Closed models hold the lowest chat price in my tests: GPT-6 Luna costs 0.06¢ per answer. The cheapest open-weight result is Qwen3.8 Flash at 0.08¢. Open-weight models (GLM, Qwen, MiniMax, DeepSeek, Kimi, Llama) span both ends of my tables, so vendor type does not predict cost. Hidden reasoning tokens matter more than list price. Prices are as of 2026-10-11 in USD.
Why list price misleads
Per-token prices are a poor guide to task cost. Take GLM 5.3 and GLM 5.3 Flash:
| Model | Input / output per 1M | Chat tokens (in / out / reasoning) | Chat cost |
|---|---|---|---|
| GLM 5.3 | $0.039 / $4.80 | 48 / 391 / 1,391 | 0.86¢ |
| GLM 5.3 Flash | $0.15 / $0.50 | 48 / 391 / 1,391 | 0.09¢ |
Both models have the same token counts in my data (the GLM 5.3 figures are estimated, the Flash figures measured). GLM 5.3 has the lower input price of the popular models I track, but it also has a $4.80 output rate, and every reasoning token is billed at that rate. On a task where output dominates, the input discount barely matters.
Token volume matters too. DeepSeek V4.1 Flash lists at $0.30 input and $1.20 output, which looks cheap. On the chat task it generated 859 visible tokens plus 5,461 reasoning tokens and cost 0.76¢, more than eight times Luna's 0.06¢ and above the 0.24¢ median across all 326 models.
Tokenizers add another gap. For the same English document, Anthropic's tokenizer produces 1.39 times as many tokens as OpenAI's. Most other providers land between 0.98 and 1.07. In the PDF test the Claude models were billed on 18,004 input tokens, against about 13,000 for most others. Details are on the methodology page.
Chat, PDF and code: how the rankings move
I ran three tasks with fixed inputs: a chat question, a 20-page PDF summary (10,023 words of On Liberty), and a review of a 497-line Python file (CPython's textwrap.py). Medians across all 326 models were 0.24¢ for chat, 0.83¢ for the PDF and 1.4¢ for the code review.
| Model | Chat (¢) | PDF summary (¢) | Code review (¢) |
|---|---|---|---|
| GPT-6 Luna | 0.06 | 0.16 | 0.31 |
| Claude Haiku 5.5 | 0.07 | 0.22 | 0.66 |
| Qwen3.8 Flash | 0.08 | 0.32 | 0.83 |
| GLM 5.3 Flash | 0.09 | 0.29 | 0.87 |
| MiniMax M3 | 0.12 | 0.47 | 0.45 |
| GPT-6.1 Sol | 0.67 | 3.1 | 4.9 |
| DeepSeek V4.1 Flash | 0.76 | 0.63 | 2.1 |
| Claude Sonnet 5.5 | 1.3 | 4.5 | 8.0 |
| Kimi K3 | 2.1 | 4.3 | 14 |
| Claude Fable 5.1 | 6.3 | 22 | 40 |
Three patterns stand out.
- Chat rewards low output volume. Claude Sonnet 5.5 used no reasoning tokens but wrote 1,256 visible tokens at $10.00 per 1M, which lands at 1.3¢. GPT-6.1 Sol used only 103 reasoning tokens and cost 0.67¢. The cheapest tier combines a low output rate with short answers.
- PDF summaries are mostly input. Each model read roughly 13,000 tokens (18,004 for Claude), so input price and tokenizer count carry the ranking. Luna, with a $0.10 input rate, leads at 0.16¢. Claude Haiku 5.5 has the same input rate but its larger token count puts it at 0.22¢.
- Code review is where reasoning blows up. Several models hit my 16,000-token output cap.
The code review trap
Qwen3.8 Flash returned zero visible output and used all 16,000 tokens on reasoning. GLM 5.3 Flash produced 15,998 reasoning tokens and 2 visible tokens. Llama 4 Maverick, DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813, GLM 5.3 and Qwen3.8 Max Prime also hit the cap.
These runs look cheap on paper. Qwen3.8 Flash at 0.83¢ and GLM 5.3 Flash at 0.87¢ are both below the 1.4¢ median. But they delivered no review. DeepSeek V4.1 Flash cost 2.1¢ for the same non-result, and Qwen3.8 Max Prime cost 21¢.
