A 1,500-word blog post costs between 0.12¢ and 19¢ in my tests, depending on the model. The median across all 326 models is 0.87¢ per post. A blog post is mostly generated text, so the output price (including reasoning tokens) sets the bill.
Why output price decides the bill
My blog post task sends a short prompt and gets back a long article. For GPT-6 Luna, that was 59 input tokens, 2,217 output tokens and 143 reasoning tokens. Reasoning tokens are billed as output, so the billable output is about 2,360 tokens.
Luna lists at $0.10 per 1M input tokens and $0.50 per 1M output tokens. The 59 input tokens cost almost nothing. The roughly 2,360 output tokens at $0.50 per 1M come to about 0.12¢, which matches my measured result.
Three things follow from this:
- Compare output prices first. Input price barely moves the total for a blog post.
- Models with the same list price can still cost different amounts, because they generate different numbers of tokens.
- Reasoning tokens add to the bill without adding a word to the article.
What the blog post runs show
The table uses my blog post results. "Measured" means one run per task through OpenRouter with default settings and output capped at 16,000 tokens. "Estimated" means the figure was derived rather than run directly.
| Model | Cost per post | Cost per 1,000 posts | Output tokens | Reasoning tokens | Basis |
|---|---|---|---|---|---|
| GPT-6 Luna | 0.12¢ | $1.19 | 2,217 | 143 | Measured |
| Claude Haiku 5.5 | 0.19¢ | $1.89 | 3,560 | 200 | Measured |
| GPT-6.1 Sol | 2.3¢ | $23.00 | 2,116 | 172 | Measured |
| Claude Sonnet 5.5 | 3.7¢ | $36.58 | 3,401 | 239 | Measured |
| Claude Opus 5.5 | 7.3¢ | $73.16 | 3,401 | 239 | Estimated |
| Claude Fable 5.1 | 18¢ | $182.90 | 3,401 | 239 | Estimated |
The live table below shows the current top results for the same task.
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  DeepSeek V4.1 FlashDeepSeek | 0.04¢$0.44 / 1,000 | 1475 | 81 / 3,039 | measured best value | |
| 2 |  GPT-6 LunaOpenAI | 0.12¢$1.19 / 1,000 | 1443 | 59 / 2,360 | measured | |
| 3 |  GLM 5.3 FlashZ.ai | 0.14¢$1.39 / 1,000 | 1474 | 62 / 6,923 | measured | |
| 4 |  Claude Haiku 5.5Anthropic | 0.19¢$1.89 / 1,000 | - | 90 / 3,760 | measured | |
| 5 |  MiMo-V2.6-ProXiaomi | 0.20¢$2.04 / 1,000 | 1479 | 220 / 2,232 | estimated best value | |
| 6 |  MiniMax M3MiniMax | 0.27¢$2.72 / 1,000 | 1441 | 219 / 2,216 | measured |
Claude Haiku 5.5 and GPT-6 Luna share the same list prices, $0.10 input and $0.50 output. Their blog post costs are still 0.19¢ and 0.12¢. Haiku wrote 3,560 output tokens against Luna's 2,217. Part of that gap is the tokenizer, covered below.
GPT-6.1 Sol and Claude Sonnet 5.5 also share list prices, $2.00 input and $10.00 output. Sol measured 2.3¢ per post and Sonnet 3.7¢. Sonnet produced 3,401 output tokens against Sol's 2,116.
Reasoning tokens are the hidden cost
Some models think before they write, and every thinking token is billed at the output rate. Here is how that played out on the blog post task:
- Mistral Large 4 produced 2,064 output tokens and 5,705 reasoning tokens, for a measured 1.6¢ per post. Reasoning was nearly three times the visible text.
- Gemini 3.1 Pro Preview used 2,522 output tokens and 1,606 reasoning tokens, for a measured 5.0¢ per post.
- GLM 5.3 Flash used 2,174 output tokens and 4,749 reasoning tokens, for a measured 0.35¢ per post. A cheap rate absorbs a lot of thinking.
