Gateway price check before your weekend AI bill doubles

You built a small bot over the weekend, and the early days feel almost free. Then the gateway changes a rate, and the bill jumps. The model name is part of the price; the gateway sets the rest, and it can change without your code changing.

With a near-zero budget, the question is what the gateway will charge per million tokens, when that rate ends, and what happens after. A short check can stop a surprise.

The gateway is the price switch

Vercel's AI Gateway applied its 50% GPT-5.6 Sol discount to OpenAI's newly lowered list price through September 18. OpenAI reduced GPT-5.6 Sol list pricing on 21 August 2026 to $4 per million input tokens and $20 per million output tokens, with the rate stated to hold at least through 21 November 2026. On Vercel's AI Gateway, the default GPT-5.6 Sol rate was $2.00 input and $10.00 output per million tokens, down from a prior gateway price of $2.50/$15.00. The GPT-5.6 Sol model ID remained unchanged on Vercel's AI Gateway, so existing requests would bill at the new price automatically.

Cloudflare announced a 50% GPT-5.6 Sol discount on 19 August 2026, with promotional rates of $2.50 input and $15 output per million tokens against a standard price of $5 input and $30 output. Cloudflare's GPT-5.6 Sol promotion was scheduled to end on 18 September 2026, after which requests would return to standard pricing. Cloudflare's GPT-5.6 Sol promotion applied to Unified Billing and not to BYOK.

For the same model, the gateway can change the effective rate, the promo end date, and the post-promo price. The check asks for the effective rate, not the headline label.

The check is the weekend habit

Keep this list next to your deploy notes. Each item catches a different way the bill can move. Put the promo end date in your calendar before the bill rises.

• Confirm the effective input and output price per million tokens for the exact model you call.
• Find the promo end date, not just the discount label.
• Note the post-promo standard rate.
• Check whether the model ID changed; if it did not, old code may bill at the new price automatically.
• Set a budget cap or alert before you deploy.

Run the list before you push code, and again when a promo date passes.

Small changes keep the bill small

Speed can be a separate charge. OpenAI's API Fast service tier charges twice the standard price for up to 2.5 times faster processing, and the tier primarily applies to the Sol model. If your weekend bot does not need faster answers, the standard tier is the cheaper default.

Anthropic's batch processing option reduces both input and output costs by 50% for work that can wait an hour. Keep the repeated part of your prompt in cache. Prompt cache hits cost one-tenth of the base input price. Anthropic charges $10 for each 1,000 web searches in addition to token costs, making ten searches per day about $3 per month.

For a triage job processing 200 emails per day with roughly 1,200 input tokens and 200 output tokens per call, Anthropic Haiku 4.5 costs about 44 cents per day or $13 per month, Sonnet 5 costs about $26 per month, and Opus 5 costs about $66 per month. A cap makes that range clear before it becomes a surprise.

Anthropic's Sonnet 5 was in an introductory pricing phase at $2/$10 per million tokens, with the price scheduled to rise to $3/$15 per million tokens after August 31. If your weekend project depends on that model, the end date belongs in the same note as the gateway promo.

If the answer is unclear, pause the deploy and ask the gateway's docs for the exact rate. You can still ship the weekend project with a cap and a calendar date.