By contrast, MiniMax M3 (4,762 in, 2,588 out, no reasoning) and GPT-6 Luna (337 out, 4,921 reasoning) finished the job at 0.45¢ and 0.31¢. A cheap price only counts if the model returns an answer.
Open-weight versus closed
The open-weight group is not cheap as a block.
- Low-cost open-weight: Qwen3.8 Flash (0.08¢ chat, 0.32¢ PDF), GLM 5.3 Flash (0.09¢, 0.29¢), MiniMax M3 (0.12¢, 0.47¢) and Llama 4 Maverick (0.11¢, 0.42¢, though its figures are estimated). All sit well under the medians on chat and PDF work.
- Mid-range open-weight: DeepSeek V4.1 Flash, GLM 5.3 and DeepSeek V4 Pro 0813 cost 0.76¢ to 1.3¢ on chat, mostly because of long reasoning or higher output rates.
- High-cost open-weight: Kimi K3 costs 2.1¢ on chat and 14¢ on code review. It has a $15.00 per 1M output rate and produced 7,522 reasoning tokens in the review. Qwen3.8 Max Prime costs 2.0¢ on chat and 8.4¢ on the PDF.
Closed models show the same spread. GPT-6 Luna and Claude Haiku 5.5 are the two cheapest chat results. GPT-6.1 Sol and Claude Sonnet 5.5 sit in the middle. Claude Fable 5.1 is the most expensive model in all three tests, at 6.3¢, 22¢ and 40¢.
Quality is a separate question
Cost per task says nothing about answer quality. LMArena Text scores in my data give a rough picture:
| Model | LMArena Text | Chat cost (¢) |
|---|---|---|
| Claude Fable 5.1 | 1501 | 6.3 |
| Kimi K3 | 1488 | 2.1 |
| GPT-6.1 Sol | 1484 | 0.67 |
| GLM 5.3 | 1478 | 0.86 |
| Claude Sonnet 5.5 | 1476 | 1.3 |
| DeepSeek V4.1 Flash | 1474 | 0.76 |
| GPT-6 Luna | 1443 | 0.06 |
| MiniMax M3 | 1440 | 0.12 |
| Llama 4 Maverick | 1327 | 0.11 |
GLM 5.3 and DeepSeek V4.1 Flash score within a few points of Sonnet 5.5 at lower chat cost. Llama 4 Maverick is cheap but scores well below the rest. Luna and MiniMax M3 give up some score for a much lower bill. Check your own prompts before treating a leaderboard gap as meaningful.
How to choose
- Measure tokens, not prices. Run your real prompts and record visible output and reasoning tokens. The per-task tables show how far the two can diverge from list price.
- Cap output and check for empty answers. A model that spends its whole budget reasoning can return nothing. Lower the cap or use a non-reasoning model where the task does not need deep thinking. Models such as MiniMax M3 and Qwen3.8 Flash list reasoning as none, but behavior in practice varies, so test it.
- Account for tokenizer differences. If you run Claude models on long documents, expect about 1.39 times the input tokens of an OpenAI model on the same text.
- Use the calculator. The cost calculator takes your own token counts. Per-model breakdowns are on the task pages for chat answers, PDF summaries and code review. Individual models, such as GLM 5.3 Flash and Claude Haiku 5.5, have their own pages.
Only 12 of the models in these tables are measured, with one run per task, default settings and a 16,000-token output cap. The rest are estimated, so treat small gaps as noise.
FAQ
Are open-weight models always cheaper per task?
No. Qwen3.8 Flash is one of the cheapest on chat at 0.08¢, but Kimi K3 costs 14¢ on code review, more than Claude Sonnet 5.5 at 8.0¢ and GPT-6.1 Sol at 4.9¢.
Why does a cheap model sometimes cost more than an expensive one?
Reasoning tokens are billed as output. DeepSeek V4.1 Flash has a $1.20 output rate but used 5,461 reasoning tokens on a chat question, ending at 0.76¢ against 0.67¢ for the much pricier GPT-6.1 Sol.
Which model should I start with for chat?
Start with GPT-6 Luna, the cheapest chat answer at 0.06¢. Compare it with Qwen3.8 Flash as the open-weight option, and test both on your own prompts.