SPACEXAI
4/5 workers samplingPOSSIBLE_MODEL_CHANGE · gpt-5
Observatory

SpaceXai Global Bench Universe

Independent measurement of intelligence, real task completion, serving speed, reliability, cost, and system efficiency — under one evidence framework. Not a press-release table.

Open ArenaMethodology
Killer metric · effective work rate

Grok 4 on xAI completed 78% of the GBU mix, at $0.09/success, 11.4 seconds median, 3.2 tool calls, 0.4 retries, 99.94% endpoint availability. Effective work rate 0.92 tasks per dollar-second.

A model that scores 2% higher but costs 8× more and takes 5× longer does not automatically look better. EWR = correct completed tasks / (dollar · second).

Grok 4 EWR
0.925
$0.095/correct
GPT-5 EWR
0.268
$0.262/correct
Flash EWR
20.2
Cheap × fast × good-enough
Opus EWR
0.104
Best agents, expensive finish
01 Capability

What it can do

Math, reasoning, coding, agents, science, long context, private vault. Scores stay attached to a run.

02 Performance

How it serves

TTFT, tokens/sec, p50/p95/p99, bursts, cold/warm. Five regions, nine scenarios, continuous sampling.

03 Economics

What it costs to finish

$/1M tokens is a vendor line. $/correct task is the number that should move a purchase.

04 Reliability

Whether it holds

Availability, refusals, broken tools, invalid JSON, replay variance, silent regressions.

Ranked by effective work rate

Full board
EndpointEWREvidence
Llama 4 70B
Groq LPU
64.3LIVE_MEASURED
Llama 4 70B
Together
32.3LIVE_MEASURED
Gemini 2.5 Flash
Google API
20.2LIVE_MEASURED
Grok 3 Mini
xAI API
16.6LIVE_MEASURED
Llama 4 405B
Self-host B200
13.5REPRODUCED
GPT-5 Mini
OpenAI API
8.50LIVE_MEASURED
Llama 4 405B
Groq LPU
8.05LIVE_MEASURED
Llama 4 405B
Self-host H100
4.76REPRODUCED

Live speed samples

streaming
  • Claude 4.1 Sonnet
    asia · streaming · HTTP 200
    535 ms
    96 tok/s
  • Grok 4
    us-east · large-prompt · HTTP 200
    616 ms
    114 tok/s
  • Llama 4 70B
    asia · large-prompt · HTTP 200
    120 ms
    700 tok/s
  • Llama 4 70B
    us-east · long-context · HTTP 200
    103 ms
    636 tok/s
  • Llama 4 405B
    us-east · long-context · HTTP 200
    366 ms
    56 tok/s
  • GPT-5
    us-west · tool-call · HTTP 200
    704 ms
    78 tok/s
  • Gemini 2.5 Pro
    us-east · concurrent-burst · HTTP 200
    718 ms
    89 tok/s
Speed observatory

Silent change watch

  • GPT-5

    Same model string. Canary drift after an undeclared August deployment. Coding −2.8, science −4.0, refusals up.

    POSSIBLE_MODEL_CHANGE
  • Grok 4

    Within daily variance band. No adapter or tokenizer shift.

    NOISE
  • Gemini 2.5 Pro

    Long-context improved; tool-call latency dropped. Possible silent serving change.

    POSSIBLE_MODEL_CHANGE
  • Claude 4.1 Opus

    Provider changelog matched canary. Computer-use stack bumped; TTFT slightly worse.

    CONFIRMED

Immutable runs

Frontier models
15
Live endpoints
21
Regional workers
5
Last Grok 4 sample
11 min ago
Grok 4 · TTFT 24h sparkp50, US-East