to $15.00. No promo code, no plan change, no code change: if your requests already route through the Gateway, the discount is applied automatically.
This matters because Sol was the one tier OpenAI left untouched in its July 30 price cut. Luna got 80% cheaper and Terra got 20% cheaper, but Sol stayed at $5/$30. For one month, the flagship costs close to what Terra costs at list price.
If you want to take advantage of the window, this post covers the exact numbers, the fine print that decides whether you qualify, and how to point your existing OpenAI client (or an API client like Apidog) at the Gateway endpoint in about two minutes.
The exact numbers
Vercel discounts all three Sol service tiers, not only the default one:
| Service tier | Normal price (per MTok, input/output) | Promo price (per MTok, input/output) |
|---|---|---|
| Default | $5.00 / $30.00 | $2.50 / $15.00 |
| Flex | $2.50 / $15.00 | $1.25 / $7.50 |
| Priority (fast mode) | $10.00 / $60.00 | $5.00 / $30.00 |
The discount runs from August 17 to September 18, 2026 and covers cached tokens, cache writes, long-context requests, and all regions.
Two details worth noticing in the table:
- Flex at $1.25/$7.50 is the quiet headline. If your workload tolerates slower completion (batch analysis, offline evals, doc generation), you’re getting flagship-quality output at under half of Terra’s list price.
- Priority at the promo rate ($5/$30) costs exactly what default Sol costs normally. For a month, “fast mode Sol” is free in the sense that you pay yesterday’s standard price for it.
For reference, OpenAI’s own API pricing still lists Sol at $5/$30, and that’s what you pay going direct or using your own key. We keep a full breakdown of the current tier structure in our GPT-5.6 pricing guide.
What Vercel AI Gateway is (30-second version)
AI Gateway is Vercel’s unified model API. One API key gets you hundreds of models across OpenAI, Anthropic, Google, and others, with automatic failover between providers, spend monitoring, and, per Vercel, zero markup on tokens: you normally pay exactly what the provider charges.
That last point is what makes this promo unusual. The Gateway isn’t discounting its fee (there isn’t one on tokens). It’s selling Sol tokens below OpenAI’s own list price. During the window, the Gateway is the cheapest legitimate way to call Sol.
The Gateway speaks three dialects:
- The AI SDK natively (Vercel’s TypeScript toolkit, covered in our Vercel AI SDK guide)
- OpenAI Chat Completions and Responses, at an OpenAI-compatible base URL
- Anthropic Messages, for clients built against Claude
The OpenAI-compatible surface is the one most teams will use, because it means your existing OpenAI client works with a one-line change.
How to use it
With the AI SDK
If you’re on Vercel’s AI SDK, the model string is all you need:
import { streamText } from 'ai';
const result = streamText({
model: 'openai/gpt-5.6-sol',
prompt: 'Investigate the failing tests and open a PR with a fix.',
});
With any OpenAI client
Point the client at the Gateway’s base URL and swap the key:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.getenv("AI_GATEWAY_API_KEY"),
base_url="https://ai-gateway.vercel.sh/v1",
)
response = client.chat.completions.create(
model="openai/gpt-5.6-sol",
messages=[{"role": "user", "content": "Summarize this OpenAPI diff."}],
)
With curl
curl -X POST "https://ai-gateway.vercel.sh/v1/chat/completions" \
-H "Authorization: Bearer $AI_GATEWAY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.6-sol",
"messages": [{"role": "user", "content": "Hello, Sol."}]
}'
Note the model id format: openai/gpt-5.6-sol, with the provider prefix. A plain gpt-5.6-sol returns a 404 from the /models route. If you’re new to the model family itself, start with our GPT-5.6 API guide.
Testing the endpoint before you commit traffic
Before you flip production traffic to a new base URL, it’s worth verifying the Gateway behaves identically to the OpenAI endpoint you’re leaving: same response shape, same tool-calling behavior, same streaming chunks. This is a five-minute job in Apidog: create an environment variable for the base URL, add your Gateway key as a vault secret, and run your existing OpenAI request collection against https://ai-gateway.vercel.sh/v1. Because the Gateway implements the Chat Completions spec, your assertions on response structure should pass unchanged, and if they don’t, you’ve found the incompatibility before your users did.
