2026 model API pricing review: who actually got cheaper

Three structural changes this generation

Change one: mid-tier prices kept falling

Over the past year or so, every vendor's mid tier (the Claude Sonnet, mid GPT and Gemini Pro level) has continued to get cheaper. The strategy is plain: move volume on the mid tier, take the premium on the flagship tier.

What that means for you: route almost everything through the mid tier by default. Only step up to a flagship when the task genuinely needs deeper reasoning — flagship output prices are usually more than three times the mid tier's.

Change two: cache pricing became the dominant cost variable

This is the most commonly overlooked point. Most vendors now give a substantial discount on cache-hit input tokens, and some charge as little as 10% of the standard input price.

What that means: if your workload is "fixed system prompt + a long document + several follow-up questions", your real cost can be a third of what the standard price implies.

Conversely, if your prompt is different every time, you get none of that benefit and your cost is the standard price. Same model, same token count, and the unit cost can differ several times over between the two workloads.

The conclusion: when costing, do not look at the price table alone. Work out your cache hit rate first.

Change three: time-of-day pricing and promotional windows

Some vendors have introduced time-of-day pricing (off-peak at roughly half the peak rate). Others cut prices via a "promotional rate that reverts at a set date".

The implication for budgeting: never treat a promotional price as your baseline. Build annual budgets on the standard price and count promotional savings as upside. Budget on the promotional price and your bill will double when it expires.

How vendors differ in style

What this means for choosing a relay service

1. A single flat "rate" is getting harder to sustain. Vendors cut prices on different schedules, so a service advertising "0.5 across the board" may be expensive for DeepSeek and cheap for Claude. Look at rates per family, not one number.
2. Whether cache discounts pass through is now a key differentiator. The vendor charges 10% on cache hits — does the relay bill you the discounted amount or the standard one? On a high-cache workload this matters a lot.
3. Promotional expiries show up as price rises. If a service's low price rests on the vendor's promotional price, it will have to raise its own price when that ends. Judging long-term cost from the "registered rate" will mislead you then.

Three practical recommendations

• Budget on the mid tier, keep the flagship as a fallback. Use a mid-tier model for 90% of requests and hold a flagship for the difficult 10%.
• Optimise the prompt structure before comparing prices. Raising your cache hit rate usually saves more than switching to a service that is 10% cheaper.
• Compute real cost, not the nominal rate. Put your own input/output ratio into the calculator on our rankings page.

The full vendor pricing comparison is in the model price list.

Promotional deadlines and time-of-day rules are marked there; for long-term budgeting use the standard price.