Most conversations about AI cost start in the wrong place. They compare headline per-token prices as if that number is what you actually pay, then stop. The real picture requires understanding input/output ratios, reasoning-token overhead, caching behavior, and where the API-vs-self-host crossover actually lands for your workload. None of that fits in a benchmark table.
This breakdown focuses on what matters for production decisions in 2026.
Update, October 15, 2026: We checked every price in this guide against the providers’ pricing pages and replaced the June figures with current list prices. The June version used GPT-5.5, Claude Opus 4.8 and DeepSeek V4-Flash as reference points; GPT-6, the Claude 5.5 models and DeepSeek V4.1 Flash have arrived since, and the cheapest closed models now list below DeepSeek’s hosted rates. The June text also had some prices wrong: GPT-5.4 Mini launched at $0.75 / $4.50, not $0.50 / $2.00, and $1.50 / $7.50 was Claude Sonnet 4.6’s batch rate, not its $3 / $15 list price. OpenAI API pricing Claude pricing All figures below are standard list prices per million tokens unless stated.
The Rate Card Isn’t the Bill
Every major provider publishes input and output rates separately, and for good reason: output tokens cost more — often far more. The ratio matters enormously in practice.
GPT-6 Sol lists at $2 / $10 per million tokens (input / output), and GPT-5.5, this guide’s June reference point, is still listed at $5 / $30. OpenAI API pricing If your workload produces three output tokens per input token — typical for summarization or code generation — your effective blended cost trends toward the output rate. At 3:1 on Sol, you’re paying $8 per million across the mix, not $2: one input token at $2 plus three output tokens at $10, divided by four. The same mix on GPT-5.5 comes to $23.75. OpenAI’s Batch and Flex modes discount both rates 50%, putting Sol at $1 / $5 — that discount is real and worth routing work through when latency tolerance allows. At the top of the range, GPT-6 Astra lists at $10 / $50 and GPT-5.5 Pro at $30 / $180: at that output rate, token economics become the dominant engineering constraint, not capability. OpenAI API pricing
Claude Opus 5.5 runs $4 / $20, down from $5 / $25 for Opus 4.8 and Opus 5. Claude pricing On paper it sits between GPT-6 Sol and GPT-5.5. The catch is thinking. Thinking tokens bill at the output rate; Anthropic’s API reports them as part of the billed output count. Extended thinking Turn thinking on, and output tokens can multiply 2–5x per task. A task that produces 2,000 output tokens without thinking may generate 8,000–10,000 tokens with it before returning an answer. On Claude 4.7 and later models, adaptive thinking replaces fixed token budgets, and the effort setting decides how much the model thinks; Opus 5.5 defaults to medium. Extended thinking The capability is often worth it; the cost structure requires you to model it explicitly. Claude Sonnet 5.5 at $2 / $10 is a more predictable choice when you don’t need the reasoning depth — it’s where most production workloads should start. Claude pricing
At the small end, GPT-6 Luna and Claude Haiku 5.5 both list at $0.10 / $0.50. Haiku 5.5’s rate applies to prompts up to 100,000 tokens; above that, the whole request is billed at five times the rate. OpenAI API pricing Claude pricing Gemini 3.8 Flash is $0.75 / $3.75 through December 31, 2026, and Google’s pricing page lists $1.50 / $7.50 from January 1, 2027. Gemini API pricing These models fit high-volume, low-complexity tasks: classification, intent detection, structured extraction, routing logic. At scale, the difference between $0.10 and $2 on input is not marginal.
Reasoning Tokens as a Hidden Multiplier
Extended reasoning — the ability to “think” before responding — is now table-stakes across frontier models, but the billing mechanics vary and the differences are significant.
Z.ai’s GLM 5.2 lists at $1.40 / $4.40 with cached input at $0.26, and its successor GLM 5.3 lists at the same rates. Z.ai pricing Reasoning is optional, which lets you dial it in per request and model cost precisely. Moonshot’s Kimi K2.7 Code runs $0.95 / $4.00 (cached $0.19), with reasoning always on — thinking tokens always bill as output tokens, so the effective output cost is always higher than the listed $4.00 depending on task complexity. Kimi API pricing There’s no mode where you’re paying $0.95 in and $4.00 out flat; you’re paying $0.95 in and some multiple of $4.00 out.
Understanding how tokens are counted matters here more than anywhere else in the stack. Reasoning tokens are opaque — you don’t see them in the completion, only in the billing. Any production system using extended reasoning needs instrumentation to surface actual token counts per request, or cost modeling becomes guesswork.
