Vibe Engines is free, interactive AI engineering & system design — runs in your browser. Pick a learning path
Model Comparison Table
Claude, GPT, Gemini, DeepSeek, Kimi and Llama side by side: context windows, list prices, open-vs-closed, and a blended per-token cost computed at your real input:output ratio — because a RAG app and a story generator should not read the same pricing table.
| Model | Context | $/1M in | $/1M out | Blended | Relative cost | Notes |
|---|---|---|---|---|---|---|
| GPT-4o mini | 128K | $0.150 | $0.600 | $0.262 | high-volume workhorse | |
| DeepSeek-V3open | 128K | $0.270 | $1.10 | $0.478 | open frontier-class MoE | |
| Llama 3.1 70Bopen | 128K | $0.600 | $0.600 | $0.600 | the fine-tunable classic | |
| Gemini 2.5 Flash | 1M | $0.300 | $2.50 | $0.850 | cheap long-context | |
| Kimi K2open | 128K | $0.600 | $2.50 | $1.07 | open, agent/tool-use focus | |
| Claude Haiku 3.5 | 200K | $0.800 | $4.00 | $1.60 | fast + cheap tier | |
| Gemini 2.5 Pro | 1M | $1.25 | $10.00 | $3.44 | huge context window | |
| GPT-4o | 128K | $2.50 | $10.00 | $4.38 | multimodal all-rounder | |
| Claude Sonnet 4 | 200K | $3.00 | $15.00 | $6.00 | the balanced default | |
| Claude Opus 4 | 200K | $15.00 | $75.00 | $30.00 | frontier reasoning & coding |
A model's price is two numbers, and which one matters depends on the shape of your traffic. This table lists ten models from the tool's own snapshot of indicative list prices, last reviewed in July 2026, with their context windows, and folds input and output prices into one blended cost per million tokens at the input-to-output ratio you set. Rows re-sort as you move the slider.
What it computes
blended $/1M = (ratio × input price + output price) ÷ (ratio + 1)
spread = highest blended ÷ lowest blended (rows shown)
bar width = blended ÷ highest blended in the full table- Ratio: input tokens per output token, from 1:1 to 20:1. The tool's guide is chat about 3:1, RAG with big prompts and short answers about 10:1, and long-form generation about 1:1.
- Input and output price: dollars per million tokens. Across this table output costs between 1 and 8.3 times input.
- Open weights only: filters to DeepSeek-V3, Kimi K2 and Llama 3.1 70B, priced at typical serverless host rates.
Blended cost is the price of one average token in your mix, so multiplying it by all the tokens you process, input plus output, gives spend.
A worked example
At the default 3:1 ratio, Claude Sonnet 4 ($3 in, $15 out) blends to (3 × 3 + 15) ÷ 4 = $6.00, and GPT-4o mini ($0.15, $0.60) to (0.45 + 0.60) ÷ 4 = 0.2625, shown as $0.262. The full sort:
| Model | Context | In | Out | Blended |
|---|---|---|---|---|
| GPT-4o mini | 128K | $0.150 | $0.600 | $0.262 |
| DeepSeek-V3 | 128K | $0.270 | $1.10 | $0.478 |
| Llama 3.1 70B | 128K | $0.600 | $0.600 | $0.600 |
| Gemini 2.5 Flash | 1M | $0.300 | $2.50 | $0.850 |
| Kimi K2 | 128K | $0.600 | $2.50 | $1.07 |
| Claude Haiku 3.5 | 200K | $0.800 | $4.00 | $1.60 |
| Gemini 2.5 Pro | 1M | $1.25 | $10.00 | $3.44 |
| GPT-4o | 128K | $2.50 | $10.00 | $4.38 |
| Claude Sonnet 4 | 200K | $3.00 | $15.00 | $6.00 |
| Claude Opus 4 | 200K | $15.00 | $75.00 | $30.00 |
The summary names GPT-4o mini as cheapest at $0.262 per 1M blended and says the most expensive row costs 114× more per token: $30.00 ÷ $0.2625 = 114.3.
Now move the slider. At 1:1, Llama 3.1 70B ($0.600) is cheaper than DeepSeek-V3 ($0.685), because Llama charges the same for input and output. From 2:1 up DeepSeek-V3 is cheaper, and from 7:1 Gemini 2.5 Flash, whose output costs 8.3 times its input, overtakes Llama as well. The spread narrows from 120× at 1:1 to 104× at 20:1.
