Best LLM for every budget, updated daily
TL;DR Highlight
A daily-updated site plots LLM benchmark scores against price to show the best model per budget.
Who Should Read
Backend developers choosing an LLM API who want to cut cost without losing quality
Core Mechanics
- It plots a value frontier
- It places Artificial Analysis Intelligence Index scores against blended API price.
- Only models with no cheaper and smarter alternative sit on the frontier.
- It gives a pick per budget
- Find your budget row; the pick is the highest-scoring model at that price.
- The runner-up is the next best that also fits.
- Blended price uses a 3:1 ratio
- Price per 1M tokens blends input and output at 3:1, as Artificial Analysis does.
- Cached-input discounts, batch pricing and fast modes are excluded.
- There is a cost-per-task view
- Models with the same token price can differ several times in cost because one spends far more tokens.
- The figure is taken as published by Artificial Analysis.
- Reasoning settings are separate rows
- Each effort level (low through max, plus non-reasoning) is benchmarked as its own model.
Community Reactions
- Per-token price is naive, some said
- Commenters said some models need 2-3x the tokens for the same intelligence.
- Several preferred Artificial Analysis's own cost per task.
- Build your own private eval, one said
- They argued benchmark noise grows when aggregated and combined with cost.
- They suggested testing hard cases from your own use.
- Overnight local model experience
- One user ran a quantized Qwen3.8 27B on a 32GB M1 Max overnight and said it debugged a Swift app well.
- They saw no real difference from q8 except speed.
- Subscription-based comparison is missing
- Commenters said most people use subscriptions or token packages and want usage-limit efficiency.
- Others noted some newest models and the coding and math tabs were out of date.
How to Apply
- Find your budget row and list candidates
- Note the pick and runner-up for your monthly budget.
- Raise Min score to your service's quality floor to filter weak models.
- Recheck candidates in cost-per-task view
- Compare the reasoning-effort rows of each candidate.
- Check whether a lower effort is better value than the top one.
- A/B test candidates on your traffic
- Run 100 real requests through the pick and runner-up and record tokens and accuracy.
- Recompute monthly cost with cache discounts and batch pricing.
Try It
# Paste into Claude or ChatGPT (quick model selection check)
I run a service that uses an LLM API and need to choose a model.
Monthly budget: {{monthly budget USD}}
Main task: {{task description}}
Candidate models and price per 1M tokens: {{model1, price}}, {{model2, price}}
Put the following in a table.
1. Estimated input and output tokens per task, with assumptions
2. Cost per task and monthly cost per model
3. The effect if any model uses more tokens
4. Three items I should verify myself in my service
Next to try: measure token counts on 20 real requests per model and ask to revise the assumptionsTerminology
Related Papers
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length
Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware