PayPerQ flame logo
PayPerQ
Model status

Model status

Measured by PPQ · last 7 days (Nov 11 – Nov 18)

Availability is the share of requests that succeeded, hour by hour over the last 72 hours (higher is better). First response is how long until the model starts sending its answer (lower is better). Speed is how many tokens per second arrive from sending the request until the last token, the wait for the first token included (higher is better).

All systems normal

latest hour, all chat models
Chat availability last 72 h99.8%

100% chats ran inside our enclave router. PayPerQ can’t see the content of your queries and each reply also carries a receipt proving which provider and model we sent it to. How it works →

  • Availability
  • ≥ 99%
  • ≥ 95%
  • < 95% or failures
  • no data
  • this hour so far
  • Speed
  • < 30%≥ 100% of usual speed
  • no data
  • Claude Fable 5.2
    anthropic/claude-fable-5.2
    Availability last 72 h
    100%
    First response p50
    3.4 s
    About 1 in 10 wait over 24.8 s
    Speed p50
    93 tok/s
    Speed last 72 h
    About 1 in 10 slower than 73 tok/s
    Cache hit rate last 7 days
    81%

    $2.15/M input vs $8.44 list · 74% cost savings

  • Claude Opus 5.5
    anthropic/claude-opus-5.5
    Availability last 72 h
    99.9%
    First response p50
    3.1 s
    About 1 in 10 wait over 12 s
    Speed p50
    92 tok/s
    Speed last 72 h
    About 1 in 10 slower than 57 tok/s
    Cache hit rate last 7 days
    82%

    $1.22/M input vs $4.22 list · 71% cost savings

  • Claude Sonnet 5.5
    anthropic/claude-sonnet-5.5
    Availability last 72 h
    99.9%
    First response p50
    2.7 s
    About 1 in 10 wait over 14.2 s
    Speed p50
    118 tok/s
    Speed last 72 h
    About 1 in 10 slower than 82 tok/s
    Cache hit rate last 7 days
    89%

    $0.48/M input vs $2.11 list · 77% cost savings

  • DeepSeek V4.1 Flash 1017 (Ultra-fast)
    deepseek-v4.1-flash-1017-ultrafast
    Availability last 72 h
    100%
    First response p50
    0.9 s
    About 1 in 10 wait over 2.4 s
    Speed p50
    392 tok/s
    Speed last 72 h
    About 1 in 10 slower than 130 tok/s
    Cache hit rate last 7 days
    97%

    $0.13/M input vs $2.53 list · 94% cost savings

  • deepseek/deepseek-v4-pro-0813
    Availability last 72 h
    100%
    First response p50
    1.5 s
    About 1 in 10 wait over 3.1 s
    Speed p50
    93 tok/s
    Speed last 72 h
    About 1 in 10 slower than 48 tok/s
    Cache hit rate last 7 days
    85%

    $0.10/M input vs $0.12 list · 10% cost savings

  • deepseek/deepseek-v4.1-flash-1017
    Availability last 72 h
    99.7%
    First response p50
    1 s
    About 1 in 10 wait over 3 s
    Speed p50
    171 tok/s
    Speed last 72 h
    About 1 in 10 slower than 69 tok/s
    Cache hit rate last 7 days
    95%

    $0.021/M input vs $0.32 list · 93% cost savings

  • google/gemini-3.1-pro-preview
    Availability last 72 h
    99.9%
    First response p50
    3 s
    About 1 in 10 wait over 8.6 s
    Speed p50
    84 tok/s
    Speed last 72 h
    About 1 in 10 slower than 58 tok/s
    Cache hit rate last 7 days
    75%

    $0.68/M input vs $2.11 list · 67% cost savings

  • Google: Gemini 3.7 Flash
    google/gemini-3.7-flash
    Availability last 72 h
    99.9%
    First response p50
    2.4 s
    About 1 in 10 wait over 5 s
    Speed p50
    110 tok/s
    Speed last 72 h
    About 1 in 10 slower than 77 tok/s
    Cache hit rate last 7 days
    38%

    $0.52/M input vs $0.79 list · 34% cost savings

  • google/gemini-3.8-flash
    Availability last 72 h
    99.9%
    First response p50
    2.4 s
    About 1 in 10 wait over 4.8 s
    Speed p50
    93 tok/s
    Speed last 72 h
    About 1 in 10 slower than 55 tok/s
    Cache hit rate last 7 days
    64%

    $0.37/M input vs $0.87 list · 57% cost savings

  • Google: Gemini 4 Argon
    google/gemini-4-argon
    Availability last 72 h
    99.9%
    First response p50
    5.4 s
    About 1 in 10 wait over 19.6 s
    Speed p50
    43 tok/s
    Speed last 72 h
    About 1 in 10 slower than 25 tok/s
    Cache hit rate last 7 days
    76%

    $0.58/M input vs $2.11 list · 72% cost savings

  • google/gemini-4-carbon
    Availability last 72 h
    99.9%
    First response p50
    7.8 s
    About 1 in 10 wait over 33.9 s
    Speed p50
    43 tok/s
    Speed last 72 h
    About 1 in 10 slower than 17 tok/s
    Cache hit rate last 7 days
    78%

    $0.55/M input vs $2.11 list · 73% cost savings

  • MoonshotAI: Kimi K3
    moonshotai/kimi-k3
    Availability last 72 h
    99.9%
    First response p50
    2.1 s
    About 1 in 10 wait over 5.8 s
    Speed p50
    41 tok/s
    Speed last 72 h
    About 1 in 10 slower than 25 tok/s
    Cache hit rate last 7 days
    93%

