vs
1:1
overall capability wins (0 ties across 2 scored). Price/speed category wins do not decide overall. Tied overall.
- Specs1:1
Verdict
Agnes-3.0-Flash Preview offers larger context (262K vs 256K).
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Agnes-3.0-Flash Preview
- -
- Atria Dawn Preview
- -
Price per 1M tokens
Full comparison
Organization
Agnes AI
Shanghai AI Lab
Family
Other
Other
License
Apache 2.0
MIT
Open weights
Yes
Yes
Release
Sep 13, 2026
Sep 11, 2026
Knowledge cutoff
-
-
API / provider
No primary API price listed
No primary API price listed
Modalities
text, image, video → text
text → text
Specs
Context window
262K
256K
Max output
-
66K
Parameters
33B
744B
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
85.0
—
Humanity's Last Exam
—
—
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
23.0
—
AA-LCR v1.1
68.3
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
74.2
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
59.6
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
38.1
—
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | Agnes-3.0-Flash Preview | Atria Dawn Preview | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Agnes AI | Shanghai AI Lab | — |
| Family | Other | Other | — |
| License | Apache 2.0 | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Sep 13, 2026 | Sep 11, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | No primary API price listed | — |
| Modalities | text, image, video → text | text → text | — |
| Specs | |||
| Context window | 262K | 256K | A +6K |
| Max output | - | 66K | — |
| Parameters | 33B | 744B | B +711 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 85.0 | — | — |
| Humanity's Last Exam | — | — | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 23.0 | — | — |
| AA-LCR v1.1 | 68.3 | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | 74.2 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | 59.6 | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 38.1 | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
Agnes-3.0-Flash Preview
- Context
- 262K
- Parameters
- 33B
- Price
- —
- Speed
- —
- Modalities
- text, image, video → text
- License
- Apache 2.0
Agnes AI's (Sapiens AI, Singapore) first open-weight release: a 33B-parameter hybrid-attention model (54 gated delta-rule recurrent layers alternating with 18 global-attention layers, 3:1) under Apache 2.0, with a 262K-token context window and text/image/video input. This open-weight Preview checkpoint is distinct from Agnes AI's production "Agnes 3.0 Flash" API model, which runs a different checkpoint and is not open-weight; the two should not be conflated.
Atria Dawn Preview
- Context
- 256K / 66K out
- Parameters
- 744B
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- MIT
Shanghai AI Laboratory's open-weight agentic MoE (744B total, MIT license, 256K context), built on a GLM-5.2 base and trained by the cross-institutional ATRIA initiative (with Fudan NLP Lab) for long-horizon research and engineering workflows: problem analysis, tool use, code, experiments, and failure recovery. Benchmark scores are from the model's own technical report and have not yet been independently verified by a third-party evaluator.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.