Skip to main content
BenchLM
Advertise

Model comparison

DeepSeek V3.2 (Thinking) vs GPT-5.4 mini

Updated July 30, 2026. Public scores include evidence status and uncertainty. They are not guarantees for a specific workload.

DeepSeek V3.2 (Thinking)

DeepSeek

57.3/100

Estimated · Public rank #72

90% interval 45.8–68.8

GPT-5.4 mini

OpenAI

55.8/100

Estimated · Public rank #81

90% interval 44.3–67.3

DeepSeek V3.2 (Thinking) has the higher public score estimate, 57.32 versus 55.79, but the 90% score intervals overlap. Treat that as a lead, not a settled winner.

1 results are shared. Category rows based on different benchmark sets are marked directional and do not name a winner.

Which one for your work

Recommendations appear only when a shared evidence basis or an explicit operating constraint supports the call. Secondary and unsupported use cases stay disclosed below the initial list.

  • Long documents

    Prompts that approach the documented context limit

    GPT-5.4 mini

    GPT-5.4 mini has the larger documented context window.

    Confidence: documented

  • Chat turn cost

    1K fresh input + 500 output tokens

    DeepSeek V3.2 (Thinking)

    DeepSeek V3.2 (Thinking) has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates

Show secondary and unsupported calls

  • Repository review cost

    50K fresh input + 3K output tokens

    DeepSeek V3.2 (Thinking)

    DeepSeek V3.2 (Thinking) has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates

  • Coding work

    Code generation, repair, and software-engineering tasks

    Not enough matched evidence

    No shared weighted benchmark basis supports a winner.

    Confidence: limited

  • Agentic work

    Tool use, computer use, and multi-step task completion

    Not enough matched evidence

    No shared weighted benchmark basis supports a winner.

    Confidence: limited

  • Cache-heavy agent loop cost

    200K cached + 20K fresh input + 10K output tokens

    Not enough matched evidence

    The page does not recommend a cost winner because at least one model cannot fit the stated workload in one request. DeepSeek V3.2 (Thinking) does not fit this workload in one request.

    Confidence: listed-rates

What is actually comparable

Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.

Shared results
1
DeepSeek V3.2 (Thinking) only
0
GPT-5.4 mini only
13
Like-for-like categories
0 / 8

Category results, on a stated basis

Each row states whether both averages use the same weighted benchmark set. Directional and not-comparable rows remain visible, but they never receive a winner in this template.

Agentic

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
65.7
Weighted basis
0 vs 2 rows
Reading
Not comparable

Coding

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
Not measured
Weighted basis
0 vs 0 rows
Reading
Not comparable

Reasoning

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
Not measured
Weighted basis
0 vs 0 rows
Reading
Not comparable

Knowledge

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
47.8
Weighted basis
0 vs 2 rows
Reading
Not comparable

Math

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
21.7
Weighted basis
0 vs 2 rows
Reading
Not comparable

Multilingual

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
Not measured
Weighted basis
0 vs 0 rows
Reading
Not comparable

Multimodal

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
76.6
Weighted basis
0 vs 1 rows
Reading
Not comparable

Instruction following

Not comparable
DeepSeek V3.2 (Thinking)
Not measured
GPT-5.4 mini
Not measured
Weighted basis
0 vs 0 rows
Reading
Not comparable

Shape of the matched evidence

Only shared public evidence is shown. Sparse evidence stays a ruled list rather than being closed into a radar shape.

A shared-evidence shape is not available.

BenchLM does not draw a radar or infer missing axes when the matched evidence is too sparse.

What each workload costs

Three fixed token mixes turn per-token rates into comparable decisions. Each scenario states context fit and whether cached input had to fall back to the published list-input rate.

Chat turn

1K fresh input + 500 output tokens

DeepSeek V3.2 (Thinking)
$0.00165
Fits in one request
GPT-5.4 mini
$0.003
Fits in one request

DeepSeek V3.2 (Thinking) has the lower modeled cost

Costs use the listed standard API rates.

Repository review

50K fresh input + 3K output tokens

DeepSeek V3.2 (Thinking)
$0.03407
Fits in one request
GPT-5.4 mini
$0.051
Fits in one request

DeepSeek V3.2 (Thinking) has the lower modeled cost

Costs use the listed standard API rates.

Cache-heavy agent loop

200K cached + 20K fresh input + 10K output tokens

DeepSeek V3.2 (Thinking)
$0.0609
Does not fit in one request
GPT-5.4 mini
$0.075
Fits in one request

DeepSeek V3.2 (Thinking) does not fit this workload in one request.

