GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

DeepSeek R1-0528: AIME 2025 Jumps From 70% to 87.5% in Open-Source Reasoning Upgrade

DeepSeek pushed a checkpoint update to its R1 reasoning model on May 28, date-stamped R1-0528. The headline number: AIME 2025 improved from 70% to 87.5% — a 17.5-point gain on the mathematical reasoning benchmark that has become the standard test for reasoning model progress.

The update is available on Hugging Face and via API at $0.50/M input, $2.15/M output. 671B total parameters, 37B active per inference pass (MoE architecture). Context: 164K tokens. Open-source.

What Changed

The model card identifies two mechanisms behind the improvement:

  1. More compute at post-training — the update spends significantly more RL compute after pretraining, sharpening the model’s internal reward optimization without changing the base weights architecture
  2. Algorithmic optimization — targeted improvements to how R1 generates and verifies intermediate reasoning steps, aimed at math, coding, and formal logic specifically

A secondary improvement: instruction following is more reliable, and the verbose chain-of-thought traces that made early R1 awkward to deploy in production have been reduced. The model spends more reasoning budget reaching a better answer rather than displaying intermediate work at length.

The Trade-Off

The accuracy gain comes at a token cost. R1-0528 uses more tokens per question on average — a direct consequence of spending more reasoning budget to get from premise to conclusion. For high-stakes tasks (formal verification, competitive math, structured code generation) that trade-off is generally worth it. For high-volume applications, V4 Flash at $0.14/M output is still the cost-optimal path.

Competitive Position

At 87.5% AIME 2025, R1-0528 pulls level with GPT-5.4 on this specific benchmark and ahead of the original R1 by a significant margin. The comparison that matters: GPT-5.4 is $3/M input, $15/M output. R1-0528 is $0.50 input, $2.15 output — a 6x blended cost advantage.

For developers building reasoning-heavy pipelines who need o1-class math performance at a fraction of the cost, this update changes the routing calculus. The open weights on Hugging Face also mean self-hosted deployment on H100 hardware is viable for organizations with inference infrastructure.

Benchmark Context

ModelAIME 2025Cost (output/M)
R1-052887.5%$2.15
R1 (original)70.0%$2.15
GPT-5.4~87%$15.00
DeepSeek V4 Pro—$0.87

The update does not change R1’s position on SWE-bench or agentic benchmarks — those remain V4 Pro’s territory. R1-0528 is a reasoning model improvement, not a general-purpose model refresh.

DeepSeek has not announced a next major generation (R2). R1-0528 is iterative maintenance on the existing checkpoint — a sign the lab is managing a live model rather than staging a launch campaign.