The Open-Weight Cost Floor
DeepSeek V4 redrew the cost curve this year, but its prices have moved twice since June. DeepSeek raised V4 rates in August and split them into peak and off-peak tiers, then replaced V4-Flash with V4.1 Flash in September. V4.1 Flash (deepseek-flash) now lists at $0.15 / $0.60 off-peak and $0.30 / $1.20 during peak hours, which run 01:00–04:00 and 06:00–10:00 UTC on weekdays. V4-Pro is $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak. DeepSeek pricing
In June, V4-Flash at $0.14 / $0.28 was the cheapest hosted API among models with serious capability. At list price that is no longer true: GPT-6 Luna and Claude Haiku 5.5 charge $0.10 / $0.50, below V4.1 Flash’s off-peak rates on both lines. DeepSeek still wins on cached input, at $0.003 off-peak against $0.01 for both. The June version also called DeepSeek’s prices real rather than promotional. The August increase showed otherwise: the low V4-Pro rate began as a launch discount and lasted under three months as the official price. DeepSeek’s pricing history
Z.ai GLM 5.2 and Kimi K2.7 sit above DeepSeek on price. All three publish open weights under MIT or modified-MIT licenses (Kimi’s adds an attribution requirement for very large products), which is the other half of the story: you can self-host them. GLM 5.3 is different; its custom license adds a clause for the largest model-as-a-service vendors.
The open source AI models guide covers a broader set — Gemma 4, Qwen 3.5, MiniMax M2.5, NVIDIA Nemotron 3 among others — but DeepSeek V4, GLM 5.2, and Kimi K2.7 are the ones serious enough to take capable work from closed models. Their hosted APIs are no longer the cheapest per token. The licenses mean you’re not locked to those APIs.
Hosted API vs. Self-Host
Self-hosting an H100 or Blackwell GB200-class cluster has a clear economic structure: high fixed cost, low marginal cost per token. The crossover against hosted APIs depends entirely on utilization. A cluster running at 80%+ utilization on a large-context model starts to undercut even the cheapest hosted rates in this guide at sufficient scale. A cluster running at 30% utilization while you figure out load patterns is not cheaper — it’s a capital allocation problem disguised as a cost problem.
The non-cost argument for self-hosting open weights has strengthened considerably, and it’s mostly about data residency. Hosted Chinese open-weight APIs — DeepSeek, GLM 5.2, Kimi — raise legitimate questions about where inference traffic and prompt content actually go. The permissive licenses exist precisely to let you sidestep that: run the weights in your own infrastructure, no data leaves your perimeter. For regulated industries or any workload handling personally identifiable information, that distinction is not optional. When “your own infrastructure” is a desk rather than a datacenter, our best local LLM machines guide covers the hardware that runs these models privately — and why memory, not raw compute, decides what you can run.
Our AI solutions work almost always starts with a model routing architecture rather than a single-model choice, because no single model wins across cost, latency, and capability simultaneously.
Caching and Its Actual Impact
Prompt caching is one of the few genuine free lunches in this space. GLM 5.2’s cached input rate drops from $1.40 to $0.26 — an 81% reduction. Kimi K2.7’s cached input drops from $0.95 to $0.19, an 80% reduction. Anthropic goes further on its newest models: a cache hit on Sonnet 5.5 or Opus 5.5 costs 5% of the base input price, a 95% reduction, although writing to the 5-minute cache costs 1.25 times the base rate. Claude pricing GPT-6 Sol and Luna charge 10% of the input rate for cached input, and DeepSeek V4.1 Flash’s off-peak cache hit, $0.003 against $0.15, is a 98% reduction. OpenAI API pricing DeepSeek pricing
Caching only helps when you have a large, stable prompt prefix — system instructions, retrieved documents, few-shot examples — that repeats across many requests. RAG architectures with a consistent system prompt and rotating retrieved chunks are a natural fit. One-off requests with highly variable context are not, and no amount of caching optimization will change that.
What the Numbers Tell You
Frontier models have genuinely compressed in price relative to capability over the past two years, but list prices still span 300x or more from top to bottom. GPT-5.5 Pro charges $30 / $180; GPT-6 Luna charges $0.10 / $0.50. A system using GPT-5.5 Pro where GPT-6 Luna would suffice isn’t making a capability decision — it’s making a cost decision carelessly.
The practical decision tree isn’t complex: match model capability to task complexity, use caching aggressively where context is reusable, route high-volume low-complexity work to the cheapest capable model, and audit reasoning-token overhead in any workload that enables extended thinking. What’s hard isn’t the arithmetic — it’s getting production instrumentation in place early enough that you’re making these decisions from data rather than estimates.
The projects that go wrong on cost do so not because the pricing is opaque but because token economics stay invisible until they show up on an invoice.