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.
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-5.5 sits at $5 / $30 per million tokens (input / output). 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, you’re paying roughly $23.75 per million across the mix, not $5. OpenAI’s batch and flex modes discount both rates 50%, landing at $2.50 / $15 — that discount is real and worth routing work through when latency tolerance allows. GPT-5.5 Pro, if you need it for complex reasoning, pushes to $30 / $180: at that output rate, token economics become the dominant engineering constraint, not capability.
Claude Opus 4.8 runs approximately $5 / $25. On paper it looks similar to GPT-5.5 standard. The catch is extended thinking. Enable it, and output tokens can multiply 2–5x per task, because thinking tokens bill at the output rate. A task that produces 2,000 output tokens in standard mode may generate 8,000–10,000 tokens in extended thinking mode before returning an answer. The capability is often worth it; the cost structure requires you to model it explicitly. Claude Sonnet 4.6 at $1.50 / $7.50 is a more predictable choice when you don’t need the reasoning depth — it’s where most production workloads should start.
GPT-5.4 Mini at $0.50 / $2.00 occupies the small-model niche for high-volume, low-complexity tasks: classification, intent detection, structured extraction, routing logic. At scale, the difference between $0.50 and $5 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. 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. 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. DeepSeek V4-Flash at $0.14 / $0.28 is the current cost floor for hosted API inference among models with serious capability. V4-Pro at $0.145 / $1.74 trades some speed for more headroom. These prices are real, not promotional — they’re enabled by architectural efficiency and scale economics from the Chinese market.
Z.ai GLM 5.2 and Kimi K2.7 sit in a similar tier, though not quite as cheap. All three are MIT-licensed open weights, which is the other half of the story: you can self-host them.
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 displace GPT-5.4 Mini on capable tasks. The hosted APIs are cheap. The MIT license means 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 DeepSeek V4-Flash 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 MIT license exists 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.
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. Anthropic’s prompt caching delivers similar magnitude savings on Sonnet 4.6 and Opus 4.8.
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 the variance across tiers is still 100x from top to bottom. A system using GPT-5.5 Pro where GPT-5.4 Mini 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.