- Qwen3.8 Flash spent the full 16,000-token cap on reasoning and returned no article. It still billed a measured 0.75¢.
- Qwen3.8 Max Prime hit the same cap. Its estimated 19¢ per post was the highest in my blog post results.
A low list price does not guarantee a cheap task, and a capped run is a paid run that delivered nothing. Run a sample of your real topics before committing a model to production.
Product descriptions: same logic, different mix
My product description task asks for 10 descriptions per run. The median across all 326 models is 0.63¢.
| Model | Cost per task | Cost per 1,000 tasks | Basis |
|---|---|---|---|
| GPT-6 Luna | 0.05¢ | $0.54 | Measured |
| Claude Haiku 5.5 | 0.21¢ | $2.12 | Measured |
| Claude Sonnet 5.5 | 4.4¢ | $44.16 | Measured |
| Claude Opus 5.5 | 8.8¢ | $88.32 | Estimated |
| Claude Fable 5.1 | 22¢ | $220.79 | Estimated |
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 FlashDeepSeek | 0.04¢$0.41 / 1,000 | 1475 | 307 / 2,712 | measured best value | |
| 2 | GPT-6 LunaOpenAI | 0.05¢$0.54 / 1,000 | 1443 | 283 / 1,021 | measured | |
| 3 | GPT-5.4 NanoOpenAI | 0.13¢$1.33 / 1,000 | 1401 | 283 / 1,021 | estimated | |
| 4 | GLM 5.3 FlashZ.ai | 0.14¢$1.42 / 1,000 | 1474 | 289 / 7,041 | measured | |
| 5 | MiMo-V2.6-ProXiaomi | 0.16¢$1.59 / 1,000 | 1479 | 444 / 1,611 | estimated best value |
Sonnet 5.5 costs more on the 10 descriptions (4.4¢) than on the full blog post (3.7¢). On descriptions it used 1,866 output tokens and 2,459 reasoning tokens. On the blog post it used 3,401 output tokens and 239 reasoning tokens. Billable output was therefore higher on the shorter job, because of the added thinking.
Haiku 5.5 shows the same pattern at a smaller scale. It used 1,803 output tokens and 2,342 reasoning tokens on the descriptions, and cost 0.21¢ against 0.19¢ for the blog post.
Tokenizers change the count, not just the price
The same English text becomes a different number of tokens depending on the model family. With OpenAI at 1.00, my tokenizer ratio for Anthropic is 1.39, so the same document produces about 39% more tokens on Claude. Other ratios are close to OpenAI's: Google 0.98, MiniMax 0.99, Mistral 1.03 and xAI 1.07.
A higher token count raises the bill even at an identical list price. That is one reason Sonnet 5.5 costs more than GPT-6.1 Sol on the blog post task despite matching $2.00 and $10.00 list prices. Compare cost per finished task, not per-token rates. See the methodology for how I count.
How to cut the cost of AI blog posts
- Choose by cost per finished post, not list price. The calculator does this comparison.
- Turn reasoning down or off for drafting jobs that do not need it, where the model allows.
- Watch the output cap. Test that your model returns an article before you batch hundreds of posts.
- Keep prompts short. Input is minor for one post, but it adds up across large batches.
- Test your own topics. My figures come from one run per task at default settings, so your results will vary.
For the task definitions behind these numbers, see write a blog post and product descriptions. For subscription options, see plans.
FAQ
Is a cheap model good enough for blog posts?
It can be. GPT-6 Luna and Claude Haiku 5.5 were the two cheapest in my blog post results, at 0.12¢ and 0.19¢, and both were measured. My figures cover cost only, so check output quality against your own standards before switching.
Why does a blog post cost more than a proofread?
A blog post generates a long article, while a one-page proofread returns text about the length of the page. Output is priced higher than input, so more generated text means a higher bill. The median is 0.87¢ for a blog post and 0.26¢ for a proofread.
Do the costs include reasoning tokens?
Yes. Reasoning tokens are billed as output tokens, and my per-task costs include them. The token columns list output and reasoning separately so you can see how much of the bill comes from thinking.