The LLM API pricing comparison matrix: OpenAI vs Anthropic vs Gemini vs Grok (and the rest)

This is an LLM API price comparison across the eight providers that matter for production traffic — OpenAI, Anthropic, Google, xAI, DeepSeek, Alibaba (Qwen), Mistral, and Meta (Llama) — organized by tier so you compare flagship against flagship and workhorse against workhorse. The sticker prices are one screen of numbers; the caveats underneath them are what actually decide the winner, so both are here.

All prices are USD per 1M tokens. "Cached input" is the discounted rate for reading a previously cached prompt prefix. Where a value is not published in the source dataset it is shown as "—".

Flagship tier

The frontier models, plus each provider's headline offering. Note the spread between Claude Fable 5 at the top and DeepSeek V4 Pro at the bottom — more than 20× on input, over 50× on output. These are not substitutes on quality, but the price gap is real.

Read the flagship tier as three bands: frontier Western models ($5 input and up — GPT-5.6 Sol, Claude Opus 5, Claude Fable 5), value flagships ($1.50–$2.50 input — Gemini 3.1 Pro, Grok 4.5, Qwen-Max, Mistral Medium 3.5), and DeepSeek V4 Pro alone at the bottom on price. The caveats section below explains why the cheapest rows come with strings attached.

Gemini vs Grok: the two value flagships

Gemini 3.1 Pro and Grok 4.5 both list $2.00 input, so head-to-head the difference is output: Grok 4.5 is $6.00 against Gemini 3.1 Pro's $12.00 — half the output cost — which makes Grok materially cheaper for generation-heavy work. Two caveats reorder that. Gemini 3.1 Pro is preview-priced and can change at general availability, and it adds a long-context surcharge above 200K tokens (input 2×, output 1.5×). Grok 4.5 doubles both input and output above 200K tokens and tops out at a 500K window versus Gemini's 1M. Under 200K tokens and output-heavy, Grok wins on price; if you need the 1M window, Gemini is the one that has it.

Workhorse (mid) tier

This tier carries most production traffic — good enough for the majority of tasks, priced for volume. It is also the most competitive, with a 5× spread on input alone.

Claude Sonnet 5's $2 / $10 is an introductory price that expires 2026-08-31, after which it lists $3 / $15. Grok 4.3 is the standout on output at $2.50, a fraction of GPT-5.4's $15.00 — for chat and generation workloads where output dominates, it is the cheapest capable mid-tier model on a US-hosted API. Qwen-Plus bills output at $4.80 in thinking mode; the $1.60 shown is the standard rate.

Reasoning models: you pay for the thinking

Dedicated reasoning models spend extra output tokens "thinking" before they answer, so their output rate — and how many output tokens they burn — matters more than input. OpenAI sells these as a separate line; Anthropic and Google fold extended thinking into their main models, where the thinking tokens bill at the standard output rate.

o4-mini and o3 are the practical reasoning workhorses; o3-pro and GPT-5.5 Pro are maximum-effort variants whose output rates ($80–$180) make them expensive to let run long. Because reasoning models emit far more output tokens than a normal chat turn, their effective cost per task can exceed a flagship's even when the per-token rate looks moderate.

The budget tier, in one line

Below the workhorse tier sits a field of sub-dollar models — GPT-5.6 Luna, GPT-5.4 nano, Gemini Flash-Lite, DeepSeek V4 Flash, Qwen-Flash, Mistral Small 4, Llama 4 Scout, and legacy GPT-5 nano — where the cheapest options run 10–15× less than the priciest "budget" model. That tier has its own ranked matrix; see the cheapest-LLM guide linked below.

The caveats that reorder the ranking

Every cheap row above carries a footnote. Read these before you rank on price:

• ◦Preview pricing: Gemini 3.1 Pro is priced as a preview and can change at general availability.
• ◦Off-peak vs peak: DeepSeek's rates are off-peak; the peak rate (09:00–12:00 and 14:00–18:00 Beijing time, UTC+8) is 2× on both V4 Pro and V4 Flash.
• ◦Long-context surcharge: Gemini 3.1 Pro doubles input and adds 50% to output above 200K tokens; Grok 4.5 doubles input and output above 200K.
• ◦Promotions: Qwen-Max may carry a limited 50%-off promo (about $1.25 / $3.75) — confirm it is active before you rely on it.
• ◦Third-party hosting: Meta does not sell Llama tokens directly; the Llama prices shown are Together AI's serverless rates, and other hosts differ.
• ◦Tokenizer skew: per-token prices are not comparable across tokenizer generations — Anthropic's newest models emit roughly 30% more tokens for the same text. Compare cost per task, not price per token.

Which should you pick

Match the model to the job, not to the top of a leaderboard:

• ◦Cheapest capable flagship: DeepSeek V4 Pro off-peak ($0.435 / $0.87) sits an order of magnitude below the Western flagships — if its peak-hours 2× and China-hosted API fit your constraints.
• ◦Best value flagship on a US-hosted API: Grok 4.5 ($2.00 / $6.00) for output-heavy work; Gemini 3.1 Pro ($2.00 / $12.00) when you need the 1M window and can accept preview pricing.
• ◦Frontier quality: GPT-5.6 Sol and Claude Opus 5 ($5.00 input each); Claude Fable 5 ($10.00 / $50.00) when you want Anthropic's most capable model and cost is secondary.
• ◦Workhorse default: Claude Sonnet 5 on its intro price or Gemini 3.6 Flash; Grok 4.3 ($1.25 / $2.50) is the cheapest of the mid tier for output-heavy traffic.

What is the cheapest LLM API?

On sticker price, DeepSeek V4 Pro off-peak ($0.435 / $0.87) is the cheapest capable flagship, and in the budget tier legacy GPT-5 nano ($0.05 / $0.40) is the cheapest overall. But "cheapest" depends on your input:output ratio, whether you hit peak-hours surcharges, and how your text tokenizes — price your own workload rather than trusting a single headline number.

Are the prices the same across providers for the same quality?

No. At the flagship tier the input rate ranges from $0.435 (DeepSeek V4 Pro) to $10.00 (Claude Fable 5) — more than a 20× spread — and the models are not interchangeable on quality, latency, tool support, or data residency. Price is one axis; the caveats above are the others.

Why do some models cost more above a token threshold?

Serving very long contexts is more expensive, so several providers add a long-context surcharge. Gemini 3.1 Pro doubles input and adds 50% to output above 200K tokens; Grok 4.5 doubles both above 200K. If your prompts are long, the headline rate understates your bill — model the surcharge tier explicitly.

How current are these prices?

They are list prices as of 2026-08-03 and they move often — Claude Sonnet 5's intro price expires 2026-08-31, and OpenAI cut GPT-5.6 Luna sharply in late July 2026. Re-verify against each provider before you commit a budget.