ProgramBench — Mercor evaluation (ProgramBench (Mercor))

We show this table for reference; we do not rank on it.

Reconstructing a program from its executable behavior and documentation. This table shows Mercor-run configurations for reference and is excluded from model rankings.

Evaluation results published by Mercor. Published evaluation aggregates only; task contents and dataset license grants are not included.

Pass@1 on ProgramBench (Mercor) — October 15, 2026 capture

We mirror the published pass@1 view for ProgramBench (Mercor). Opus 5.5 leads the public snapshot at 20.00%, followed by Fable 5.1 (11.30%) and Fable 5.1 (10.70%). We do not use these results to rank models overall.

Opus 5.5

Anthropic

mini-swe-agent (1000 max steps, 6 hour time limit) · max reasoning

Published margin ±6.7 · 300 reported samples

20.00%

Fable 5.1

Anthropic

mini-swe-agent (1000 max steps, 6 hour time limit) · high reasoning

Published margin ±5.7 · 300 reported samples

11.30%

Fable 5.1

Anthropic

mini-swe-agent (1000 max steps, 6 hour time limit) · max reasoning

Published margin ±5 · 300 reported samples

10.70%

24 configurationsCodingDated Mercor configurationsDisplay onlyUpdated October 15, 2026 capture

Pass@1 table (24 configurations)

Pass@1 results for ProgramBench (Mercor)
Rank	Model / configuration	Score	Parameters (B)	Open / closed
1	Opus 5.5Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±6.7 · 300 reported samples	20.00%	Not reported	Unknown
2	Fable 5.1Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±5.7 · 300 reported samples	11.30%	Not reported	Unknown
3	Fable 5.1Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±5 · 300 reported samples	10.70%	Not reported	Unknown
4	GPT-6 AstraOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±4.7 · 300 reported samples	8.00%	Not reported	Unknown
5	GPT-6.1 SolOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±4.2 · 300 reported samples	6.30%	Not reported	Unknown
6	Opus 5Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±3.3 · 300 reported samples	4.30%	Not reported	Unknown
7	Haiku 5.5Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±2.3 · 300 reported samples	3.00%	Not reported	Unknown
8	GPT-6 SolOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±2.5 · 300 reported samples	2.30%	Not reported	Unknown
9	Grok 4.6xAImini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±2.2 · 300 reported samples	1.70%	Not reported	Unknown
10	Gemini 3.7 FlashGooglemini-swe-agent (1000 max steps, 6 hour time limit) · medium reasoningPublished margin ±1.5 · 300 reported samples	1.00%	Not reported	Unknown
11	GPT-5.5OpenAImini-swe-agent (1000 max steps, 6 hour time limit) · xhigh reasoningPublished margin ±1.5 · 300 reported samples	1.00%	Not reported	Unknown
12	GPT-5.6 SolOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · xhigh reasoningPublished margin ±1.5 · 300 reported samples	1.00%	Not reported	Unknown
13	Gemini 3.8 FlashGooglemini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±1.3 · 300 reported samples	1.00%	Not reported	Unknown
14	GPT-6 LunaOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±1.5 · 300 reported samples	1.00%	Not reported	Unknown
15	GPT-5.6 TerraOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±1.5 · 300 reported samples	1.00%	Not reported	Unknown
16	Sonnet 5Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±0.5 · 300 reported samples	0.30%	Not reported	Unknown
17	Gemini 3.6 FlashGooglemini-swe-agent (1000 max steps, 6 hour time limit) · medium reasoningPublished margin ±0.5 · 300 reported samples	0.30%	Not reported	Unknown
18	DeepSeek-V4-Pro-0813DeepSeekmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±0.5 · 300 reported samples	0.30%	Not reported	Unknown
19	GLM-5.2Zhipumini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±0.5 · 300 reported samples	0.30%	Not reported	Unknown
20	Fable 5Anthropicmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±0 · 300 reported samples	0.00%	Not reported	Unknown
21	InklingThinking Machinesmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±0 · 300 reported samples	0.00%	Not reported	Unknown
22	MiniMax-M3MiniMaxmini-swe-agent (1000 max steps, 6 hour time limit) · high reasoningPublished margin ±0 · 300 reported samples	0.00%	Not reported	Unknown
23	GPT-5.6 LunaOpenAImini-swe-agent (1000 max steps, 6 hour time limit) · max reasoningPublished margin ±0 · 300 reported samples	0.00%	Not reported	Unknown
24	Grok 4.7xAImini-swe-agent (1000 max steps, 6 hour time limit) · xhigh reasoningPublished margin ±0 · 300 reported samples	0.00%	Not reported	Unknown

How to read this Mercor evaluation

We captured Mercor's ProgramBench results on October 15, 2026. The table preserves 24 published configurations, their source identifiers, reasoning settings, and disclosed pass@1 values. These are results from Mercor's evaluation setup; we did not rerun them.

This dated source table is display only and does not enter overall or category rankings. Source configurations retain their external identity, including Pro descriptors and different effort settings.

The headline is the source page's selected Pass@1 metric. The task set, harness, tools, and grader belong to this Mercor run and may differ from another evaluator's results.

The row labels retain published error margins and sample counts when supplied. Missing margins and counts remain unreported; a sample count is not used to reconstruct a task denominator. Except where the dedicated source defines an interval, the confidence level is unreported. Equal displayed scores and overlapping margins do not establish statistical ties. Capture time does not establish individual evaluation dates.

The general methodology lists programmatic unit tests (all must pass) and a judge of None. Its run-count label is k = 3. The linked result page takes precedence where its setup differs, as noted above.

Snapshot

Pass@124 configurationsMercor evaluationDisplay only

Mercor ProgramBench results Mercor APEX benchmark index Mercor methodology Source paper Upstream evaluation code Upstream dataset

The published ProgramBench (Mercor) snapshot places Opus 5.5 first at 20.00%. The third row is 9.30 points behind. The broader top-10 range is 19.00 points, so the table still separates the published systems.

24 configurations are shown for ProgramBench (Mercor). The benchmark falls in the Coding category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. ProgramBench (Mercor) is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About ProgramBench (Mercor)

Tasks

100 source-reported tasks

Format

Pass@1

Difficulty

Source-specific evaluation

The selected headline is Pass@1. Source model identifiers, effort settings, published margins, and reported sample counts remain attached to the source configuration. Public-set results and Mercor task extensions occupy separate tables.

ProgramBench evaluation results Public benchmark source

Freshness and provenance

Version

ProgramBench; Mercor Pass@1

Refresh cadence

Reviewed dated source capture

Staleness state

Dated Mercor configurations

Question availability

Upstream task access and terms; only aggregate results mirrored

Dated Mercor configurationsDisplay 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 ProgramBench (Mercor) measure?

Reconstructing a program from its executable behavior and documentation. This table shows Mercor-run configurations for reference and is excluded from model rankings.

Which model leads the published ProgramBench (Mercor) snapshot?

Opus 5.5 currently leads the published ProgramBench (Mercor) snapshot with 20.00% pass@1. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on ProgramBench (Mercor)?

The October 15, 2026 capture snapshot contains 24 source configurations.

Last updated: October 15, 2026 capture · 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 5,600+ readers.

One email each week. Unsubscribe anytime.