LLM API Pricing Compared: Cost per 1M Tokens

Anthropic prices Claude on a strict three-rung ladder: Opus 4.8 at $30 combined, Sonnet 5 at $12, Haiku 4.5 at $6. It is the most consistent lineup in the dataset — output costs exactly 5x input at every tier ($5/$25, $2/$10, $1/$5) — which makes cost forecasting unusually predictable: whatever a task's input/output mix, moving one rung down cuts the bill by half or more.

Claude Opus 4.8 · $5 In / $25 Out per 1M Tokens

Opus 4.8 mirrors GPT-5.5's $5 input rate but charges $25 for output instead of $30, landing at $30 combined. That $5-per-million output gap compounds fast: a service generating 1B output tokens a month saves $5,000 monthly on Opus at identical volume. For agentic coding and long-form reasoning — the workloads flagship pricing is aimed at — that makes Opus the cheaper of the two frontier options on every generation-heavy job.

Claude Sonnet 5 · $2 In / $10 Out per 1M Tokens

Sonnet 5's $12 combined undercuts its closest OpenAI rival, GPT-5.4 ($17.50), by $5.50 — about 31% — while keeping input at just $2 per 1M tokens. It is the tier most coding assistants and production chat products actually run on: capable enough for multi-file code edits, priced low enough to serve as a default rather than an escalation.

Claude Haiku 4.5 · $1 In / $5 Out per 1M Tokens

Haiku 4.5 closes the ladder at $6 combined. Against other providers' light tiers it reads mid-priced — above GPT-5.4-mini ($5.25) and well above Gemini 2.5 Flash ($2.80) — but it keeps the same 5x output ratio as its Opus and Sonnet siblings. Teams already on Claude use it for triage, moderation, and short-answer endpoints where a $1 input rate keeps high-frequency calls affordable.

Google prices Gemini a full class below the other US flagships: top-tier 2.5 Pro costs $11.25 combined — roughly a third of Claude Opus 4.8 — while Flash-lite bottoms out at $0.50. The trade hiding in the lineup is the output ratio: Gemini output runs 4x to more than 8x its input price, so the cheap headline rates favor read-heavy workloads over write-heavy ones.

Gemini 2.5 Pro · $1.25 In / $10 Out per 1M Tokens

At $11.25 combined, Gemini 2.5 Pro is the cheapest flagship in the comparison — but its 8x output-to-input ratio is the highest of any frontier model here. The pricing rewards a specific shape of work: long-context reading, document analysis, and RAG over large corpora, where inputs dwarf outputs and the $1.25 input rate does the heavy lifting. Flip the mix toward generation and the $10 output rate still undercuts Opus ($25) and GPT-5.5 ($30), but by less than the input numbers suggest.

Gemini 2.5 Flash · $0.30 In / $2.50 Out per 1M Tokens

Flash lands at $2.80 combined — 47% cheaper than GPT-5.4-mini ($5.25) and less than half of Claude Haiku 4.5 ($6). It shares a $2.50 output price with Grok 4.3 while charging under a quarter of Grok's $1.25 input rate. For high-throughput production endpoints — chat, summarization, structured extraction — Flash is the price-performance anchor most teams benchmark everything else against.

Gemini 2.5 Flash-lite · $0.10 In / $0.40 Out per 1M Tokens

Flash-lite's $0.10 input rate is the lowest of any model in the dataset, and its $0.50 combined cost is beaten only by DeepSeek V4 Flash ($0.42). At these prices token math nearly disappears from the meeting agenda: a billion input tokens costs $100. That makes it the default candidate for firehose workloads — log triage, spam filtering, bulk labeling — where per-call quality demands are modest but volume is enormous.

xAI is the only provider in this comparison listing a single price point: Grok 4.3 at $1.25 input and $2.50 output per 1M tokens. There is no flagship-to-nano ladder to route across — which simplifies integration to one decision, but leaves developers without an in-house budget fallback when volume spikes.

Grok 4.3 · $1.25 In / $2.50 Out per 1M Tokens

Grok 4.3's $3.75 combined places it between the mini and mid tiers — above Gemini 2.5 Flash ($2.80), below GPT-5.4-mini ($5.25). The standout number is its output multiple: at 2x input, it carries the lowest output premium of any US-provider model here, against 5x at Anthropic, 6x at OpenAI, and up to 8x at Google. For generation-heavy jobs — long answers, content drafting, verbose code — that ratio makes Grok cheaper in practice than its combined rank suggests.

