# Terminal-Bench

12 Hacker News commenters recommended Terminal-Bench and 11 criticized it, November 2025 to November 2026. It was recommended in 10 of 13 months.

**Verdict:** Recommended as a free, open-source test for comparing AI coding agents on real terminal tasks; critics say it is saturated and its tasks are often ambiguous or overly specific.

## Why people recommend it

- Best available proxy for coding agents (2 comments): "I think Terminal-Bench is currently the best proxy for "how good is this harness+agent combo at coding"" ([source](https://news.ycombinator.com/item?id=46296429)) "that bench has a lot of room to improve, but it is the best we can find now" ([source](https://news.ycombinator.com/item?id=49234642))
- Compares agent harnesses on one model (2 comments): "I think Terminal-Bench is currently the best proxy for "how good is this harness+agent combo at coding"" ([source](https://news.ycombinator.com/item?id=46296429)) "pretty popular for comparing different harnesses using the same underlying model" ([source](https://news.ycombinator.com/item?id=49429620))
- Covers longer, multi-step tasks (2 comments): "I found Terminal-Bench to be the most relevant for me, even for tasks that go far outside the terminal." ([source](https://news.ycombinator.com/item?id=45821072)) "They cover these longer horizon tasks that need more than just writing a bit of code in one file." ([source](https://news.ycombinator.com/item?id=47540332))
- Shows how far LLMs have come (2 comments): "The benchmark is actually pretty good. It shows in an extremely comprehensible way how far LLMs have come." ([source](https://news.ycombinator.com/item?id=46903862)) "This is a great way to become well-versed in benchmarking." ([source](https://news.ycombinator.com/item?id=48001612))

## What people warn about

- Tasks don't reflect real-world coding (3 comments): "it's painfully obvious how bad those benchmarks are and why they dont represent real world coding" ([source](https://news.ycombinator.com/item?id=49006193)) "I've looked at the Terminal-Bench questions and I just think they're irrelevant." ([source](https://news.ycombinator.com/item?id=48313578)) "some of the benchmarks they highlight might not represent the reality" ([source](https://news.ycombinator.com/item?id=50212575))
- Saturated; no longer separates models (2 comments): "2.1 is saturated - there is no world in which Gemini 3.8 Flash is producing better code" ([source](https://news.ycombinator.com/item?id=49575568)) "I'm certain has been saturated in post-training to no end" ([source](https://news.ycombinator.com/item?id=48870319))
- Ambiguous instructions, overly specific tests (2 comments): "the instructions are often ambiguous while the test cases are overly specific" ([source](https://news.ycombinator.com/item?id=48838823)) "verifiers are overly specific" ([source](https://news.ycombinator.com/item?id=48902453))
- Scoring flaws and reward hacking (2 comments): "40 of 83 gold-passing tasks still score 1 when the run also performs a destructive accident" ([source](https://news.ycombinator.com/item?id=49055101)) "both can be reward-hacked with simple exploits" ([source](https://news.ycombinator.com/item?id=47818130))

## Alternatives hackers recommend

- [SWE-Bench Pro](https://hackerapproved.com/p/swe-bench-pro.md): 5 recommend
- [SWE-bench](https://hackerapproved.com/p/swe-bench.md): 5 recommend

## Cite as

Hacker Approved, "Terminal-Bench", https://hackerapproved.com/p/terminal-bench/ (counts from 1,795,640 Hacker News comments, November 2025 to November 2026).