vs
13:1
overall capability wins (3 ties across 17 scored). Granite 4.2 30B leads overall.
- Reasoning5:1
- Coding6:0
- Tool use1:0
- Specs1:0
Verdict
Granite 4.2 30B leads coding (widest gap: +14.2 on SWE-Pro).
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Granite 4.2 30B
- -
- Granite 4.2 8B
- -
Price per 1M tokens
Full comparison
Organization
IBM
IBM
Family
Granite
Granite
License
Apache 2.0
Apache 2.0
Open weights
Yes
Yes
Release
Aug 25, 2026
Aug 25, 2026
Knowledge cutoff
-
-
API / provider
No primary API price listed
No primary API price listed
Modalities
text → text
text → text
Specs
Context window
131K
131K
Max output
8K
8K
Parameters
30B
8B
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
77.6
74.0
GPQA Diamond
64.4
63.1
Humanity's Last Exam
11.2
9.7
AIME 2025
89.2
86.7
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
10.1
11.2
AA-LCR v1.1
49.0
45.0
CritPt
0.3
0.3
MMMU-Pro
—
—
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
57.0
47.7
SWE-bench Pro
33.3
19.1
SWE-bench Multilingual
41.9
30.8
LiveCodeBench
75.8
73.2
Terminal-Bench 2.1
26.6
18.4
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
37.8
31.5
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | Granite 4.2 30B | Granite 4.2 8B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | IBM | IBM | — |
| Family | Granite | Granite | — |
| License | Apache 2.0 | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Aug 25, 2026 | Aug 25, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | No primary API price listed | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 131K | 131K | tie |
| Max output | 8K | 8K | tie |
| Parameters | 30B | 8B | A +22 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | 77.6 | 74.0 | A +3.6 pts |
| GPQA Diamond | 64.4 | 63.1 | A +1.3 pts |
| Humanity's Last Exam | 11.2 | 9.7 | A +1.5 pts |
| AIME 2025 | 89.2 | 86.7 | A +2.5 pts |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 10.1 | 11.2 | B +1.1 pts |
| AA-LCR v1.1 | 49.0 | 45.0 | A +4.0 pts |
| CritPt | 0.3 | 0.3 | tie |
| MMMU-Pro | — | — | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | 57.0 | 47.7 | A +9.3 pts |
| SWE-bench Pro | 33.3 | 19.1 | A +14.2 pts |
| SWE-bench Multilingual | 41.9 | 30.8 | A +11.1 pts |
| LiveCodeBench | 75.8 | 73.2 | A +2.5 pts |
| Terminal-Bench 2.1 | 26.6 | 18.4 | A +8.2 pts |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 37.8 | 31.5 | A +6.3 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
Granite 4.2 30B
- Context
- 131K / 8K out
- Parameters
- 30B
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- Apache 2.0
IBM's Aug 25, 2026 flagship dense reasoning model (Apache 2.0) — native 128K context with a 512K long-context extension, switchable thinking / non-thinking / low-effort modes, and agentic RL on software-engineering, terminal, and search tasks. IBM reports SWE-bench Verified 57.00 and Terminal-Bench 2.1 29.24 under its NeMo Evaluator setup.
Granite 4.2 8B
- Context
- 131K / 8K out
- Parameters
- 8B
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- Apache 2.0
Mid-size Granite 4.2 dense reasoning model (Apache 2.0, Aug 25, 2026) with native 128K context and the same thinking-mode switch as the 30B flagship. The 8B and 30B sizes received agentic RL; IBM reports SWE-bench Verified 47.67 and Terminal-Bench 2.1 20.56.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.