Data study · 2026

The State of AI API Pricing in 2026

We track 30 models from 10 providers and re-check every rate against the provider's own pricing page. Here is what the numbers say — the spread, the mixes where the cheapest model flips, and the discount almost nobody prices in.

Finding 1: the spread is 147×, and it never shrinks

On a 10M-input / 3M-output month, the cheapest model we track (Qwen-Flash) costs $1.70 and the most expensive (Claude Fable 5) costs $250. The category average is $42.4.

The interesting part is what happens as you scale: nothing. Per-token pricing is linear, so there are no volume tiers and no bulk discounts — the gap between cheapest and priciest is the same multiple whether you spend $2 a month or $5,000. That is the opposite of tiered SaaS, where the cheapest tool changes as you grow. Picking wrong here does not get cheaper to fix later; it scales linearly with you.

Finding 2: your token mix picks the winner, not your volume

Output tokens cost several times more than input on every model here, and the multiple differs by provider. That means two models can swap places purely on the shape of your workload — a document-summarising app and a chat agent burning the same total tokens will not have the same cheapest option. We found 15 pairs that trade the lead at a realistic output share.

A concrete case: Gemini 3.1 Flash-Lite is cheaper than Qwen-Plus while output is under 33% of your tokens. Above that line it reverses. Neither is simply "cheaper" — the answer depends on what you are building. See that comparison →

Finding 3: cache discounts vary 8× and reorder the board

Most providers discount cached input reads to about 10% of their standard input rate, but the range is far wider than that suggests. DeepSeek V4 Flash bills cached reads at 3.2% of its own input rate; Grok 4.6 stops at 25%. For anything that resends a large fixed context — RAG, agent loops, a long system prompt — that gap moves the total more than the headline rate does, and it changes which model is cheapest.

Cache reads only. Some providers also bill a cache write premium, which our figures exclude — it amortises across reuse but not at low reuse. You can model your own hit rate on the AI cost calculator.

Finding 4: 2026 has moved in both directions

Price cuts are rare in software, so it is worth noting that this category produced moves both ways in the same month. We log every change we verify against the live pricing page:

• July 2026 · reported
  DeepSeek roughly triples V4 pricing and moves to peak/off-peak billing

  DeepSeek V4 Flash went from $0.14 in / $0.28 out to $0.44 in / $1.32 out per 1M at peak, and V4 Pro from $0.435 / $0.87 to $1.32 / $3.96 — output rates rose about 4.6x. DeepSeek also introduced time-of-day billing: peak is 01:00-04:00 and 06:00-10:00 UTC, with off-peak (the other 17 hours) at half price. Cache-hit input rose in step, taking V4 Pro's cached discount from roughly 0.8% of its input rate to about 3.3%. DeepSeek remains cheap by frontier-lab standards, but it is no longer an order of magnitude below the field.

• July 2026 · reported
  OpenAI cuts GPT-5.6 Luna five-fold and drops Terra by 20%

  GPT-5.6 Luna fell from $1.00 in / $6.00 out to $0.20 / $1.20 per 1M — a five-fold cut that moves current-generation OpenAI quality into budget-tier pricing. GPT-5.6 Terra dropped from $2.50 / $15.00 to $2.00 / $12.00. Sol, GPT-5.4 Mini and GPT-5.4 Nano were unchanged. Price cuts are rare enough in this category to be worth noting: most 2026 movement has been upward.

See the full change log →

Method

Every rate is taken from the provider's own pricing or model documentation, never from an aggregator, and re-checked on a stated date (July 2026). Where a provider tiers by prompt length, time of day or region, we publish the rate that cannot understate a bill and record the alternative in a per-model note. Figures exclude batch discounts unless stated. Full detail in our methodology.