LLM models and token costs
LLM Models and Token Costs: OpenAI, Anthropic, OpenRouter and Hugging Face
A working guide to the model layer of the stack: which language models are current, how each provider prices tokens, where open models and gateways fit, and how model choice connects to the hardware and agents beneath and above it.
Prices and versions: figures are published list prices in US dollars per 1 million tokens, read from each provider’s own pages on 18 November 2026. Models, prices and promotions change often, so confirm on the provider’s page before you budget.
The four sources and what each is for
OpenAI and Anthropic make the leading closed models and sell them by the token. OpenRouter is a gateway that routes one API to many providers. Hugging Face is the home of open-weight models and a way to run them on partner hosts or dedicated GPUs.
Token prices side by side
List prices per million input and output tokens, as published on 18 November 2026. Tiers are an approximate guide to how teams use each model, not a quality ranking.
| Tier | OpenAI | Anthropic | Open models (hosted) |
|---|---|---|---|
| Budget, high volume | GPT-6 Luna: $0.10 / $0.50 | Haiku 5.5: $0.10 / $0.50 (prompts up to 100K) | gpt-oss-120b: $0.04 / $0.17; GLM-5.3-Flash: $0.15 / $0.50 |
| Value, everyday production | GPT-6.1 Terra: $1 / $5 | Haiku 4.5: $1 / $5 | DeepSeek-V4.1-Flash-1017: $0.20 / $0.60 |
| Balanced workhorse | GPT-6.1 / 6.2 Sol: $2 / $10 | Sonnet 5.5: $2 / $10 | GLM-5.3: $0.90 / $4; Kimi-K3.1-Pro: $2 / $10 |
| Strong reasoning and agents | GPT-6.2 Astra: $8 / $40 | Opus 5.5: $4 / $20 | Kimi-K3: $2.85 / $14.25 |
| Top tier | GPT-6 Astra: $10 / $50 | Fable 5.2: $8 / $40; Fable 5.1: $10 / $50 | Not directly comparable |
What a monthly bill looks like
The same example everywhere: 100 million input tokens and 20 million output tokens a month, with no caching, batching or discounts.
| Model | Monthly cost: 100M input + 20M output tokens |
|---|---|
| gpt-oss-120b (hosted) | about $7 |
| GPT-6 Luna or Claude Haiku 5.5 | $20 |
| GLM-5.3-Flash (hosted) | about $25 |
| GPT-6.1 Terra | $200 |
| GPT-6.1 Sol or Claude Sonnet 5.5 | $400 |
| Claude Opus 5.5 | $800 |
| GPT-6.2 Astra or Claude Fable 5.2 | $1,600 |
| GPT-6 Astra or Claude Fable 5.1 | $2,000 |
The spread is roughly a factor of 300 between the cheapest hosted open model and the top closed tiers. Price per task is what matters, so measure cost per successful result on your own data.
How token pricing works
- Input and output are priced separately. Output tokens cost several times more than input, so long answers and reasoning traces drive the bill.
- Caching. Reused prompt prefixes bill at a small fraction of normal input.
- Batch and flex tiers. Roughly half price for work that can wait.
- Long context. Some models charge more for very long prompts.
- Tokens are not equal. Each vendor tokenises differently, so the same text can count differently.
- Agents multiply usage. A single agent task can make dozens of model calls, so cheap tiers and caching matter more than in chat.
Choosing models for a real system
- Route, do not standardise. Send simple work to cheap models and escalate hard cases, using a router or gateway.
- Evaluate on your data. Keep a small test set and compare cost per success, not public benchmarks.
- Plan for change. Models are replaced and retired within months. Pin versions, track lifecycle dates and keep a tested fallback.
- Decide where data may go. Check logging, retention, regional processing and export or sovereignty rules for each provider.
- Know when to bring it in-house. High, steady volume, sensitive data or latency needs can justify hosted or self-run open models.
From tokens to hardware
Behind every token price is hardware: GPUs, memory, networking, power and cooling. Open models served on NVIDIA, AMD or other accelerators have a cost per token that depends on utilisation, batch size and software such as CUDA or ROCm. That is why we connect model choice to hosted AI models, neocloud capacity, open-source models and the full stack, and why the agent platform and agent security and defence work sit on top.
Where this fits in the stack
Common questions
Which LLM is the cheapest?
Among the sources we read on 18 November 2026, open models hosted by third parties are the cheapest (for example gpt-oss-120b at $0.04 input and $0.17 output per million tokens), followed by the small closed tiers such as GPT-6 Luna and Claude Haiku 5.5 at $0.10 and $0.50.
Why do output tokens cost more than input tokens?
Generating tokens uses more compute per token than reading them, so providers price output at a multiple of input, commonly three to six times.
Should I use OpenRouter or go direct to OpenAI and Anthropic?
Go direct when you use one or two models and want the simplest path and direct support. A gateway helps when you want many models behind one API, price-based routing and provider failover, with the trade-off of an extra party and credit fees.
Where do open models fit?
They are often the lowest cost per token and give more control over data, but they need evaluation, and either a hosted provider or your own GPUs to serve them.
How current are these prices?
They are list prices read from the providers’ own pages on 18 November 2026. Prices, promotions and model lineups change often, so confirm before budgeting.
Sources and dating
Sources: OpenAI developer documentation, Anthropic documentation, OpenRouter models, rankings, pricing and FAQ pages, and Hugging Face Inference Providers, pricing and Hub pages, all read 18 November 2026. See each provider page for specific caveats. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against the provider pages before we size or budget a build. All provider, product and model names are trademarks of their owners. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: Open-source AI models · Hosted AI models · AI application layer · AI stack explorer
Want a model and cost review for your workload?
Model and cost review
