As of late August 2026, the models most founders ship on cost, per million tokens (input/output): Claude Opus 5 $5/$25, Sonnet 5 $3/$15, Haiku 4.5 $1/$5; GPT-5.6 Sol $5/$30 and its budget Luna tier $0.20/$1.20; Gemini 3.1 Pro $2/$12 and Gemini 3.7 Flash $0.75/$3.75. That's the sticker price. Your actual bill is set by four things the table doesn't show — how much you write vs. read, whether you cache, whether you batch, and whether you route. Here's the comparison, then the math to turn it into a monthly number.
The comparison table#
| Model | Input $/1M | Output $/1M | Context | Best for |
|---|---|---|---|---|
| Claude Opus 5 | $5 | $25 | 1M | Hard reasoning, long-horizon agents |
| Claude Sonnet 5 | $3 | $15 | 1M | Everyday workhorse (intro-rate note) |
| Claude Haiku 4.5 | $1 | $5 | 200K | Cheap, fast classification/extraction |
| GPT-5.6 Sol | $5 | $30 | ~1M | OpenAI flagship reasoning |
| GPT-5.6 Terra | $2 | $12 | ~1M | Mid-tier, cheaper reasoning |
| GPT-5.6 Luna | $0.20 | $1.20 | ~1M | High-volume budget calls |
| GPT-5 (base) | $1.25 | $10 | 400K | Prior-gen, still cheap |
| Gemini 3.1 Pro | $2 | $12 | 1M | Long-context, multimodal |
| Gemini 3.7 Flash | $0.75 | $3.75 | ~1M | Cheap coding/agents (introductory) |
A few footnotes that matter more than they look:
- Prices move weekly. These are published list rates gathered in late August 2026; treat them as the shape of the market, not a live quote, and confirm on the provider's own page before you commit. (OpenAI reportedly trimmed GPT-5.6 Sol toward $4/$20 on Aug 22; Gemini 3.7 Flash's $0.75/$3.75 is an introductory rate that rises to $1.50/$7.50 on Jan 1, 2027.)
- Claude Sonnet 5 shows an August introductory rate of $2/$10 that is scheduled to end Aug 31, reverting to the $3/$15 listed here — budget on the number you'll actually pay in September.
- Long-context surcharges are real. Several models charge more above ~200K tokens in a single request (Gemini and GPT-5.6 both roughly double). If you routinely send huge prompts, price the high-context tier, not the headline.
- Chinese and open-weight APIs (DeepSeek, Qwen, GLM, and open models hosted on Groq/Together/Fireworks/DeepInfra) sit well below this table — often $0.05–$0.60/1M — and are covered in our Chinese-model share breakdown; the trade is frontier reasoning for price.
How to estimate your bill (the calculator)#
There is exactly one formula, and the only trick is to price input and output separately:
monthly cost = ( avg_input_tokens × input_price
+ avg_output_tokens × output_price ) × requests_per_month…where prices are dollars per 1,000,000 tokens. Work an example — a support assistant on Claude Sonnet 5 ($3 in / $15 out), sending ~1,500 input tokens and generating ~500 output tokens per request, at 100,000 requests/month:
input : 1,500 tok × $3 / 1,000,000 = $0.0045 per request
output: 500 tok × $15 / 1,000,000 = $0.0075 per request
per request = $0.0120
× 100,000 requests = $1,200 / month (before discounts)Notice that the 500 output tokens cost more than the 1,500 input tokens — because output is priced ~5× higher. That's the single most common budgeting mistake: reading the input price, ignoring the output column, and under-estimating by half. Two guardrails keep the estimate honest:
- Count output at its real rate, and shorten generations wherever you can — output length is the cheapest lever you own.
- Apply a 1.7–2× buffer over the naive number for retries, system prompts, tool schemas, and context you'll forget to count. For a fuller treatment of measuring the tokens themselves, see how to read an LLM pricing page and how to measure real LLM cost.
The three multipliers that beat picking a cheaper model#
The sticker price is the start of the number, not the end. Three levers move it far more than swapping models:
- Prompt caching (~90% off repeated input). Any stable prefix you resend — a system prompt, tool definitions, a document you ask many questions about — bills at roughly 10% of the input rate after the first call across Anthropic, OpenAI, and Gemini. For agents this is the whole game.
- Batch API (50% off, input and output). Jobs that can tolerate minutes-to-hours of latency — evals, enrichment, offline generation — run at half price on Anthropic, OpenAI, and Google. At Anthropic the batch discount stacks with cache reads, so a batchable, cache-friendly job can land near 5% of its naive list cost on the input side.
- Routing (exploit the spread). The input-price gap between a flagship and a budget tier is up to 25× (GPT-5.6 Sol $5 vs. Luna $0.20). Send classification, extraction, and simple chat to the cheap tier and reserve the frontier for hard tasks — the discipline behind a cost-aware model router, and the mechanism behind the "40% average savings" claims from the new gateway products like Ramp's Router.
Two teams on the identical model and traffic routinely see a 10–20× difference in their bill. The gap is almost never the sticker price — it's whether they cache, batch, and route.
Why agents cost 10× what your chatbot math predicts#
If you budget an agent like a chatbot, you'll be off by an order of magnitude. A chatbot turn is roughly one prompt and one answer. An agent resends its entire context every step — the system prompt, the tool schemas, and the accumulated results of every prior tool call — so across an n-step task, cumulative input grows on the order of n(n+1)/2: quadratically, not linearly. In practice, input becomes 70–85% of an agent's total spend, the mirror image of chat.
Two consequences follow. First, caching matters more for agents than for anything else — that resent prefix is exactly what cache reads discount ~90%. Second, trimming what the agent carries between steps — pruning stale tool output, summarizing history — cuts cost faster than switching models. And always measure tokens per completed task, not per call, or your budget will be fiction. Our guide to reducing agent token costs and the agent-framework token-cost comparison go deeper on both.
So which model should you actually pick?#
Cheapest-that-passes-your-evals, then route everything else to it. Start every task on the cheapest tier that clears your quality bar — often Haiku 4.5, Gemini Flash, or GPT-5.6 Luna — and promote to a flagship only where the evals demand it. Layer caching on every stable prefix and batching on every job that can wait. If your volume is steady and high enough, run the rent-vs-buy math against our GPU rental price map and self-hosting vs. API cost breakdowns — but for almost every team of one, the answer is a well-cached, well-routed API bill, not a GPU you have to keep busy. The sticker price in the table above is where the estimate starts. Caching, batching, and routing are where it lands.
Prices verified against provider pricing pages and cross-checked in late August 2026; they change often, so confirm the current rate before you commit a budget.
