GPT-4 0125 Preview vs GPT-6 Astra
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
GPT-4 0125 Preview—
GPT-6 Astra52.7
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
GPT-4 0125 Preview—
GPT-6 Astra49.5
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
GPT-4 0125 Preview—
GPT-6 Astra64.6
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
GPT-4 0125 Preview—
GPT-6 Astra · max1,450
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
GPT-4 0125 Preview—
GPT-6 Astra85.1
| Benchmark | GPT-4 0125 Preview | GPT-6 Astra |
|---|---|---|
| Humanity's Last Exam | — | 52.7 |
| Terminal-Bench 4.0 | — | 49.5 |
| AutomationBench-AA | — | 64.6 |
| Arena Text · creative writing | — | 1,450 · max |
| MMMU-Pro (vision) | — | 85.1 |
Specs
| Spec | GPT-4 0125 Preview | GPT-6 Astra |
|---|---|---|
| Input, per 1M | — | $10 |
| Output, per 1M | — | $50 |
| Cost per task? | — | $1.54 |
| Speed, tok/s | — | 59 |
| Context | — | 1.05M |
| Weights | Not verified | Closed |
| Released | — | Sep 2026 |
| See images | ? | ✓ |
| Read PDFs | ? | Partial |
| Hear audio | ? | — |
| Watch video | ? | — |
| Call tools | ✓ | ✓ |
| Force a tool call | ? | Partial |
| Return strict JSON | ? | ✓ |
| Search the web | ? | ✓ |
| Run code | ? | ✓ |
| Use a computer | ? | ✓ |
What breaks when you switch
GPT-4 0125 Preview → GPT-6 Astra
Monitoring can return HTTP 403›
Misalignment monitoring can return HTTP 403 before streaming or errors after output; match misalignment_policy_violation and handle stream errors.
No none reasoning effort›
The none reasoning effort is unsupported on GPT-6 Astra and GPT-6.1 Sol; use low instead.
Remove unsupported sampling parameters›
When reasoning effort is not none, remove temperature, top_p and top_logprobs, plus Chat Completions logprobs or Responses output logprobs inclusion.
GPT-6 Astra → GPT-4 0125 Preview