The cheapest LLM API for most everyday tasks as of 2026-10-11 is GPT-6 Luna, with Claude Haiku 5.5 close behind. Both list at $0.10 per million input tokens and $0.50 per million output tokens. List price does not decide the winner, though: models that look cheap per token can cost more per task when they spend thousands of hidden reasoning tokens before answering.
Reasoning tokens are billed as output tokens, so per-task cost is what matters. The per-task figures below are labeled measured, estimated or modeled, and the per-token prices are list rates.
Why list price misleads
Several models look cheap on a pricing page and cost more in practice.
GLM 5.3 lists at $0.039 per million input tokens, the lowest input price in my set, but its output price is $4.80 per million. On the code review task it cost 7.7¢ per task (estimated), against 0.31¢ for GPT-6 Luna (measured).
Qwen3.8 Flash lists at $0.15 input and $0.47 output. On the 500-line code review it used 16,000 reasoning tokens, hit my 16,000-token output cap and returned no visible answer. It still cost 0.83¢ (measured). DeepSeek V4.1 Flash, at $0.30 input and $1.20 output, hit the same cap and cost 2.1¢ (measured).
Mistral Large 4 lists at $0.68 input and $2.09 output. It spent 7,522 reasoning tokens on code review and cost 2.2¢ per task (measured).
Cheapest per task, by workload
The three lowest-cost models on each task, in cents per task:
| Task | Cheapest | Second | Third | Median of 326 models |
|---|---|---|---|---|
| Chat answer | GPT-6 Luna 0.06¢ | Claude Haiku 5.5 0.07¢ | Qwen3.8 Flash 0.08¢ | 0.24¢ |
| Summarize 20-page PDF | GPT-6 Luna 0.16¢ | Claude Haiku 5.5 0.22¢ | GLM 5.3 Flash 0.29¢ | 0.83¢ |
| Review 500 lines of code | GPT-6 Luna 0.31¢ | MiMo-V2.6-Pro 0.44¢ (estimated) | Gemini 3.1 Flash Lite 0.44¢ (estimated) | 1.4¢ |
| 1,500-word blog post | GPT-6 Luna 0.12¢ | Claude Haiku 5.5 0.19¢ | MiMo-V2.6-Pro 0.20¢ (estimated) | 0.87¢ |
Here is the chat answer ranking in more detail:
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GLM 5.3 FlashZ.ai | 0.04¢$0.36 / 1,000 | 1474 | 48 / 1,782 | measured best value | |
| 2 |  GPT-6 LunaOpenAI | 0.06¢$0.59 / 1,000 | 1443 | 45 / 1,179 | measured | |
| 3 |  Claude Haiku 5.5Anthropic | 0.07¢$0.66 / 1,000 | - | 61 / 1,299 | measured | |
| 4 |  Qwen3.8 FlashAlibaba Qwen | 0.08¢$0.76 / 1,000 | - | 88 / 1,598 | measured | |
| 5 |  Mistral Small 4Mistral | 0.08¢$0.83 / 1,000 | - | 40 / 1,377 | estimated | |
| 6 |  MiMo-V2.6-ProXiaomi | 0.09¢$0.90 / 1,000 | 1479 | 206 / 933 | estimated best value | |
| 7 |  DeepSeek V4.1 FlashDeepSeek | 0.09¢$0.91 / 1,000 | 1475 | 65 / 6,320 | measured | |
| 8 |  Llama 4 MaverickMeta | 0.11¢$1.06 / 1,000 | 1327 | 87 / 1,592 | estimated | |
| 9 |  MiniMax M3MiniMax | 0.12¢$1.18 / 1,000 | 1441 | 205 / 929 | measured | |
| 10 |  Command A+Cohere | 0.15¢$1.46 / 1,000 | - | 205 / 930 | estimated |
GPT-6 Luna and Claude Haiku 5.5 were measured on all four tasks above. Many other entries are estimated, and each task page flags the basis. See chatbot answers, PDF summaries, code review and blog posts.
Where cheap-per-token models lose
Three patterns repeat across tasks.
- Reasoning-heavy runs. Qwen3.8 Flash and GLM 5.3 Flash both hit the 16,000-token cap on code review. GLM 5.3 Flash cost 0.87¢ and Qwen3.8 Flash cost 0.83¢, versus 0.31¢ for GPT-6 Luna, which used 4,921 reasoning tokens.
- Agent loops. On the modeled 50-step coding agent, GPT-6 Luna cost 4.1¢ per task, the lowest in my set. Claude Haiku 5.5 cost 6.6¢. The median across 326 models is 48¢. Caching matters a lot here: GPT-6 Luna would cost $0.23 without caching, so the cached input rate affects agent work far more than single calls. See the agent task.
- High output prices. GLM 5.3 charges $4.80 per million output tokens. In the agent scenario it cost 30¢ per task, against 4.1¢ for GPT-6 Luna.
Best value picks
- GPT-6 Luna is the default choice. It was cheapest on chat, PDF summary, code review, blog post and the agent run. It has a 1,050,000-token context window and an LMArena Text score of 1443.
- Claude Haiku 5.5 is the pick if you want a second vendor with a 1,000,000-token context. It was second on chat, PDF and blog, and cost 6.6¢ on the agent run.
- MiMo-V2.6-Pro is worth a look for code review and blog work. It has no reasoning mode and cost 0.44¢ on code review and 0.20¢ on the blog post, both estimated. Its LMArena Text score is 1480.
If quality matters more than price, the list is different. Claude Opus 5.5 scores 1507 on LMArena but cost 2.5¢ on the chat task (estimated), about 40 times GPT-6 Luna's 0.06¢.
Cheap list prices are a starting point, not an answer. Run your own workload through the calculator, and read my methodology to see how I count tokens, handle caps and label estimates.
FAQ
Is GPT-6 Luna the cheapest LLM API overall?
On the four tasks in this article and the 50-step agent run, it had the lowest cost per task of the models I compared. Your workload may differ, so check the task page that matches it.
Why can a model with a low token price cost more per task?
Reasoning tokens are billed as output tokens. Models that reason at length, or hit the 16,000-token output cap, can cost more than their per-token price suggests. Qwen3.8 Flash and GLM 5.3 Flash both hit the cap on code review.
Are all the costs in this article measured?
No. Each figure is measured, estimated or modeled. Measured costs come from one run per task with default settings through OpenRouter. The agent figures are modeled, and estimated figures are flagged on each task page.