Assumptions and limits
- A dated snapshot. Prices are the tool's indicative list prices, last reviewed July 2026, and vendors change them. Trust the ratios between rows, which drift slowly, more than the cents.
- List prices only. Prompt caching and batch discounts, long-context surcharges and reasoning tokens billed as output are not modeled.
- Hosted open models. Open-weight rows use typical serverless host prices. Self-hosting turns per-token pricing into GPU time and utilization.
- Price, not quality. The table ranks cost per token and says nothing about accuracy, latency or how many tokens a model needs to finish a task.
How to use the result
- Measure your ratio first. Divide logged input tokens by output tokens over a week of real traffic; RAG and long-form generation sort the table differently.
- Shortlist, then evaluate. Take the cheapest few rows that plausibly meet your quality bar, run your own evaluation set on them and pick the cheapest that passes.
- Turn blended cost into a budget. Monthly tokens × blended rate ÷ 1,000,000 is the bill: 500 million tokens a month at 3:1 is $3,000 on Claude Sonnet 4 and about $131 on GPT-4o mini.
- Route rather than marry. With a gap of more than 100× between rows, sending easy requests to a cheap model and escalating hard ones can be the largest cost lever available.
- Check the context column. The Gemini rows list 1M tokens, Claude rows 200K and the rest 128K. A cheap model that cannot hold your prompt is not an option.
How it works
- Blended $/1M = (ratio × input + output) ÷ (ratio + 1), sorted ascending.
- Chat ≈ 3:1, RAG ≈ 10:1, long-form generation ≈ 1:1.
- Open-weights rows are downloadable and self-hostable, priced at typical hosts.
- Indicative list prices, last reviewed 2026-07, kept consistent with the pricing comparator.
Frequently asked questions
What is blended token cost and how is it calculated?
Blended cost is the average price of one token in your traffic mix: ratio times the input price, plus the output price, divided by ratio plus 1, where ratio is input tokens per output token. At 3:1, Claude Sonnet 4 at $3 in and $15 out blends to (9 + 15) divided by 4, which is $6.00 per million. It turns two prices into one number you can sort by, and multiplying it by your total tokens gives spend.
Why does the input to output ratio change which model is cheapest?
Because models price output differently relative to input. In this table output costs from 1 times input (Llama 3.1 70B) to 8.3 times (Gemini 2.5 Flash). At 1:1, Llama at $0.600 blended beats DeepSeek-V3 at $0.685; from 2:1 upward DeepSeek-V3 is cheaper, and from 7:1 Gemini 2.5 Flash passes Llama too. A RAG app sending ten input tokens per output token mostly pays input prices, while a long-form generator mostly pays output prices.
Which models in the table are open weights?
DeepSeek-V3, Kimi K2 and Llama 3.1 70B. Their weights can be downloaded, fine-tuned and self-hosted, and the prices shown are typical serverless host rates rather than a single vendor's list. With the open-only toggle on at the default 3:1 ratio, DeepSeek-V3 is cheapest at $0.478 blended and Kimi K2 costs about 2.3 times as much at $1.07. Self-hosting replaces per-token prices with hardware cost, so the comparison then depends on utilization.
How current are the prices?
They are a snapshot of indicative list prices that the tool states were last reviewed in July 2026, kept consistent with the site's pricing comparator. Vendors change prices and release new models often, so check the provider's pricing page before committing a budget. The relative geometry, such as the 114 times gap between the cheapest and most expensive rows at 3:1, changes far more slowly than individual prices and is the durable signal.
Should I just pick the cheapest model?
Pick the cheapest model that clears your quality bar for each task. Use the table to shortlist a few rows, run your own evaluation set on them, and compare cost per completed task rather than per token, since a model that needs fewer tokens or fewer retries can win despite a higher rate. Many production systems route between rows, sending easy requests to a cheap model and escalating hard ones, which puts the spread in this table to work.
How do I turn blended cost into a monthly bill?
Multiply your monthly tokens, input and output together, by the blended rate and divide by a million. At the default 3:1 ratio, 500 million tokens a month costs 500 times $6.00, or $3,000, on Claude Sonnet 4 and about $131 on GPT-4o mini. For an estimate built from request counts and separate token sizes, use a pricing calculator; this table is for ranking models at your traffic shape.