What Are Tokens? How Much Do 100M Tokens Cost? Complete 2026 AI Model Pricing Guide

Bottom line: 100M tokens costs roughly $10–14 on Flash tier, ~$540 on Sonnet, ~$2,700 on Opus—all in USD. Four tables below list June 2026 list prices. One-line takeaway under each table; audience picks at the end.

If you are sizing an Agent budget or comparing Cursor defaults to a self-hosted API stack, start here—not with benchmark scores. Every figure below is US dollars per million tokens ($/M) unless noted. Model names and tiers match what developers actually route on OpenRouter in mid-June 2026; your invoice may differ slightly with caching, routing, or enterprise discounts.

Table 1: Flash execution tier — June 2026 API rates

OpenRouter and official pricing pages. Unit: USD per million tokens ($/M). List prices below; your dashboard shows actuals after cache and routing.

Flash models are not “worse GPT”—they are MoE architectures tuned for high-volume loops. DeepSeek V4 Flash leads on raw usage because it pairs 1M context with sub-$0.10/M input. Hy3 and Kimi matter when your pipeline is CJK-heavy or tool-call dense. Gemini 2.5 Flash and GPT-4o mini are the Western-vendor equivalents at roughly 1.5–2× DeepSeek’s list price.

Table 2: Frontier review tier — June 2026 API rates

Frontier tier pricing explains why “just use the best model” stopped being viable once Agents began burning 50K–200K tokens per task. A single Sonnet review is affordable; making Sonnet the default for every file read is not. Gemini 2.5 Pro and o-series reasoning models fill niche lanes—long multimodal docs or formal proofs—not daily coding loops.

Table 3: 100M-token bill comparison

Common yardstick: 100M tokens. Three mixes: input-only / 80·20 chat / 90·10 Agent.

“100M tokens” is a useful mental unit: roughly a busy week for a small Agent pilot, or a few hours for a high-volume RAG service. Pure-input column models ingestion-heavy pipelines (search, rerank, classify). The 80/20 mix matches chat products. The 90/10 Agent column is the one to stress-test if your tool reads entire repositories before writing a short patch.

Table 4: Typical single Agent task cost

Assumption: 100K input + 10K output, 80% input cache hit. Daily bill at 500 runs.

The per-task row is what engineering leads should paste into a spreadsheet: multiply by expected daily Agent invocations, then add headroom for retries. If your product triggers an LLM on every CI failure, every support ticket, and every nightly batch job, Table 4 scales linearly—there is no “unlimited tier” on raw API pricing.

Audience picks: who you are → which row to use

Quick notes per audience

• Solo dev / IDE: Subscription tools bundle token cost—you optimize time, not $/M. Revisit API pricing only when you outgrow included fast requests.
• Indie Agent: DeepSeek default + Sonnet on merge is the lowest-friction split; budget $20–80 until usage dashboards prove otherwise.
• CJK long chains: Hy3’s tool stability often beats raw $/M; pair with Kimi when documents are mainland-Chinese heavy.
• RAG product: Cache your system prompt and doc prefixes; Flash + Sonnet review beats single-model Sonnet by an order of magnitude.
• 500+ tasks/day: Dual-route DeepSeek/Hy3 before considering Opus; Opus belongs on human-gated steps only.
• Data residency: Local 7B–14B removes per-token billing for predictable workloads; API for bursts and 200B+ MoE capability.
• Regulated: Price is secondary to audit trails—still route bulk token volume through Flash, not Opus.

In one line: price picks Flash; risk picks Sonnet/Opus. 100M tokens is the yardstick; the audience table is the answer.

Revisit this page when vendors cut Flash prices again—June 2026 already moved faster than 2025 frontier lists. Export your own usage split monthly; tables age well, but your input/output ratio is the variable that actually moves your bill. When in doubt, default to Flash, measure real usage, then promote models to production.