Where to get Claude and GPT cheaper: official list, caching, batch and resellers
The same Claude Opus 5.5 costs different amounts depending on where you buy it. Official pricing is only the baseline. Below: what discounts are fair (caching, batch, vendor promos) and how to tell a good reseller from a scheme.
Official savings
Anthropic prompt-caching reads cost $0.20 vs $4 per 1M on Opus 5.5. Batch at OpenAI and Anthropic is about −50% with hourly latency. Promo tiers — Google's Gemini 3.8 Flash until end of 2026, OpenAI's Sol until end of November — are temporary but real −50%. Well-architected apps save up to 4× with no reseller at all.
Why aggregators are cheaper
Token resellers earn on three things: volume contracts with vendors, cache hits on shared infrastructure, and geo-arbitrage of list prices. Legal multi-account aggregators (DeepInfra, Novita, Polza) sell the same Claude Opus 5.5 at a 10–30% discount or a small markup.
Where risk starts
“Claude worth $1000 for $30” usually means stolen keys, abused regional promos, vendor ToS violations, or a service that dies in a month. Red flags: no public price list, no jurisdiction, crypto-only top-ups, Telegram-only support. For production the failure cost is data and downtime, not the discount.
Reseller checklist
1) Public pricing page or API — we crawl them daily. 2) GET /models answers without a key — our free checker verifies this. 3) Minimum top-up and payment methods. 4) Stability: our monitor probes endpoints every 30 minutes. 5) Transparency: demo listings should be labelled. None of these guarantee safety; their absence guarantees trouble.
How we rank
TokenRate is an independent price aggregator. We do not sell model access; provider links are sponsored (rel="sponsored"), but sponsorship never changes sort order — ranking is built on blended cost for your workload and monitoring data. Quotes without a confirmed source are marked demo.
Published 2026-11-11. Prices are vendor list prices on the publish date; live quotes are in the catalogue.