Thinking Machines Inkling 2 Mini at launch
- Thinking Machines
- Released November 21, 2026
- 1.05M context
Thinking Machines released Inkling 2 Mini on November 21, 2026. Independent boards have not scored it on enough tests for the LLMBenchmark Index yet; below are the results Thinking Machines reported and the independent scores published so far. Thinking Machines's own announcement reports 51 benchmark results, including SWE-bench Verified (79.2%) and JobBench (49.6). It costs $0.0775 per million input tokens and $0.325 per million output tokens on OpenRouter.
What Thinking Machines reported at launch
From “Inkling 2 Mini”, November 21, 2026. The maker's own numbers, not re-run by us.
-
AA-AnalystAgent
46.3%
#2 of 2 models with a reported result
-
AA-Briefcase v1.1
1263
#9 of 10 models with a reported result
-
AA Index v5.0
20.1
#4 of 4 models with a reported result
-
AA-Omniscience v1.1 non-hallucination rate
63.4%
#2 of 2 models with a reported result
-
Apex 2025
34.4%
-
ARC-AGI v1 / v2
49.3%
-
ARC-AGI v1 / v2
84.5%
-
AutomationBench-AA v1.0.6
41.8%
-
BrowseComp Inkling 2 Mini with compaction
79.0%
-
BVB
68.4%
#2 of 2 models with a reported result
-
Chartography original / with Python
22.1%
-
Chartography original / with Python
72.5%
#2 of 2 models with a reported result
Show all 51
-
ComplexConstraints v1.1 openhands v1.45.0
33.4%
-
Covert misbehavior (Internal) ↓
8.3%
-
CritPt v2 Full / Hard
31.4%
#2 of 2 models with a reported result
-
CritPt v2 Full / Hard
17.0%
-
DeepSearchQA Inkling 2 Mini with compaction
85.3%
-
DeepSWE v1.1 mini-swe-agent v2.4.6
48.2%
#2 of 2 models with a reported result
-
Destructive actions (Internal) ↓
10.2%
-
Diligence Bench Best harness
48.6%
#2 of 2 models with a reported result
-
EnterpriseOps-Gym
42.0%
#2 of 4 models with a reported result
-
FORTRESS (adversarial) Public
89.5%
-
FORTRESS (benign) Public
96.4%
-
FrontierCode v2.0 stirrup v0.2.1
15.1%
#2 of 2 models with a reported result
-
GDP.pdf
16.8%
#9 of 10 models with a reported result
-
GDPval-AA 2.1
1319
#8 of 10 models with a reported result
-
GraphWalks BFS 256K
69.7%
-
HLE-Diamond
15.2%
#4 of 4 models with a reported result
-
JobBench
49.6
#14 of 21 models with a reported result
-
LongBench v2
66.9%
#3 of 6 models with a reported result
-
MCP Atlas (Public)
82.7
#1 of 3 models with a reported result
-
Misbehavior (Internal) ↓
9.8%
-
MMMU Pro (Standard 10)
74.9%
#2 of 4 models with a reported result
-
MRCR v2 8-needle 256K–512K
46.4%
-
MuirBench
77.5%
#2 of 3 models with a reported result
-
OmniDocBench v1.6
92.8%
#2 of 2 models with a reported result
-
PerceptionBench
57.5%
#4 of 5 models with a reported result
-
PMPP-Hard (AUC, 0–10h) pi v0.83.0
455.8
#2 of 2 models with a reported result
-
ScreenSpot Pro original / with Python
87.7%
#2 of 2 models with a reported result
-
ScreenSpot Pro original / with Python
76.5%
-
StrongREJECT (none)
97.8%
#1 of 2 models with a reported result
-
SWE-bench Verified
79.2%
#7 of 30 models with a reported result
-
SWEBench Multilingual Inkling 2 Mini with bash only
73.7%
-
SWEBench Pro v2 (Hard) mini-swe-agent v2.4.6
77.9%
#2 of 2 models with a reported result
-
Terminal Bench v2.1 Inkling 2 Mini in pi v0.83.0
71.6%
-
Terminal Bench v4.0 mini-swe-agent v2.4.6
5.6%
-
Vals Finance Index
50.1%
-
Vals Index v2.2
46.5%
#4 of 4 models with a reported result
-
Vibe Code Bench v1.1 openhands (vals)
77.9%
#2 of 2 models with a reported result
-
ZeroBench (main) original / with Python
17.0%
-
ZeroBench (main) original / with Python
34.0%
#2 of 2 models with a reported result
Independent results so far
29 scores from 29 boards that test models themselves. Ranks compare Inkling 2 Mini with every model on the same board.
