Skip to main content
Radar

Every change to the models you run, with its source and its date. Releases, price changes, retirements, API changes, and incidents.Every change to the models you run, with its source.

Follow model changes

Terminal-Bench 3.0

We mirror this table; we do not rank on it.

A continuously maintained benchmark for difficult computer work, including coding, deep learning, finance, engineering, math, and science tasks.

Current release

Terminal-Bench 4.0 is the current release. It recalibrates task resources, fixes 19 tasks, and removes eight, so its scores are not directly comparable with Terminal-Bench 3.0.

View Terminal-Bench 4.0

Tasks completed on Terminal-Bench 3.0 — September 18, 2026 snapshot

We mirror the published tasks completed view for Terminal-Bench 3.0. Claude Opus 5 leads the public snapshot at 42.7%, followed by GPT-5.6 Sol (34.6%) and Claude Fable 5 (34.0%). We do not weight the raw table directly. A normalized version can contribute through our external agentic consensus gate.

12 modelsAgenticCurrentDisplay onlyUpdated September 18, 2026 snapshot

Tasks completed table (12 models)

Score
1
Claude Opus 5Anthropic · Closedmax reasoning
42.7%
2
GPT-5.6 SolOpenAI · Closedmax reasoning
34.6%
3
Claude Fable 5Anthropic · Closedmax reasoning
34.0%
4
GLM-5.3Z.AI · Open weightmax reasoning
32.4%
5
Grok 4.6xAI · Closedhigh reasoning
26.5%
6
Claude Opus 4.8Anthropic · Closedmax reasoning
21.1%
7
GPT-5.6 TerraOpenAI · Closedmax reasoning
20.8%
8
18.6%
9
Grok 4.5xAI · Closedxhigh reasoning
15.7%
10
Claude Sonnet 5Anthropic · Closedmax reasoning
14.6%
11
GPT-5.6 LunaOpenAI · Closedmax reasoning
14.3%
12
GLM-5.2Z.AI · Open weightmax reasoning
4.6%

How we use Terminal-Bench 3.0

We mirror the Terminal-Bench 3.0 version 3.0 table: 74 professional computer-work tasks across 7 domains. The Terminal-Bench and Harbor team first released this benchmark under the FrontierBench name.

Each row measures a model and agent harness together, including Codex, Claude Code, mini-SWE-agent, or Cursor CLI. We therefore keep the raw table display-only. The normalized result counts as one external agentic benchmark family, and the ranking pipeline requires another benchmark family and evaluator before it can affect a model score.

Snapshot

12 launch rows74 tasks7 domainsv3.0Agentic consensus input

The published Terminal-Bench 3.0 snapshot places Claude Opus 5 first at 42.7%. The third row is 8.7 points behind. The broader top-10 range is 28.1 points, so the table still separates the published systems.

12 models have been evaluated on Terminal-Bench 3.0. The benchmark falls in the Agentic category. This category carries a 22% weight in BenchLM.ai's overall scoring system. Terminal-Bench 3.0 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About Terminal-Bench 3.0

Year

2026

Tasks

74 professional computer-work tasks across 7 domains

Format

Task completion rate

Difficulty

Frontier autonomous knowledge work

Terminal-Bench 3.0, formerly FrontierBench, launched as version 0.1 with 74 tasks across seven domains. Each public row combines a model with an agent harness, so we keep the raw results display-only. Its normalized score counts as one external agentic benchmark family and cannot make a category eligible without independent corroboration.

Freshness and provenance

Version

Terminal-Bench 3.0

Refresh cadence

Continuous

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Questions

What does Terminal-Bench 3.0 measure?

A continuously maintained benchmark for difficult computer work, including coding, deep learning, finance, engineering, math, and science tasks.

Which model leads the published Terminal-Bench 3.0 snapshot?

Claude Opus 5 currently leads the published Terminal-Bench 3.0 snapshot with 42.7% tasks completed. BenchLM keeps the raw table display-only. A normalized version can contribute through the external agentic consensus gate, which still requires independent corroboration.

How many models are evaluated on Terminal-Bench 3.0?

The September 18, 2026 snapshot snapshot contains 12 AI models.

Last updated: September 18, 2026 snapshot · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 2,000+ readers.

One email each week. Unsubscribe anytime.