Benchmarks / Microsoft STATE-Bench

Reported by Microsoft STATE-Bench

Microsoft STATE-Bench

Share of customer support tasks completed correctly, averaged over repeated runs.

Data older than 90 days: last updated 1 Jul 2026

Results dated
1 Jul 2026
Models
2
Unit
% of tasks
Licence
MIT License

Microsoft STATE-Bench in brief

What it measures. Microsoft STATE-Bench is a third-party benchmark, shown here with credit. Measured here: Share of customer support tasks completed correctly, averaged over repeated runs.

Latest results. , 2 models ranked.

Top 2 on Customer support task success (pass@1) (% of tasks, higher is better):

  1. GPT-5.4 (high reasoning) (OpenAI): 59.0%
  2. GPT-5.5 (high reasoning) (OpenAI): 59.0%

The first 2 are too close to tell apart, so read the order among them as a tie.

How to cite. Microsoft STATE-Bench, via Spring Prompt, results dated 1 July 2026. https://springprompt.com/benchmarks/statebench These results belong to Microsoft STATE-Bench, not to us. Licence: MIT License. Credit Microsoft STATE-Bench first.

Customer support task success (pass@1):

2 results · % of tasks, higher is better · ≈ cannot be told apart from the leader. Choose a model to highlight it.Clear highlight

  1. 1≈ GPT-5.4 (high reasoning)OpenAI 59.0%
  2. 1≈ GPT-5.5 (high reasoning)OpenAI 59.0%

Full results

STATE-Bench v0.8: customer support task success (pass@1), % of tasks, higher is better
#ModelCustomer support task success (pass@1)
% of tasks, higher is better
Price
$ per million tokens, in / out
1≈ GPT-5.4 (high reasoning)OpenAI
59.0%
$2.50 / $15
1≈ GPT-5.5 (high reasoning)OpenAI
59.0%
$5 / $30

Swipe the table sideways for more columns.

Source: Microsoft STATE-Bench, via Spring Prompt, springprompt.com/benchmarks/statebench. Test your own prompt on these models →

Ranks follow the score as shown, so equal numbers share a rank. Results as published by Microsoft STATE-Bench; we do not re-run them.

What it measures

Share of customer support tasks completed correctly, averaged over repeated runs.

What it does not measure

Not the model alone: it runs inside STATE-Bench's agent loop against simulated users, and versions of the benchmark are not comparable.

4 results from Microsoft STATE-Bench not ranked here · show why

We rank a result only when we can tie it to a specific model you can use. These are left out:

  • Agent systems with memory, not models on their own: 4

STATE-Bench by Microsoft and STATE-Bench contributors, MIT License.