Coding LLM comparison

Claude Opus 5.5 leads this coding group on SciCode at 66.90%. Qwen3.8-Flash-Next has the lowest illustrated cost at $0.0025; DeepSeek leads configured output speed at 178.3 tokens/s.

Updated · 12 sources · Methodology

NeedPickWhy
Highest AA SciCode scoreClaude Opus 5.566.90%2
Lowest example costQwen3.8-Flash-Next$0.00257
Output speed (tokens/s)DeepSeek V4.1 Flash 1017178.3 tokens/s
AA time to first answer token (s)DeepSeek V4.1 Flash 101712.41 s

Model capabilities and output performance

ProviderSciCode (%)Output speed (tokens/s)AA time to first answer token (s)
Claude Opus 5.5 (Max, Default Fallback)266.9098.3687.99
Kimi K3.1 Pro160.7635.758.90
GLM-5.3 (Max)559.0374.229.16
DeepSeek V4.1 Flash 1017 (Max)355.90178.312.41
Qwen3.8-Flash-Next450.5855.438.17

Prices and request cost

ProviderInput / 1M tokensOutput / 1M tokensContext window (tokens)Max output (tokens)
Claude Opus 5.5 · Anthropic6$4$201,000,0006128,0006
Kimi K3.1 Pro · Novita AI9$2$101,048,5761,048,576
GLM-5.3 · Novita AI10$1.4$4.41,048,576131,072
DeepSeek V4.1 Flash 1017 · Novita AI8$0.3$1.21,048,576393,216
Qwen3.8-Flash-Next · Lyceum7$0.2$0.5262,14411131,07212
API documentationNovita AI · Chat API
Cost example: 10,000 input + 1,000 billed output tokens
Claude Opus 5.5$0.06
Kimi K3.1 Pro$0.03
GLM-5.3$0.0184
DeepSeek V4.1 Flash 1017$0.0042
Qwen3.8-Flash-Next$0.0025

Options, strengths and limitations

  1. Claude Opus 5.5

    Claude Opus 5.5 combines 66.90% SciCode with 98.3 tokens/s in this coding comparison.2

    Input / 1M tokens
    $4
    Output / 1M tokens
    $20
    Output speed (tokens/s)
    98.3

    Strengths

    • The highest SciCode result here is 66.90%.2

    Limitations

    • The $0.06 request example is the highest cost in this group.6

    Sources:26

  2. Kimi K3.1 Pro

    For this coding shortlist, Kimi K3.1 Pro records 60.76% SciCode and a separate 35.7 tokens/s output rate.1

    Input / 1M tokens
    $2
    Output / 1M tokens
    $10
    Output speed (tokens/s)
    35.7

    Strengths

    • SciCode is 60.76%, above DeepSeek’s 55.90%.13

    Limitations

    • Its 58.90 s first-answer wait exceeds DeepSeek’s 12.41 s.13

    Sources:19

  3. GLM-5.3

    GLM-5.3: 59.03% SciCode; its separate AA output speed is 74.2 tokens/s for the coding comparison.5

    Input / 1M tokens
    $1.4
    Output / 1M tokens
    $4.4
    Output speed (tokens/s)
    74.2

    Strengths

    • SciCode is 59.03%, above DeepSeek’s 55.90%.53

    Limitations

    • The $0.0184 example costs more than DeepSeek’s $0.0042.108

    Sources:510

  4. DeepSeek V4.1 Flash 1017

    The coding comparison includes DeepSeek V4.1 Flash 1017 at 55.90% SciCode and 178.3 tokens/s output.3

    Input / 1M tokens
    $0.3
    Output / 1M tokens
    $1.2
    Output speed (tokens/s)
    178.3

    Strengths

    • Its first-answer wait of 12.41 s is the shortest in this group.3

    Limitations

    • SciCode 55.90% trails Claude Opus 5.5 at 66.90%.32

    Sources:38

  5. Qwen3.8-Flash-Next

    In the coding shortlist, Qwen3.8-Flash-Next has 50.58% SciCode and 55.4 tokens/s in the separate performance test.4

    Input / 1M tokens
    $0.2
    Output / 1M tokens
    $0.5
    Output speed (tokens/s)
    55.4

    Strengths

    • The $0.0025 token example costs less than Claude’s $0.06.76

    Limitations

    • SciCode 50.58% trails Claude Opus 5.5 at 66.90%.42

    Sources:47

Methodology and sources

SciCode is AA’s scientific-coding evaluation, reported as a percentage; higher is better. It is separate from the 10,000-input-token performance test and is not an in-editor Cline evaluation. Quality and performance use the configurations shown beside the model names. A current AA v5 Intelligence Index is displayed only where both models have a published value. AA performance uses the 10,000-input-token benchmark at each model’s displayed configuration; reasoning settings vary by model. Output speed is the median output rate. Answer latency measures the first answer token after reasoning, not the initial streamed reasoning chunk. For non-streamed reasoning and non-reasoning models, AA reports the same first-chunk and first-answer event. AA estimates answer latency from initial response and reasoning duration; it is not a separately sampled median. These are reference deployments, not Novita endpoint guarantees. The uncached cost example bills 10,000 input and 1,000 output tokens at each named seller’s tariff; it is not the cost of the AA benchmark. Prices, context and output limits use the cited seller sources.

  1. Artificial Analysis · Kimi K3.1 Pro · quality and performance
  2. Artificial Analysis · Claude Opus 5.5 (Max, Default Fallback) · quality and performance
  3. Artificial Analysis · DeepSeek V4.1 Flash 1017 (Max) · quality and performance
  4. Artificial Analysis · Qwen3.8-Flash-Next · quality and performance
  5. Artificial Analysis · GLM-5.3 (Max) · quality and performance
  6. Anthropic · Claude Opus 5.5 · API pricing and limits
  7. Lyceum · Qwen3.8-Flash-Next · API pricing and limits
  8. Novita AI · DeepSeek V4.1 Flash 1017 · endpoint pricing and limits
  9. Novita AI · Kimi K3.1 Pro · endpoint pricing and limits
  10. Novita AI · GLM-5.3 · endpoint pricing and limits
  11. Lyceum · Qwen3.8-Flash-Next · native context specification
  12. Lyceum · serverless request limits

Build with Novita Model APIs

Create an API key and connect your application.

Start BuildingExplore Models