It also matches Gemini 2.5 Pro's $1.25 input rate exactly while charging a quarter of its $10 output price. The catch is scope: with a single listing, teams that outgrow Grok's quality ceiling or need a sub-dollar tier must cross providers rather than move down a rung — a real integration cost that the per-token price doesn't capture.

DeepSeek's two listings — V4 Pro at $1.305 combined and V4 Flash at $0.42 — sit at the very bottom of the ladder. Both hold output at exactly 2x input, and even the Pro tier undercuts OpenAI's cheapest model, GPT-5.4-nano ($1.45). This is where the roughly 83x top-to-bottom spread of the whole comparison is anchored.

DeepSeek V4 Pro · $0.435 In / $0.87 Out per 1M Tokens

V4 Pro prices its premium tier at $1.305 combined — less than 4% of GPT-5.5's $35, and cheaper than every other provider's budget model except Gemini 2.5 Flash-lite ($0.50). For teams whose evaluations show acceptable quality on their task, the arithmetic is blunt: a workload costing $10,000 a month on GPT-5.4 reruns here for roughly $750 at the same token volume — with the usual caveats around latency, hosting jurisdiction, and compliance review.

DeepSeek V4 Flash · $0.14 In / $0.28 Out per 1M Tokens

V4 Flash defines the floor: $0.42 combined, roughly 1/83rd of GPT-5.5. A pipeline pushing 1M input and 1M output tokens a day costs $12.60 a month here versus $1,050 on GPT-5.5. At that spread the question stops being which model is best and becomes which tasks genuinely need frontier reasoning — because everything that doesn't can run at prices that barely register on a cloud bill.

Token costs are the primary variable expense for AI products. For anything built on hosted LLM APIs, the per-token rate multiplied by monthly volume defines gross margin — and with rates spanning $0.42 to $35 per combined 1M tokens, two teams shipping the same feature can carry cost structures 83x apart. That is why pricing tables are now required reading for founders, CFOs, and engineering leads alike.

Who Pays per Token · Builders, Not Subscribers

Consumer subscriptions like a $20-a-month chatbot plan hide token economics behind a flat fee. API customers get no such buffer: every startup embedding an LLM in its product, every enterprise team running document pipelines, and every agent framework calling a model in a loop pays for exactly the tokens it moves. For these builders — SaaS products with AI features, coding assistants, support bots, internal automation — the pricing table is the cost of goods sold.

That is also why per-token rates matter more than subscription prices for this audience: a subscription caps one user's cost, while an API bill scales with every customer, every request, and every retry. A support bot answering 100,000 tickets a month at roughly 2,000 tokens each moves 200M tokens — about $1,750 on GPT-5.4 versus $280 on Gemini 2.5 Flash for the same traffic.

Where Token Volume Explodes · Agents, RAG, Pipelines

The workloads growing fastest are also the hungriest. Agent loops re-read their context on every step, RAG systems stuff retrieved documents into each prompt, and batch pipelines process entire archives — token volume compounds in ways a human chat session never does. A single agentic coding task can burn several hundred thousand tokens before it finishes, so the difference between a 2x and an 8x output premium starts deciding the architecture, not just the invoice.

The upside is that the market now offers substitutes at every tier. Five providers field 13 models across flagship, mid, mini, and nano classes, so buyers can arbitrage capability against cost — GPT-5.4-mini against Gemini 2.5 Flash, Claude Sonnet 5 against GPT-5.4 — in ways that were impossible when only one or two frontier models existed.

How to Read the Numbers · One Combined Metric

Providers quote input and output prices separately, which hides true cost — especially with output premiums ranging from 2x at DeepSeek and xAI to 8x at Google. Combining 1M input + 1M output tokens into one number levels the field for ranking, even if individual workloads skew toward one side. Read-heavy systems like RAG and long-document summarization should weight the input column; generation-heavy systems like code and drafting should weight output — and the ranking can flip between the two.

LLM API pricing now spans roughly 83x from top to bottom — $35 for GPT-5.5 down to $0.42 for DeepSeek V4 Flash per combined 1M input + 1M output tokens — making model selection the single largest cost lever in AI development.

The flagship race is tight: Claude Opus 4.8 ($30 combined) trails GPT-5.5 ($35) by $5 with identical $5 input pricing, so the real differentiation at the top happens on output rates and capability, not input cost.

Output premiums are a provider signature, not a constant: 2x input at DeepSeek and xAI, 5x at Anthropic, 6x at OpenAI, and up to 8x at Google. Real-world bills depend on how much a model writes, not just how much it reads — so match the ratio to your workload's shape.

For buyers, the winning strategy is routing, not loyalty: send each task to the cheapest tier that clears its quality bar, and revisit the table often — with five providers competing across 13 models, these numbers move fast.