Scale Labs SWE-Bench Pro V2 HARD (SWE-Bench Pro V2 HARD)
We show this table for reference; we do not rank on it.
A Scale Labs public leaderboard mirrored as display-only reference data. It does not affect BenchLM rankings.
Scale score on SWE-Bench Pro V2 HARD — November 12, 2026 snapshot
We mirror the published scale score view for SWE-Bench Pro V2 HARD. Opus 5 (Claude Code) xhigh leads the public snapshot at 98%, followed by Fable 5.1 (Claude Code) high (92.2%) and Gemini 4 Argon (mini-swe-agent) high (92.2%). We do not use these results to rank models overall.
Opus 5 (Claude Code) xhigh
Anthropic
SWE-Bench Pro V2 HARD
Fable 5.1 (Claude Code) high
Anthropic
SWE-Bench Pro V2 HARD
Gemini 4 Argon (mini-swe-agent) high
OpenAI
SWE-Bench Pro V2 HARD
18 modelsCodingCurrentDisplay onlyUpdated November 12, 2026 snapshot
Scale score table (18 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Opus 5 (Claude Code) xhighAnthropicSWE-Bench Pro V2 HARD | 98% | Not reported | Unknown |
| 2 | Fable 5.1 (Claude Code) highAnthropicSWE-Bench Pro V2 HARD | 92.2% | Not reported | Unknown |
| 3 | Gemini 4 Argon (mini-swe-agent) highOpenAISWE-Bench Pro V2 HARD | 92.2% | Not reported | Unknown |
| 4 | GPT-6 Astra (Codex) highOpenAISWE-Bench Pro V2 HARD | 90.2% | Not reported | Unknown |
| 5 | Muse Wonder 1.0 (mini-swe-agent) xhighOpenAISWE-Bench Pro V2 HARD | 90.2% | Not reported | Unknown |
| 6 | Kimi-K3 (mini-swe-agent) maxMoonshot AISWE-Bench Pro V2 HARD | 88.2% | Not reported | Unknown |
| 7 | Sonnet 5 (Claude Code) xhighAnthropicSWE-Bench Pro V2 HARD | 88.2% | Not reported | Unknown |
| 8 | GPT-5.6 Terra (Codex) xhighOpenAISWE-Bench Pro V2 HARD | 86.3% | Not reported | Unknown |
| 9 | Inkling 2 SmallThinking Machines LabSWE-Bench Pro V2 HARD | 86.3% | Not reported | Closed |
| 10 | Minimax M3p1 Flash (mini-swe-agent) xhighOpenAISWE-Bench Pro V2 HARD | 84.3% | Not reported | Unknown |
| 11 | GLM-5.3 (mini-swe-agent) maxZ.AISWE-Bench Pro V2 HARD | 84.3% | Not reported | Unknown |
| 12 | GPT-5.6 Sol (Codex) xhighOpenAISWE-Bench Pro V2 HARD | 82.4% | Not reported | Unknown |
| 13 | Muse Spark Open (mini-swe-agent) xhighOpenAISWE-Bench Pro V2 HARD | 82.3% | Not reported | Unknown |
| 14 | Sonnet 5.5 (mini-swe-agent) xhighAnthropicSWE-Bench Pro V2 HARD | 70.6% | Not reported | Unknown |
| 15 | Gemini 3.8 Flash (mini-swe-agent) highGoogleSWE-Bench Pro V2 HARD | 58.8% | Not reported | Unknown |
| 16 | Inkling (mini-swe-agent) xhighThinkingmachinesSWE-Bench Pro V2 HARD | 56.9% | Not reported | Unknown |
| 17 | Haiku 5.5 (mini-swe-agent) highAnthropicSWE-Bench Pro V2 HARD | 52.9% | Not reported | Unknown |
| 18 | Haiku 4.5 (Claude Code) xhighAnthropicSWE-Bench Pro V2 HARD | 25.5% | Not reported | Unknown |
Results reported by model developersUnranked
Labs ran these themselves. They take no rank here, and they are not comparable to the board score above.
| Model | Reported value | What was measured | Against the official board | Who ran it |
|---|---|---|---|---|
| SWE-2 FlashCognition | 89.9%Resolved rate | SWE-Bench Pro v2 HardCognition’s own run; no harness or trial details published | Not comparable to the board score | Reported by Cognition |
| Ripple-65B PreviewRicursive Intelligence | 79.1%Resolved rate | SWE-bench Pro v2.0 HardTerminus 2 harness; full BF16 precision, maximum effort | Not comparable to the board score | Reported by Ricursive Intelligence |
A lab’s own result is shown for reference. It never takes a rank on this board.
How we label sourcesHow to read this leaderboard
Compare the published configurations as complete evaluation systems. The source can combine a base model, agent scaffold, tools, budget, and inference setting in each result.
Operator receipt: 18 sourced rows are currently displayable on this page; the leading published row is Opus 5 (Claude Code) xhigh at 98%.
Honest limit: This Scale table is display-only context, not benchmark provenance or a weighted model-only comparison.
How BenchLM shows SWE-Bench Pro V2 HARD
BenchLM mirrors 18 published rows from Scale Labs’ public SWE-Bench Pro V2 HARD leaderboard, captured on November 12, 2026 snapshot.
The table is display only. It is useful context for a published agent or model configuration, but it does not enter BenchLM’s overall or category rankings.
Snapshot
The published SWE-Bench Pro V2 HARD snapshot places Opus 5 (Claude Code) xhigh first at 98%. The third row is 5.8 points behind. The broader top-10 range is 13.7 points, so the table still separates the published systems.
18 models have been evaluated on SWE-Bench Pro V2 HARD. 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. SWE-Bench Pro V2 HARD is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About SWE-Bench Pro V2 HARD
Year
2026
Tasks
18 published rows
Format
Published Scale leaderboard score
Difficulty
External agent and model evaluation
BenchLM mirrors 18 published rows from the SWE-Bench Pro V2 HARD public table captured on November 12, 2026 snapshot.
Freshness and provenance
Version
SWE-Bench Pro V2 HARD 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
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 SWE-Bench Pro V2 HARD measure?
A Scale Labs public leaderboard mirrored as display-only reference data. It does not affect BenchLM rankings.
Which model leads the published SWE-Bench Pro V2 HARD snapshot?
Opus 5 (Claude Code) xhigh currently leads the published SWE-Bench Pro V2 HARD snapshot with 98% scale score. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on SWE-Bench Pro V2 HARD?
The November 12, 2026 snapshot snapshot contains 18 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.