ZMIME
Comparison · 3 models · Updated Nov 18, 2026

GPT-5-Codex vs Kimi K2 Instruct vs Qwen3 235B-A22B Thinking 2507

No independent capability benchmark covers every model here yet. The score uses inputs & features and context window.

Price in rating: Excluded · Include price

  1. OpenAI

    GPT-5-Codex

    Released Sep 15, 2025

    56/100
    • ECI—
    • Price$1.25 / $10.00
    • Context400K
  2. Moonshot AI

    Kimi K2 Instruct

    Released Jul 11, 2025

    28/100
    • ECI141.3
    • Price$0.57 / $2.30
    • Context131K
  3. Alibaba (Qwen)

    Qwen3 235B-A22B Thinking 2507

    Released Jul 25, 2025

    38/100
    • ECI143.3
    • Price$0.30 / $2.90
    • Context262K
01 — Verdict

Capability evidence is incomplete

There is not enough shared capability evidence to name an overall winner. The scores (GPT-5-Codex 56/100, Qwen3 235B-A22B Thinking 2507 38/100, Kimi K2 Instruct 28/100) compare the available specifications only. The score weighs inputs & features 60%, context window 40%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.

  • CapabilityNot enough dataNo independent benchmark covers every model here yet
  • Lowest priceQwen3 235B-A22B Thinking 2507Qwen3 235B-A22B Thinking 2507 $0.95 · Kimi K2 Instruct $1.00 · GPT-5-Codex $3.44 per 1M tokens (3:1 blend)
  • Longest contextGPT-5-CodexGPT-5-Codex 400,000 · Qwen3 235B-A22B Thinking 2507 262,144 · Kimi K2 Instruct 131,072 tokens
  • Widest inputsGPT-5-CodexGPT-5-Codex: Text, Images · Kimi K2 Instruct: Text · Qwen3 235B-A22B Thinking 2507: Text
  • Self-hostingKimi K2 Instruct and Qwen3 235B-A22B Thinking 2507Publishes downloadable weights (Modified MIT and Apache 2.0)
How the score is built
MeasureWeightGPT-5-CodexKimi K2 InstructQwen3 235B-A22B Thinking 2507
Inputs & features60%602535
Context window40%503243
Specs only100%56/10028/10038/100

Left out because at least one model lacks the data: capability. The remaining weights were rescaled.

02 — Side by side

Every spec in one table

Highlighted cells lead their row. Dashes mean the data is not published.

GPT-5-Codex vs Kimi K2 Instruct vs Qwen3 235B-A22B Thinking 2507 specifications side by side
SpecificationGPT-5-CodexOpenAIKimi K2 InstructMoonshot AIQwen3 235B-A22B Thinking 2507Alibaba (Qwen)
Capability
Capabilities Index (ECI)—141.3143.3 (best)
ECI rank—#108 of 158#100 of 158 (best)
GPQA DiamondGraduate-level science questions——80.1%
OTIS Mock AIME 2024–2025Competition mathematics——86.7%
SimpleQA VerifiedShort factual questions——40.4%
Price per million tokens
Input$1.25$0.57$0.30 (best)
Output$10.00$2.30 (best)$2.90
Cached input———
Blended (3:1)$3.44$1.00$0.95 (best)
Long-context rateSame rateSame rateSame rate
Price sourceMedian of 9 providersMedian of 11 providersMedian of 11 providers
Limits
Context window400,000 tokens (best)131,072 tokens262,144 tokens
Max output128,000 tokens (best)32,768 tokens81,920 tokens
Inputs and features
TextYesYesYes
ImagesYesNoNo
PDFsNoNoNo
AudioNoNoNo
VideoNoNoNo
ReasoningYesNoYes
Tool callingYesYesYes
Structured outputNoNoNo
Availability
WeightsProprietaryOpenModified MITOpenApache 2.0
API model ID———
API providers101112 (best)
ReleasedSep 15, 2025Jul 11, 2025Jul 25, 2025
Knowledge cutoffSep 30, 2024——
03 — Cost

What would a month cost?

Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.

  • GPT-5-Codex$32.50
  • Kimi K2 Instruct$10.30
  • Qwen3 235B-A22B Thinking 2507$8.80
04 — Questions

Which should you choose?

Which is better: GPT-5-Codex, Kimi K2 Instruct or Qwen3 235B-A22B Thinking 2507?

There is not enough shared capability evidence to name an overall winner. The scores (GPT-5-Codex 56/100, Qwen3 235B-A22B Thinking 2507 38/100, Kimi K2 Instruct 28/100) compare the available specifications only. The score weighs inputs & features 60%, context window 40%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.

Which is cheaper, GPT-5-Codex, Kimi K2 Instruct or Qwen3 235B-A22B Thinking 2507?

Qwen3 235B-A22B Thinking 2507 is cheaper at $0.30 input / $2.90 output per million tokens (median across 11 API providers). Kimi K2 Instruct costs $0.57 input / $2.30 output per million tokens (median across 11 API providers); GPT-5-Codex costs $1.25 input / $10.00 output per million tokens (median across 9 API providers). At a typical mix of three input tokens to one output token, that is $0.95 per million tokens for Qwen3 235B-A22B Thinking 2507 versus $1.00 for Kimi K2 Instruct (1.1× as much) and $3.44 for GPT-5-Codex (3.6× as much).

Which scores higher on benchmarks?

There is no independent benchmark that covers all three models yet. GPT-5-Codex has not been scored yet, Kimi K2 Instruct has an ECI of 141.3 and Qwen3 235B-A22B Thinking 2507 has an ECI of 143.3.

Which is better for coding?

There are no published SWE-bench Verified results for GPT-5-Codex, Kimi K2 Instruct and Qwen3 235B-A22B Thinking 2507 yet, so there is no like-for-like coding score. Test both on a sample of your own tasks. All three support tool calling for agent workflows.

Which has the bigger context window?

GPT-5-Codex has the largest context window at 400,000 tokens, against 262,144 for Qwen3 235B-A22B Thinking 2507 and 131,072 for Kimi K2 Instruct. Maximum output per response: GPT-5-Codex up to 128,000, Kimi K2 Instruct up to 32,768, Qwen3 235B-A22B Thinking 2507 up to 81,920 tokens.

Which can read images, PDFs, audio or video?

GPT-5-Codex accepts text and images; Kimi K2 Instruct accepts text; Qwen3 235B-A22B Thinking 2507 accepts text. GPT-5-Codex handles the widest range of inputs.

Are any of these open source?

Kimi K2 Instruct and Qwen3 235B-A22B Thinking 2507 publishes its weights (Modified MIT and Apache 2.0) and can be self-hosted; GPT-5-Codex is proprietary.

Which is newer?

GPT-5-Codex is the newest, released Sep 15, 2025. Qwen3 235B-A22B Thinking 2507 came out Jul 25, 2025; Kimi K2 Instruct came out Jul 11, 2025. Knowledge cutoff: GPT-5-Codex Sep 30, 2024.

How do you decide the winner?

This rating uses inputs & features (60%, image, PDF, audio and video input, tool calling, structured output and reasoning); context window (40%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 18, 2026.