Model
Model explorer

GPT-6 Astra Mini

CLOSED
OpenAI · GPT-6 family · released Sep 24, 2026

OpenAI's September 24, 2026 volume tier for the GPT-6 line — Astra Mini at $1.50/$6 per Mtok, roughly a seventh of the flagship's $10/$50 for what OpenAI's own docs describe as 80-85% of its eval profile. Same 1.05M context and the same five-effort reasoning selector as full Astra; the architecture is billed as a distill rather than a separate training run. The published table mirrors Astra's suite row-for-row at reduced figures — a cleaner family story than the GPT-5.6 tier split, and the first GPT-6 row cheap enough for high-volume agent loops.

ReasoningCodingVisionFunction callingTool useAgentic
2731.6
Elo · rank #18
Parameters
Undisclosed
Active params
—
Context
1.05M tokens
Architecture
Proprietary distilled tier of GPT-6 Astra (size undisclosed); 1.05M-token context, text+image in / text out, five selectable reasoning efforts (low → max)
License
Proprietary
Languages
—
API price (in/out)
$1.5 / $6
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
Agents' Last ExamAgents47.8%#8
best: GPT-6 Astra · 59.3%
ARC-AGI-2Reasoning82.4%#7
best: GPT-6 Astra · 94.6%
AutomationBenchAgents33.6%#6
best: Qwen3.8-Max-0902 · 50.8%
best: GPT-6 Astra · 95.9%
ExploitBenchAgents88.4%#2
best: GPT-6 Astra · 100.0%
best: Claude Fable 5 · 64.9%
FrontierCodeCoding41.7%#8
best: Claude Fable 5 · 53.5%
GPQA DiamondReasoning89.4%#40
best: GPT-6 Astra · 96.0%
Humanity's Last ExamReasoning48.9%#14
best: Claude Opus 5 · 64.7%
best: Muse Spark 1.3 · 98.1%
best: GPT-6 Astra · 100.0%
ScreenSpot-ProVision84.3%#3
best: GPT-6 Astra · 92.7%
SRE-BenchCoding71.2%#2
best: GPT-6 Astra · 88.0%
SWE-bench ProCoding66.8%#7
best: Claude Fable 5.1 · 81.2%
Terminal-Bench 4.0Agents44.2%#5
best: Claude Mythos 5.1 · 60.9%
Run it locally
Closed weights — available via API only. No local deployment.
Input / M tok
$1.5
Output / M tok
$6
GPT-6 family
Elo progression across releases
API price $1.5/$6 · each benchmark row carries its own source badge (see methodology)