Cloudflare's 50% Off GPT-5.6 Sol Promo Lands at OpenAI's Regular List Price
Cloudflare is promoting GPT-5.6 Sol through AI Gateway at what it calls half price, running through September 18, 2026. No promo code, no application. Point your requests at the gpt-5.6-sol model and the rate applies automatically.
The promo rates are $2 per million input tokens, $10 per million output tokens, and $0.25 per million cache read tokens. Cloudflare gives the regular rates as $4, $20, and $0.50.
Those numbers are worth holding up against the source. OpenAI’s published price for GPT-5.6 Sol is $2.00 per million input tokens and $10.00 per million output. OpenRouter lists the same figures, with cache reads at $0.20. So the discounted price here is the standard price everywhere else, and the promotional cache read rate is actually a few cents above what OpenAI charges.
That leaves the question of what the $4 and $20 refer to. The obvious reading is that Cloudflare’s unified billing normally carries roughly a 2x markup over the underlying provider rate, and the promotion waives it. The alternative, that the comparison is against Sol’s max reasoning configuration, doesn’t hold up either. Independent measurement puts that tier at $5 input and $30 output, which matches neither the promo nor the stated regular price.
The exclusion is the point
Read the fine print and the offer clarifies itself. The discount applies only to unified billing customers. Bring your own key and you get nothing. Cloudflare wants you loading credits onto its platform rather than passing your OpenAI key through it, and the promotion is priced to make that switch feel free.
That’s a customer acquisition offer for a billing relationship, not a discount on a model. Which is a legitimate thing to run. It just isn’t what the headline says, and anyone budgeting off the promo copy will be surprised when the thirty days end and their per-token cost doubles rather than returns to par.
What the gateway is actually worth
Strip out the pricing claim and there’s still a product underneath. A single API surface across thirty or so providers, response caching, rate limiting, and request logging. Swapping models without touching integration code has real value once you’re running more than one and comparing them seriously. So does having usage logs in one place instead of three provider dashboards.
For production traffic at volume, that operational layer is often worth a markup on its own terms. For someone making a few hundred calls a week, it’s a middleman between you and an API you could call directly in an afternoon.
Measure the thing that matters
If you are evaluating Sol against what you’re already running, per-token rates are the wrong unit anyway. Sol is a reasoning model, and reasoning models spend output tokens generating thinking before they answer. On agentic work, the volume of tokens consumed varies by multiples between models, which swamps small differences in the posted rate. Cost per completed task is the number to track.
Latency deserves the same scrutiny. Benchmarks put Sol’s throughput below the median for its price class, with a long wait before the first token appears at higher reasoning settings. That’s tolerable for a background job and unpleasant in anything interactive.
Test it if it fits your stack. Just don’t test it because you think it’s half price.