Coding (9)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| CyberBench (Vals AI) | 74.6% | #9 of 42 | — | Vals AI boards |
| ProgramBench (Vals AI) | 0.5% | #18 of 62 | — | Vals AI boards |
| SWE-bench Multimodal (Vals AI) | 37.9% | #18 of 21 | — | Vals AI boards |
| SkillsBench (Vals AI) | 43.2% | #27 of 34 | OpenHands | Vals AI boards |
| FrontierCode 2.0 | 15.1% | #41 of 62 | — | Artificial Analysis |
| Vibe Code Bench (Vals AI) | 77.9% | #41 of 110 | OpenHands | Vals AI boards |
| IOI (Vals AI) | 11.6% | #42 of 43 | — | Vals AI boards |
| Terminal-Bench 4 (Vals AI) | 6.6% | #53 of 63 | Mini-SWE-agent | Vals AI boards |
Show all 9
| Terminal-Bench 4.0 (Artificial Analysis run) | 5.6% | #58 of 117 | — | Artificial Analysis |
|---|
Agentic work (7)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Vals Science Index | 23.6% | #27 of 27 | — | Vals AI boards |
| SREBench (Vals AI) | 0.0% | #30 of 34 | — | Vals AI boards |
| BioMysteryBench (Vals AI) | 47.8% | #33 of 33 | — | Vals AI boards |
| Harvey's Legal Agent Benchmark (Vals AI) | 5.0% | #33 of 40 | — | Vals AI boards |
| Terminal-Bench-Science (Vals AI) | 0.0% | #42 of 45 | Mini-SWE-agent | Vals AI boards |
| Tax Agent Bench (Vals AI) | 14.1% | #65 of 83 | — | Vals AI boards |
| Legal Research Bench (Vals AI) | 17.8% | #70 of 89 | — | Vals AI boards |
General (4)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| SAGE (Vals AI) | 85.9% | #14 of 29 | — | Vals AI boards |
| MedScribe (Vals AI) | 86.4% | #24 of 84 | — | Vals AI boards |
| Vals Index | 46.5% | #37 of 40 | — | Vals Index |
| GDPval-AA | 1319 | #51 of 118 | — | Artificial Analysis |
Reasoning (3)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| MysteryMechanism (Vals AI) | 23.0% | #22 of 32 | — | Vals AI boards |
| CritPt v2, hard subset (Artificial Analysis run) | 17.0% | #48 of 62 | — | Artificial Analysis |
| HLE-Diamond (Artificial Analysis run) | 15.2% | #48 of 79 | — | Artificial Analysis |
Web design (3)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Design Arena: 3D | 1308 | #23 of 127 | — | Design Arena |
| Design Arena: UI components | 1292 | #25 of 130 | — | Design Arena |
| Design Arena: game dev | 1287 | #29 of 134 | — | Design Arena |
Knowledge (2)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| AA-Omniscience | -8.6 | #51 of 65 | — | Artificial Analysis |
| MedCode (Vals AI) | 25.7% | #80 of 82 | — | Vals AI boards |
Intelligence (1)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| AA Intelligence Index | 20.1 | #44 of 65 | — | Artificial Analysis (via OpenRouter) |