AI Value Frontier: where does more money stop buying more quality?
We ran 24 prompts on every benchmarked model and recorded the score and what the runs cost. Plot one against the other and a pattern shows: on this benchmark, 36 of 38 models were matched (within 0.05 points) or beaten by a cheaper one. Find yours.
Benchmark run 2026-11-05. Costs on this page are what those runs cost on that day, not today's list prices. For a forecast at today's prices, use the AI Cost Simulator.
Score against cost
Green dots and the dashed line are the frontier: no cheaper model scored higher. Hover a dot for its numbers.
Am I overpaying for the model I use?
Points out of 10 below your model's score. Zero means as good or better only.
Benchmark-sized requests: short prompts and short answers. For your real token counts use the simulator.
GPT-5.5 scored 9.63 and cost $2.65 per 1,000 runs ($265 at 100K requests a month).
9 cheaper models scored at least as well, within 0.05 points:
| Model | Score | Cost per 1,000 | You would save | A month |
|---|---|---|---|---|
| GPT-5.4 Nano OpenAI | 9.75 (+0.13) | $0.096 | 96% | $255 |
| MiniMax M2.5 Minimax | 9.75 (+0.13) | $0.351 | 87% | $230 |
| MiniMax M3 Minimax | 9.79 (+0.17) | $0.352 | 87% | $230 |
| GPT-5.4 mini OpenAI | 9.58 (-0.04) | $0.436 | 84% | $221 |
| DeepSeek V4 Pro DeepSeek | 9.63 (+0.00) | $0.606 | 77% | $204 |
| Claude Sonnet 5 Anthropic | 9.79 (+0.17) | $1.03 | 61% | $162 |
| GPT-5.4 OpenAI | 9.67 (+0.04) | $1.26 | 52% | $139 |
| Grok 4.5 xAI | 9.67 (+0.04) | $1.62 | 39% | $102 |
A match on this benchmark is a reason to test, not a promise. Check the cheaper model on your own examples first.
The frontier models
The models nobody cheaper beat, from cheapest to most expensive.
| Model | Score | Per 1,000 runs | Check pass rate | Latency |
|---|---|---|---|---|
| GLM 5.3 Flash Z Ai | 9.38 | $0.021 | 97% | 1.0s |
| Llama 4 Maverick Meta · 3.8x the cost of GLM 5.3 Flash for 0.04 more points | 9.42 | $0.079 | 94% | 362ms |
| DeepSeek V4 Flash 0423 DeepSeek · 1.2x the cost of Llama 4 Maverick for 0.04 more points | 9.46 | $0.095 | 92% | 832ms |
| GPT-5.4 Nano OpenAI · 1x the cost of DeepSeek V4 Flash 0423 for 0.29 more points | 9.75 | $0.096 | 97% | 321ms |
| MiniMax M3 Minimax · 3.7x the cost of GPT-5.4 Nano for 0.04 more points | 9.79 | $0.352 | 97% | 1.2s |
How to read this honestly
- Each model ran 24 prompts: three for each of eight task types. They are short. Long documents, long chains of steps and unusual domains are not covered.
- Answers were scored by claude-haiku-4.5 against a rubric, plus automatic checks. A judge model has its own biases.
- Most passing models score within a point of each other. The benchmark is better at spotting a model that fails than at ranking two that pass.
- Cost is what the runs cost on the day they ran. Prices have moved since for some models, so use the AI Cost Simulator for a current forecast.
- Benchmark run on 2026-11-05. See the Model Playground for every prompt and answer.
Questions people ask
What is the value frontier?
Take every benchmarked model and plot its score against what it cost to run. The frontier is the set of models where nobody cheaper scored higher. Every model off the frontier has a cheaper one that did at least as well on our tests.
Does this mean I should switch to the cheapest model?
No. Our benchmark is short, with three prompts per task type, and most models that pass score near the top. It tells you which models are worth testing on your own work, not which one to trust blindly. Run a few hundred of your own examples before you switch.
Where does the cost come from?
It is what each benchmark run actually cost when we ran it, so a model that writes long answers or thinks a lot costs more here than its price per token suggests. It reflects prices on the day of the run, shown on the page. For a forecast at your own volume, use the AI Cost Simulator.
Why are some models missing?
Only models we have benchmarked appear, and a model is left out of the frontier if any of its benchmark runs failed. Other models may be good. We simply have no scores for them, and we do not guess.