The same setup makes cost comparison concrete: send one prompt to openai/gpt-5.6-sol, openai/gpt-5.6-terra, and openai/gpt-5.6-luna side by side, diff the outputs, and check the token counts in the usage block. Download Apidog free if you want to run that experiment; it takes one collection and three environment switches.
The fine print that decides whether you save anything
The changelog is short, but three conditions matter:
- Direct Gateway requests only, not BYOK. If you’ve configured Bring Your Own Key so the Gateway forwards requests on your OpenAI account, you bill at OpenAI’s rates and the promo doesn’t apply. To get the discount, your requests must run on Vercel’s keys, meaning the Gateway handles billing.
- It ends September 18. Budget forecasts built on $2.50/$15 will be wrong by 2x in October. If you’re doing a one-off (a large backfill, an eval sweep, a dataset relabel), the window is a gift. If you’re setting a permanent unit-economics baseline, use list price.
- Sol’s list price didn’t change. OpenAI hasn’t cut Sol. If you read this as a signal about where Sol pricing is heading, that’s speculation; the only committed number is Vercel’s, and only until September 18.
When paying Sol prices is worth it at all
The discount only matters if Sol is the right tier for the job. Sol’s edge over Terra and Luna shows up on the hard end of the distribution: long-horizon agent tasks, complex tool orchestration, and, per recent third-party testing, vision.
Roboflow’s benchmarks published this week found Sol hit 46.2 mAP@50 on object detection, up from 13.8 for GPT-5.5, and 73% on object counting. At normal prices that capability costs about 2.5 cents per image. During the promo window it’s closer to 1.2 cents, which moves “run Sol over our whole document archive” from a budget request to an afternoon experiment. Our Sol benchmarks roundup has the broader numbers if you’re deciding between tiers.
The honest counterpoint from the same Roboflow data: Gemini 3.5 Flash still beats Sol on detection while costing 0.8 cents per image. A discount narrows the gap; it doesn’t erase it. Run your own eval on your own data before you commit either way.
The play for API teams
If you build or test APIs for a living, here’s the practical checklist for the next month:
- Front-load Sol-heavy work. Eval sweeps, synthetic test-data generation, spec-review passes, doc drafts: anything you’d run through Sol eventually, run before September 18 at half cost.
- A/B the tiers while the flagship is cheap. The usual argument against comparing Sol to Terra on your workload is that the comparison itself costs money. Right now Sol is priced near Terra, so the experiment is nearly free. Keep the results; they stay valid after the promo ends.
- Watch your usage dashboard. The Gateway’s spend monitoring shows per-model cost, so you’ll see the discounted rate reflected there. If your numbers still show $5/$30, check whether BYOK is silently routing you to your own OpenAI account.
FAQ
Do I need to change my code to get the discount? No. If your requests already run through AI Gateway on Vercel’s billing, the discount applies automatically from August 17. New users need a Vercel account and a Gateway API key, then point any OpenAI-compatible client at https://ai-gateway.vercel.sh/v1.
Does the discount apply to cached tokens and long-context requests? Yes. Vercel’s changelog states the 50% off covers cached tokens, cache writes, long-context requests, and all regions.
Does it apply if I bring my own OpenAI key? No. BYOK requests bill at OpenAI’s provider rates ($5/$30 for default Sol) and are excluded from the promo.
Is this the same as OpenAI cutting Sol’s price? No. OpenAI’s list price for Sol is unchanged. This is Vercel discounting Gateway traffic for one month. The last change OpenAI itself made was the July 30 cut to Luna and Terra, which skipped Sol.
How do I verify I’m getting Sol and not a fallback model? The Gateway supports provider routing and fallbacks, so pin the model explicitly as openai/gpt-5.6-sol and check the model field in the response body. In Apidog you can add an assertion on that field to every request in your collection, so a silent fallback fails the test instead of quietly billing you for a different tier. If you’re still choosing a tier, our what is GPT-5.6 Sol explainer covers where the flagship earns its price.
Worth grabbing, with a calendar reminder
A 50% discount on a flagship model with no markup underneath it is rare, and the conditions are clean: route through the Gateway, don’t use BYOK, done by September 18. Front-load the expensive work, run the tier comparison you’ve been putting off, and set the reminder for when the math reverts.
And before any of that traffic moves, test the endpoint like you’d test any other API dependency. Apidog gives you the environment switching, secret storage, and response assertions to swap base URLs without guesswork. Try it free and point your first collection at the Gateway.
