LLM API pricing in 2026, normalized and sourced

Last verified 2026-07-14 against each provider's official pricing page (linked per row). Prices change weekly — treat this as a dated snapshot, not a contract.

Most LLM pricing tables have three problems: they go stale within weeks, they compare $ per token when different models tokenize the same text differently, and they're published by vendors with a thumb on the scale. This page fixes all three. Every price below is verified against the provider's own page and dated; each row links its source; and it adds a column no other table publishes — cost per 1,000 English words, normalized for the fact that some models emit ~30% more tokens for identical text.

Disclosure: TierUp is an LLM routing API, so we have a horse in the "route to the cheapest adequate model" race. That's exactly why we keep the reference tables scrupulously vendor-neutral — our thesis only holds if the numbers are honest. Our own view is in one clearly-marked section at the end; everything above it is just the data.

Frontier tier — flagship general models

ModelIn / 1MOut / 1MCache readOut / 1k words*Src
GPT-5.6 SolOpenAI$5$30$0.50$0.0400
Claude Opus 4.8Anthropic · +30% tokenizer$5$25$0.50$0.0433
Claude Fable 5Anthropic · +30% tokenizer$10$50$1$0.0867
Gemini 3.1 ProGoogle · ≤200k ctx; $4/$18 above$2$12$0.20$0.0160

Mid tier — balanced workhorses

ModelIn / 1MOut / 1MCache readOut / 1k words*Src
GPT-5.6 TerraOpenAI$2.50$15$0.25$0.0200
Claude Sonnet 5Anthropic · intro thru Aug 31 → $3/$15 Sep 1 · +30% tokenizer$2$10$0.20$0.0173
Gemini 3.5 FlashGoogle$1.50$9$0.0120
GPT-5.6 LunaOpenAI$1$6$0.10$0.0080

Cheap tier — small / high-volume / open

ModelIn / 1MOut / 1MCache readOut / 1k words*Src
Claude Haiku 4.5Anthropic$1$5$0.10$0.0067
GPT-5.4 miniOpenAI$0.75$4.50$0.0060
Gemini 3.1 Flash-LiteGoogle$0.25$1.50$0.0020
DeepSeek V4 FlashDeepSeek · reasoning-native$0.14$0.28$0.0028$0.0004
Gemini 2.5 Flash-LiteGoogle$0.10$0.40$0.0005

*Cost per 1,000 English words (output) assumes ~0.75 words/token (≈1,333 tokens per 1,000 words); rows marked "+30% tokenizer" use ≈1,733 tokens, per Anthropic's stated ~30% higher token count for the same text on its newest models. This is the apples-to-apples number most tables omit. Prices are standard on-demand list rates; batch (−50%) and cache discounts apply on top.

Five things the raw table hides

  1. The tokenizer tax is real and unpriced everywhere else. Claude Opus 4.8's $5/$25 sticker is effectively ~$6.50/$32.50 per equivalent unit of English, because its tokenizer emits ~30% more tokens for the same text (Anthropic's own note). Compare cost-per-word, not cost-per-token, or you'll mis-rank the newest Claude models.
  2. Output, not input, is your bill. Output runs 5–6× input across the frontier, and reasoning models bill hidden chain-of-thought as output. GPT-5.5 Pro at $180/1M output is ~200× more expensive to emit than DeepSeek V4 Flash at $0.28. The input column you skim is the least important number for agent workloads.
  3. The real chasm is frontier → cheap, not frontier → mid. Sonnet 5 sits at ~40% of Opus 4.8 and GPT-5.6 Terra at ~50% of Sol — the flagship-to-mid gap has collapsed. But DeepSeek V4 Flash is 36× cheaper on input and 107× on output than GPT-5.6 Sol, and Gemini 2.5 Flash-Lite pushes that further. That's why mixed routing typically cuts spend 60–80%.
  4. Caching beats provider-switching for repeated context. A cache hit is 90% off at Anthropic and Google, ~98% at DeepSeek; batch is another 50%. Cached Opus 4.8 input ($0.50) undercuts Gemini 3.1 Pro's uncached input. If you have a big system prompt or RAG context, caching is a bigger lever than the flagship you pick.
  5. For identical open weights, the host swings price up to 5×. The same model can cost 5× more or less depending on who serves it, while aggregators pass through provider prices for a small top-up fee. Buying "the model" is not buying "a price" — the routing layer is where savings live.