Specification differences

Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.

Cached-input rate

A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.

DeepSeek V3.2 (Thinking)

$0.14 per 1M cached input tokens

GPT-5.4 mini

$0.075 per 1M cached input tokens

OpenAI pricing

Provider availability

DeepSeek V3.2 (Thinking)

Not sourced

GPT-5.4 mini

Generally Available · OpenAI Responses API

OpenAI model catalog

Reasoning profile

DeepSeek V3.2 (Thinking)

Reasoning

GPT-5.4 mini

Reasoning

Weight access

DeepSeek V3.2 (Thinking)

Open Weight

GPT-5.4 mini

Proprietary

License

DeepSeek V3.2 (Thinking)

Open Weight

GPT-5.4 mini

Proprietary

Release date

DeepSeek V3.2 (Thinking)

2025-12-01

GPT-5.4 mini

2026-03-17

If you already use one of these models
Deployment change
The models list different providers, so authentication, endpoint behavior, limits, and feature support may change.
Quality signal
DeepSeek V3.2 (Thinking) has the higher public score estimate, 57.32 versus 55.79, but the 90% score intervals overlap.
Workload cost
Repository review: $0.03407 vs $0.051. Cache-heavy agent loop: $0.0609 vs $0.075.
Context tradeoff
GPT-5.4 mini has the larger documented window (400K).

Run the same representative tasks against both endpoints before changing production traffic.

Benchmark evidence

The full public result ledger is available for audit without forcing a wide desktop table onto a phone.

Browse raw public benchmark evidence14 rows

Agentic

  • Terminal-Bench 2.0

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini60%
    Source

    Not directly comparable

  • OSWorld-Verified

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini72.1%
    Source

    Not directly comparable

  • MCP Atlas

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini57.7%
    Source

    Not directly comparable

  • Toolathlon

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini42.9%
    Source

    Not directly comparable

  • τ²-bench results

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini93.4%
    Source

    Not directly comparable

Coding

  • Vibe Code Bench

    Shared source
    DeepSeek V3.2 (Thinking)5.11%
    GPT-5.4 mini47.97%

    GPT-5.4 mini leads this result

  • FrontierCode 1.1 Main

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini27.0%
    Source

    Not directly comparable

Knowledge

  • GPQA

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini88%
    Source

    Not directly comparable

  • HLE

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini41.5%
    Source

    Not directly comparable

  • HLE w/o tools

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini28.2%
    Source

    Not directly comparable

Math

  • FrontierMath v2 (Tiers 1-3)

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini28.280%
    Source

    Not directly comparable

  • FrontierMath v2 (Tier 4)

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini2.080%
    Source

    Not directly comparable

Multimodal

  • MMMU-Pro

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini76.6%
    Source

    Not directly comparable

  • MMMU-Pro w/ Python

    DeepSeek V3.2 (Thinking)—
    GPT-5.4 mini78%
    Source

    Not directly comparable

Frequently asked questions

Which is better, DeepSeek V3.2 (Thinking) or GPT-5.4 mini?

DeepSeek V3.2 (Thinking) has the higher public score estimate, 57.32 versus 55.79, but the 90% score intervals overlap. The higher estimate is not a decisive winner because the uncertainty ranges overlap.

Which is better for coding, DeepSeek V3.2 (Thinking) or GPT-5.4 mini?

The published evidence does not provide a shared weighted coding basis for both models, so BenchLM does not name a coding winner.

Which is better for agentic tasks, DeepSeek V3.2 (Thinking) or GPT-5.4 mini?

The published evidence does not provide a shared weighted agentic tasks basis for both models, so BenchLM does not name a agentic tasks winner.

Which costs less, DeepSeek V3.2 (Thinking) or GPT-5.4 mini?

For the stated presets, chat costs $0.00164 on DeepSeek V3.2 (Thinking) and $0.003 on GPT-5.4 mini; repository review costs $0.03407 and $0.051; the cache-heavy agent loop costs $0.0609 and $0.075. DeepSeek V3.2 (Thinking) does not fit this workload in one request.

Which has the larger context window, DeepSeek V3.2 (Thinking) or GPT-5.4 mini?

GPT-5.4 mini has the larger documented context window: 400K, compared with 128K.

Related comparisons

Compare API pricing Read the methodology Open the model selector

Last updated July 30, 2026

Watch DeepSeek V3.2 (Thinking) vs GPT-5.4 mini

One weekly email when material rank, price, or benchmark evidence changes make this matchup worth revisiting.

Read a sample issue

Join 2,000+ readers.