Open Models on CursorBench 3.2: Grok 4.5, GLM 5.2, Kimi…

Cluster: CursorBench 3.2 hub · Benchmark lenses · Best model by work mode

What are open models on CursorBench 3.2?

Open-weight and xAI coding models (Grok 4.5, GLM 5.2, Kimi K2.7 Code) appear on CursorBench 3.2 with score, cost, tokens, and steps per Cursor agent task. LongCat 2.0 publishes strong SWE-bench rows on its vendor site but no CursorBench row at the time of writing.

Who it is for: Anyone comparing open-model launch posts to what Cursor sessions actually cost — founders, students on a budget, operators, and procurement reviewers alike.

What you will learn: CursorBench rows for Grok, GLM, and Kimi on the 3.2 battery, how LongCat's vendor table differs, and when public numbers justify a policy change.

Open models entered everyday conversations with two stories. Vendors cite SWE-bench and long-context wins. CursorBench cites ambiguous multi-file agent sessions with a bill attached.

Those stories overlap. They are not the same test — and you pay for the difference when defaults follow the wrong chart.

Why open models need two tables

Source	Measures	Open-model signal
CursorBench 3.2	Cursor agent loop, cost, steps; adds instruction-following + tool-use tasks	Budget reality in the IDE
Vendor evaluations	SWE-bench, Terminal-Bench, etc.	Coding quality under vendor harness
Missing row	LongCat on CursorBench	Cannot price LongCat per Cursor task yet

Read both. Do not rank LongCat against Composer on Cursor dollars until a CursorBench row exists.

CursorBench rows (benchmarked)

Benchmarked numbers (from CursorBench 3.2):

Model	Score	Cost / task	Tokens / task	Steps / task
Grok 4.5 High*	66.7%	$1.51	19,521	33
Grok 4.5 Medium*	65.4%	$1.54	18,914	34
Grok 4.5 Low*	63.5%	$1.22	15,841	31
Composer 2.5 (reference)	56.1%	$0.44	14,286	33
GLM 5.2 Max	55.0%	$1.76	35,946	58
GLM 5.2 High	51.5%	$1.19	21,829	49
Kimi K2.7 Code	49.7%	$1.43	31,247	58

* Cursor states Grok 4.5 may have an advantage because Cursor codebase snapshot was unintentionally included in training; exact impact unclear (evals disclaimer).

Grok 4.5 (xAI)

Grok 4.5 ships as Low / Medium / High effort tiers on CursorBench 3.2. Grok 4.5 High lands 66.7% at $1.51 and 33 steps — above Composer on score with similar step count, at roughly 3.4× Composer's benchmark cost.

Practical read: Grok is the headline score-per-dollar challenger to Fable on the public table, but treat the row as directional until you read Cursor's training-data footnote. Do not swap team defaults without validating on real program tasks.

GLM 5.2 (Z.ai)

Z.ai positions GLM 5.2 for long-horizon agent work with 1M-token context and MIT licensing. Vendor tables cite FrontierSWE, Terminal-Bench, and SWE-bench Pro.

On CursorBench 3.2: GLM 5.2 Max lands 55.0% at $1.76 with 58 steps. That is ~1.1 points below Composer 2.5 with nearly double the steps (58 vs 33).

Practical read: GLM remains credible when context length or self-hosting matters. On ambiguous Cursor sessions on the public table, it does not beat Composer on score or step efficiency.

Kimi K2.7 Code (Moonshot)

Kimi K2.7 Code targets agentic coding with thinking mode required in Kimi Code workflows.

On CursorBench 3.2: 49.7% at $1.43 and 58 steps. Score per dollar is mid-pack (~35 when derived from public rows), but 58 steps is heavy agent churn for a sub-50% score.

Practical read: Inexpensive exploratory model. Not a drop-in Composer replacement on this harness.

LongCat 2.0: vendor SWE-bench strength, no CursorBench row

LongCat 2.0 is Meituan's agentic coding model (MIT license, 1M context, MoE architecture). It ran on OpenRouter as Owl Alpha before launch.

Meituan publishes in-house Evaluations with SWE-bench and Terminal-Bench rows. Example coding rows from their table (vendor-reported):

Benchmark	LongCat-2.0	Notes
SWE-bench Pro	59.5	Compared against GPT-5.5 (58.6) and Gemini (54.2) in vendor table
Terminal-Bench 2.1	70.8	Tight cluster with Gemini 70.7*
SWE-bench Multilingual	77.3	Opus 4.8 leads at 84.8* in same table

* = external score per LongCat's footnotes.

CursorBench gap: LongCat does not appear in CursorBench 3.2 yet. You cannot line it up against Fable 5, Grok 4.5, or Composer on cost and steps per Cursor session.

Practical read: LongCat is a watchlist model for open coding quality. Wait for a CursorBench row before changing Cursor defaults based on vendor SWE-bench alone. See benchmark comparison.

Derived efficiency (from public CursorBench rows)

Model	Score / USD (derived)	Score / step (derived)
Composer 2.5	127.5	1.70
Grok 4.5 Low*	52.0	2.05
Grok 4.5 High*	44.2	2.02
Kimi K2.7 Code	34.8	0.86
GLM 5.2 High	43.3	1.05
GLM 5.2 Max	31.3	0.95

Composer 2.5 remains the extreme score-per-USD outlier on the table. Grok tiers trade higher benchmark cost for higher score — read Cursor's Grok footnote before treating that as a fair fight.

When to pick an open model in Cursor

Situation	Sensible pick
Daily agent tasks on budget	Composer 2.5
Higher score, still under ~2 USD/task	Grok 4.5 Low/Medium* (validate caveat)
Open-weight requirement, mid budget	Kimi K2.7 for experiments; watch steps
1M context, self-host path	GLM 5.2 outside Cursor or when context is the bottleneck
Vendor SWE-bench leader, no Cursor row	LongCat: monitor; do not assume Cursor economics

Limitations

Grok 4.5 scores may be inflated by training-data overlap per Cursor's disclaimer.
Vendor long-horizon scores may use hour-scale runs CursorBench does not simulate.
Moonshot and Z.ai harnesses differ from Cursor's agent loop.
License and hosting constraints are out of scope for this table; verify terms before production use.

Reader action

Open CursorBench 3.2 and confirm Grok, GLM, and Kimi rows still match.
Read the Grok 4.5 footnote on the evals page before changing defaults.
Run one task on Kimi or Grok Low and log steps versus Composer on the same prompt.
Bookmark LongCat evals; re-check when Cursor adds a row.
Read Fable tier pricing if you escalate from budget open models to frontier closed models.

Open Models on CursorBench 3.2: Grok 4.5, GLM 5.2, Kimi K2.7, and LongCat