What "cheapest" actually means

Sticker input price is the wrong thing to minimize. Your real cost per unit of work is: effective $ = (input tokens × input price × cache factor) + (output tokens × output price), then divide by the words of usable work produced. Three corrections most comparisons skip: normalize for the tokenizer (some models cost ~30% more than sticker), weight output far more than input, and fold in caching/batch. Do that and the "cheapest" column often reorders — and the honest answer for most apps is "different models for different calls," not one winner.

Our take (this is the vendor talking)

The data above says the same thing our product is built on: no single model is the right default, and routing the easy 60–80% of calls to a cheap tier while reserving the frontier for hard ones is where the savings are. You can build that yourself with a couple of API keys and a router. TierUp does it behind one OpenAI-compatible endpoint where you request a quality tier instead of a model name — ~50% under retail, tier-1 free, no card to try at tierup.ai/try. We're early and say so. If you'd rather wire your own routing, the tables above are yours to use.

FAQ

What is the cheapest LLM API in 2026?

On raw list price, the cheapest capable options are the small/open tier: Gemini 2.5 Flash-Lite ($0.10 input / $0.40 output per 1M tokens), Qwen-Turbo, and reasoning-native DeepSeek V4 Flash ($0.14 / $0.28). These run 20–100× below frontier models like GPT-5.6 Sol ($5 / $30) or Claude Opus 4.8 ($5 / $25). But 'cheapest' depends on the task — a cheap model that fails and needs a retry on a frontier model is not cheap. The rational architecture is routing: send easy calls to the cheap tier and hard calls to the frontier tier.

Why is output priced so much higher than input?

Across every frontier model, output costs 5–6× input ($5→$30 for GPT-5.6 Sol; $5→$25 for Opus 4.8; $2→$12 for Gemini 3.1 Pro). Output is the compute-heavy half — each token is generated sequentially. Reasoning/'thinking' models bill their hidden chain-of-thought as output tokens, so a reasoning request can silently emit 3–10× the visible answer length. For agentic or reasoning workloads, the output price is the number that decides your bill, not input.

What is the 'tokenizer tax' and why does it change the ranking?

Anthropic's newest models (Opus 4.7+, Fable 5, Sonnet 5) use a tokenizer that produces about 30% more tokens for the same English text — Anthropic states this on its own pricing page. So Claude Opus 4.8's $5/$25 sticker is effectively closer to $6.50/$32.50 per equivalent unit of text versus a model on an efficient tokenizer. Any table that compares $/token without normalizing for this overstates how cheap the newest Claude models are. That is why the table below also shows cost per 1,000 English words.

Do caching and batching matter more than switching providers?

For most real workloads, yes. A prompt-cache hit costs 10% of input price at Anthropic and Google (about 2% at DeepSeek); the Batch API takes 50% off input and output at OpenAI, Anthropic, Google, and Mistral. Stacked, that is roughly 25% of list price. Cached Opus 4.8 input ($0.50) is cheaper than Gemini 3.1 Pro's uncached input — so enabling caching often beats hunting for a marginally cheaper flagship.

Is an aggregator/router more expensive than going direct?

Generally no. OpenRouter, for example, passes through each provider's per-token price and charges only a small fee at credit top-up, not per call. Meanwhile the same open weights can cost up to ~5× more or less depending on which host you pick. So multi-provider access costs little, and the real savings live in choosing the right model per request — which is what a routing layer automates.

Related: LLM API cost calculator · How tier routing works · Pick a tier, not a model · vs OpenRouter · Cheapest AI API in 2026

Featured on Fazier