Compare

Compare models

Up to four models side by side — specs, capabilities, every covered benchmark, provenance and field context. Best value per row is highlighted.

Model A

Model B

Model C

Model D

Saved →

GPT-5.2CLOSED

2426.4#46 · OpenAI

DeepSeek-V3.1 (Thinking)OPEN

1848.3#100 · DeepSeek

empty slot

Overview

Frontier Elo2426.41848.3

Overall rank#46#100

Benchmarks tested11 / 1709 / 170

Categories covered4/75/7

Benchmark wins40

Elo per $ out1731100

Specifications

OrganizationOpenAIDeepSeek

FamilyGPT-5.2DeepSeek-V3

ReleasedDec 2025 · 9mo agoAug 2025 · 13mo ago

Statusreleaseddeprecated · Thinking · variant

Succeeds—DeepSeek-V3-0324

WeightsClosedOpen weights

LicenseProprietary · ProprietaryMIT · Permissive

ArchitectureUndisclosed architecture (presumed dense transformer, reasoning-tuned); shipped as Instant/Thinking/Pro tiersMixture-of-Experts (671B total / 37B active) — DeepSeekMoE + Multi-head Latent Attention, hybrid think/non-think chat template

Parametersundisclosed671B·37Ba

Context400K128K

Modalitiestext · visiontext

Languages——

Hosted APIYesYes

Price & deployment

Price in /M$1.75$0.56

Price out /M$14$1.68

Blended 3:1 /M$4.81$0.84

VRAM @ Q4—405 GB

VRAM @ FP16—1340 GB

Smallest GPU (Q4)API onlyNo single GPU

~tok/s RTX 4090——

Quantizations—GGUF · FP8

LinksDOCSHF GH DOCS

Capabilities

Reasoning

Coding

Vision

Function calling

Tool use

Agentic

Benchmarks

Knowledge2—93.9——field best

MMLU-Pro—84.8%self91.5%

MMLU-Redux—93.7%self95.0%

Reasoning473.850.3——field best

ARC-AGI-186.2%self—98.5%

ARC-AGI-252.9%self—95.0%

GPQA Diamond92.4%self80.1%−12.3%self96.0%

Humanity's Last Exam34.5%self15.9%−18.6%self67.7%

Coding580.873.3——field best

Aider Polyglot—76.3%self89.4%

Codeforces—2091self3471

LiveCodeBench88.9%74.8%−14.1%self93.5%

SWE-bench Pro55.6%self—89.9%

SWE-bench Verified80.0%self—97.0%

Math271.493.1——field best

AIME100.0%self93.1%−6.9%self100.0%

FrontierMath (Tiers 1-3)40.3%self—89.0%

Agents389.231.3——field best

BrowseComp—30.0%self91.2%

GDPval70.9%self—87.9%

τ²-Bench Telecom98.7%self—99.3%

Notes

Curation noteAdds GPT-5.2 Pro tier and new xhigh reasoning effort; SOTA at launch on GDPval, ARC-AGI-2, and long-context MRCRv2.Thinking mode of the hybrid V3.1 checkpoint; served via deepseek-reasoner, faster than R1-0528 at similar quality

Head to head

The short answer across rating, categories, shared benchmarks and cost

Rating leaderGPT-5.2 · +578.1 Elo

Category leadsGPT-5.2 3 · DeepSeek-V3.1 (Thinking) 1 (of 4)

Shared benches (4)GPT-5.2 4

CheapestDeepSeek-V3.1 (Thinking) · 5.7× cheaper

NewestGPT-5.2 · Dec 2025 · 3mo newer

Runs locallyDeepSeek-V3.1 (Thinking)

Best valueDeepSeek-V3.1 (Thinking) · 1100 Elo/$

Capability profile

Normalized against the tracked field · only categories a selected model was tested on

PREF hidden — untested by all selected models, so not charted.

GPT-5.24/62426.4 DeepSeek-V3.1 (Thinking)5/61848.3

Field context

Frontier Elo vs. output price across the priced field — compared models are labeled. Click a point: fills the first open slot, or opens the model's page when the table is full (or it's already in).

Scores carry per-row provenance (self-reported marks are flagged) — see methodology · data as of October 10, 2026.