    $0.50/M input vs $3.17 list · 84% cost savings

  • MoonshotAI: Kimi K3.1 Pro
    moonshotai/kimi-k3.1-pro
    Availability last 72 h
    99.9%
    First response p50
    1.8 s
    About 1 in 10 wait over 4.7 s
    Speed p50
    58 tok/s
    Speed last 72 h
    About 1 in 10 slower than 34 tok/s
    Cache hit rate last 7 days
    96%

    $0.29/M input vs $2.11 list · 86% cost savings

  • GPT-6 Astra
    openai/gpt-6-astra
    Availability last 72 h
    98.7%
    First response p50
    3.9 s
    About 1 in 10 wait over 14.8 s
    Speed p50
    34 tok/s
    Speed last 72 h
    About 1 in 10 slower than 20 tok/s
    Cache hit rate last 7 days
    87%

    $2.33/M input vs $10.55 list · 77% cost savings

  • GPT-6.1 Sol
    openai/gpt-6.1-sol
    Availability last 72 h
    99.7%
    First response p50
    2.8 s
    About 1 in 10 wait over 11.7 s
    Speed p50
    34 tok/s
    Speed last 72 h
    About 1 in 10 slower than 18 tok/s
    Cache hit rate last 7 days
    95%

    $0.22/M input vs $2.11 list · 89% cost savings

  • GLM-5.3 (Encrypted)
    private/glm-5-3
    Availability last 72 h
    99.9%
    First response p50
    1.7 s
    About 1 in 10 wait over 3.6 s
    Speed p50
    92 tok/s
    Speed last 72 h
    About 1 in 10 slower than 43 tok/s
    Cache hit rate last 7 days
    94%

    $0.56/M input vs $1.90 list · 70% cost savings

  • venice/abliteration-abliterated-model-large-flash
    Availability last 72 h
    100%
    First response p50
    3.2 s
    About 1 in 10 wait over 5.8 s
    Speed p50
    106 tok/s
    Speed last 72 h
    About 1 in 10 slower than 45 tok/s
    Cache hit rate last 7 days
    99%

    $0.15/M input vs $1.32 list · 88% cost savings

  • venice/abliteration-abliterated-model-large-v2
    Availability last 72 h
    99.9%
    First response p50
    2.2 s
    About 1 in 10 wait over 6.1 s
    Speed p50
    90 tok/s
    Speed last 72 h
    About 1 in 10 slower than 30 tok/s
    Cache hit rate last 7 days
    97%

    $0.40/M input vs $3.17 list · 87% cost savings

  • Z.AI: GLM 5.3
    z-ai/glm-5.3
    Availability last 72 h
    100%
    First response p50
    1.5 s
    About 1 in 10 wait over 5.2 s
    Speed p50
    92 tok/s
    Speed last 72 h
    About 1 in 10 slower than 43 tok/s
    Cache hit rate last 7 days
    97%

    $0.31/M input vs $1.48 list · 78% cost savings

  • z-ai/glm-5.3-prime
    Availability last 72 h
    100%
    First response p50
    2.3 s
    About 1 in 10 wait over 6.1 s
    Speed p50
    132 tok/s
    Speed last 72 h
    About 1 in 10 slower than 44 tok/s
    Cache hit rate last 7 days
    98%

    $0.44/M input vs $2.22 list · 79% cost savings

Measurements are best-effort: some requests are not timed. The ⓘ beside each model shows the share of successful streamed queries its times cover.

Methodology

First response
From the moment the request is sent upstream until the first content or reasoning token arrives.
Speed
Output tokens divided by the time from sending the request to the provider until the last token, time to first token included. Responses with very few output tokens are left out. The hourly bar compares each UTC hour’s median speed with the model’s median over the last 7 days, on a scale from red (much slower) through amber to green (usual speed or faster). An hour with too few timed answers is green without a number. The last block is the current hour so far, and the hour before it may still update.
p50 and “1 in 10”
p50 is the median: about half of the measured requests were faster. For first response, “1 in 10 wait over” is the 90th percentile (p90): about 1 in 10 measured requests took longer. For speed, “1 in 10 slower than” is the 10th percentile (p10): about 1 in 10 measured requests wrote more slowly. Percentiles are estimated, so the values are approximate. They describe the successful requests that were timed, which may not represent all traffic.
Cache hit rate
Input tokens served from the provider’s prompt cache divided by all input tokens, summed over the last 7 days, so a long prompt weighs more than a short one. Only successful queries where the provider reported cache usage count; the others are left out rather than counted as a miss. A share is shown only when there are enough such queries. A low share is not a fault: it depends on how people use the model and whether its providers cache.
Effective input price
What users paid per million input tokens on the model over the last 7 days, with cache reads and writes billed at the model’s own cache rates, compared with its list input price. The chart under the cache hit rate shows it in 12-hour windows; a window with too few requests is drawn between its neighbours. Your own price depends on how often your prompts repeat.
Availability
Successful requests divided by all requests, per UTC hour, counted in the hour each request was recorded. A failure counts against the model that was asked for, a success under the model that answered. Failures caused by the account — such as insufficient balance or an invalid request — are left out, and so are 4xx failures other than 429 (rate limit) and routing models that pick a model for you. Some hours show a colour without a percentage: green when no counted request failed, red when at least one did. The last block is the current hour so far, and the hour before it may still update. The headline is the same ratio over all 72 hours. This is availability measured on PPQ, not a provider’s own uptime.
Which models
The 20 chat models PPQ users were charged the most for in the last 7 days, among those with at least 50 queries, recomputed daily. Search above for any other chat model. A figure is shown only when there are enough measurements; otherwise you see a dash.