Wait Which Model?
← Back to directory
Amazon (Nova)

Nova Premier

Released Apr 30, 2025 · knowledge cutoff 2024-10

Status
Superseded
Location
United States
Modality
Multimodal
Context window
1M
Max output
25K
Speed
33 tok/s · 2.85s to first answer token
Price ($/MTok in / out)
$2.50 / $12.50
Cost per task
—
Open weights
No
Benchmarks4 of 8 reported

MMLU-Pro

73.3%

GPQA Diamond

56.9%

SWE-bench Verified

42.4%

AIME (latest)

17.3%

Terminal-Bench 2.1

—

Humanity's Last Exam

—

LMArena Elo

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Designed as a teacher model for Bedrock Model Distillation, so its main value for many customers is training cheaper Nova Lite/Micro students rather than being called directly
  • Strong at knowledge retrieval and visual understanding (SimpleQA, MMMU) and processes a genuinely usable 1M-token window across text, image and video in one request
  • Built-in Bedrock Guardrails and content moderation out of the box, which enterprise adopters flagged as a real time-saver over stitching in a third-party safety layer

Weaknesses

  • Not a reasoning model — can't fact-check or deliberate step-by-step, so it falls behind reasoning-enabled peers on math and multi-step problems
  • Trails Gemini 2.5 Pro on SWE-bench Verified and lags on GPQA Diamond/AIME despite the 'most capable Nova' positioning at launch
  • One of the slower non-reasoning models Artificial Analysis measured at launch (32.5 output tok/s, TTFT 2.85s), both below the median for its class

AWS's original launch announcement and TechCrunch coverage both date general availability to 2025-04-30; the official Bedrock model-card page lists 'Model launch date: Oct 31, 2025' instead, which may reflect a later model-card/reasoning update under the same amazon.nova-premier-v1:0 id (the card also lists 'Reasoning: Supported', at odds with contemporaneous reviews calling launch-day Premier a non-reasoning model) — used the earlier, better-corroborated April date as releaseDate and flagged the discrepancy here rather than guess which it refers to. Benchmark figures (MMLU-Pro 73.3, GPQA-Diamond 56.9, AIME 2025 17.3, SWE-Bench Verified 42.4) are Amazon's own numbers, republished as the 'Nova 1 Premier' comparison row in the later Nova 2 technical report (assets.amazon.science, Table in section 2.1/2.4). Terminal-Bench figure in the same table (11.3%) is version 1.0, not the tracked 2.1 metric, so left null per house style. HLE, ARC-AGI-2 and lmarenaElo were not found on any official or independent leaderboard. Speed and pricing are Artificial Analysis' measurements (artificialanalysis.ai/models/nova-premier); AWS's own pricing page does not publish per-token figures directly, only routing to Bedrock's console pricing. retirementDate (2026-09-14) and knowledgeCutoff (Oct 2024) are from the official Bedrock model-card page, which also lists the model's lifecycle as 'Legacy'. Closed weights; no license record.

For developers

API model strings

  • aws-bedrockamazon.nova-premier-v1:0

Licence

Proprietary — weights not released

Retirement

Sep 14, 2026

Lineage

No recorded predecessor

News