LLM API pricing compared: what a task really costs in 2026
Per-million-token prices hide the shape of real work. Here is what a chat reply, a document summary, a code edit and an agentic session cost on today's models, with and without prompt caching.
Every provider prices the same way, per million tokens in and per million tokens out, and every comparison stops there. That is the wrong unit. A chat reply is four hundred tokens in and three hundred out; an agentic coding session is sixty thousand in and four thousand out; the ratio between input and output price matters more than either number alone. This page prices real workloads. Numbers as of 9 Sept 2026.
The cheapest models that are still good
"Cheap" only matters above a quality floor. These are the highest-index models below fifty cents per million tokens, blended at three input tokens per output token:
| # | Model | Blended $/M | Index | Price $/M |
|---|---|---|---|---|
| 1 |  GLM-5.3-Flash Zhipu AI | $0.119/M | 61.0 | $0.119 |
| 2 |  GPT-5.6 Luna xhigh OpenAI | $0.450/M | 59.3 | $0.450 |
| 3 |  DeepSeek V4 Flash high DeepSeek | $0.168/M | 59.3 | $0.168 |
| 4 |  Qwen3.8-Flash-Next Alibaba | $0.230/M | 59.3 | $0.230 |
| 5 | Qwen3.5 397B-A17B Alibaba | $0.387/M | 58.6 | $0.387 |
As of 9 Sept 2026. Full list.
And below two dollars, where most of the practical frontier lives:
| # | Model | Blended $/M | Index | Price $/M |
|---|---|---|---|---|
| 1 |  Gemini 3.7 Flash medium Google | $1.50/M | 64.4 | $1.50 |
| 2 | Gemini 3.8 Flash medium Google | $1.50/M | 63.5 | $1.50 |
| 3 | Gemini 3 Flash thinking Google | $1.13/M | 61.5 | $1.13 |
| 4 | GLM-5.3-Flash Zhipu AI | $0.12/M | 61.0 | $0.119 |
| 5 | Gemini 3.6 Flash Google | $1.50/M | 60.7 | $1.50 |
As of 9 Sept 2026. Full list.
Quality per dollar
Kimi K3 sits within 5 points of the leader at $6.00 per million tokens, 3.7× cheaper than GPT-6 Astra (max). That ratio is the single most useful number on this site for anyone with a budget, and it moves every few weeks as prices are cut. The value board ranks every model by index per dollar; the price cuts table lists every listed price that fell in the last ninety days.
Five workloads, priced
The per-task boards price five representative jobs from each model's list price:
| Workload | Tokens in / out | What it stands for | |---|---|---| | Chat reply | 400 / 300 | one assistant turn | | Summarise a 30-page report | 12,000 / 600 | document in, short answer out | | Code edit | 6,000 / 1,500 | files and instructions in, a diff out | | Agentic coding session | 60,000 / 4,000 | one long tool-using loop | | Structured extraction | 2,000 / 200 | text in, JSON out |
Two things jump out of those tables. First, input-heavy work (summaries, agent loops) is dominated by the input price, and the input price is where providers compete hardest. Second, output-heavy work with reasoning models can cost far more than the list price suggests, because reasoning tokens are billed as output and are not shown on the price page. We do not model reasoning tokens; treat the agentic numbers as a floor for max-effort configurations.
Prompt caching changes the ranking
Most providers now sell cached input tokens at a fraction of the fresh price. For an agent loop that resends the same context on every step, three-quarters of the input can be cache hits. The "with caching" column on the cost-per-task boards applies exactly that assumption, and it reorders the table: models with an aggressive cache discount move up several places.
What the prices do not include
- Batch tiers, which are often half price for asynchronous work.
- Regional or committed-use discounts.
- Reasoning tokens for thinking models, as above.
- Provider markups and discounts on the same open-weights model, which can span 3× between hosts. The "where to run it" table on each model page lists every host's price alongside its measured speed.
How to use this
Start from the quality you need, not the price you want to pay. Take the best model under your ceiling, check its per-task cost with caching, and compare it against the leader on the compare page. Most teams find that the second-cheapest model on the frontier, not the cheapest, is the one that stops causing rework.