GPT-6 Astra catches 9 more of every 100 critical alerts than your triage model at 10% lower cost

New release·GPT-6 Astra shipped to trusted access Oct 7, 10:00·Scored 640 cases by 11:12

Your alert triage runs GPT-5.6 Sol at $1.20 per alert and catches 78.4% of critical alerts
GPT-6 Astra catches 87.4% at $1.08 and now holds 5 of the 6 points on your frontier

Recall on critical alerts: 87.4%▲ +9.0 pts
Cost per alert: $1.08▼ 0.9×
Estimated monthly cost at 150,000 alerts: $162k▼ −$18k

Claude Fable 5.1GPT-5.6 SolGPT-6.1 SolClaude Opus 5.5GPT-6 AstraProductionNew champion90%75%60%$6$4$2

Your frontierAlert triageFrontierGPT-6 AstraClaude Opus 5.5Others

What changed

All runs

21 models · 640 cases · avg@3

Model	Recall	Cost / alert	p95	Tokens / alert	Turns	Frontier
GPT-6 Astra Max	91.8%	$4.60	98.4s	96,200	64	Yes
GPT-6 Astra Extra High	90.6%	$2.85	71.2s	61,800	46	Yes
GPT-6 Astra High	89.5%	$1.72	52.6s	39,400	34	Yes
Claude Opus 5.5 Max	88.9%	$6.80	142.3s	118,400	88	No
GPT-6 Astra Medium	87.4%	$1.08	34.8s	24,600	23	Yes
Claude Opus 5.5 High	86.8%	$3.10	78.5s	54,600	49	No
Claude Fable 5.1 Max	86.1%	$6.20	176.5s	142,000	104	No
Claude Fable 5.1 High	84.7%	$4.10	118.2s	89,500	72	No
Claude Opus 5.5 Medium	84.0%	$1.90	49.8s	33,200	32	No
GPT-5.6 Sol Extra High	83.2%	$3.30	88.1s	71,500	58	No
GPT-6.1 Sol High	82.4%	$2.40	63.9s	52,100	41	No
GPT-6 Astra Low	82.1%	$0.52	19.7s	11,800	13	Yes
Claude Fable 5.1 Medium	81.3%	$2.70	82.1s	56,300	51	No
GPT-5.6 Sol High	81.0%	$1.95	59.4s	44,800	39	No
GPT-6.1 Sol Medium	79.0%	$1.30	41.3s	31,700	28	No
GPT-5.6 Sol Medium	78.4%	$1.20	38.6s	28,900	27	No
Claude Opus 5.5 Low	77.5%	$0.85	26.4s	15,100	18	No
Gemini 3.8 Flash High	76.8%	$0.95	31.2s	68,400	39	No
GPT-6.1 Sol Low	73.6%	$0.70	22.8s	14,900	16	No
Gemini 3.8 Flash Medium	72.4%	$0.48	18.9s	39,100	24	Yes
GPT-5.6 Sol Low	71.6%	$0.62	21.5s	14,200	15	No

Where each model lands

Top right is best on both axes, faded dots are off the frontier

Speed against cost

Latency p95 · cost per alert

Gemini 3.8 Flash Medium is the cheapest and fastest run, at $0.48 and 18.9s per alert.

GPT-6 Astra Low costs 4 cents more and catches 9.7 points more, at 82.1%. GPT-6 Astra Medium is the pick, catching 87.4% at $1.08 and 34.8s, both under production's $1.20 and 38.6s.

Recall against speed

Recall · latency p95

GPT-6 Astra High gives the best recall without the wait, 89.5% at a 52.6s p95.

GPT-6 Astra Max adds only 2.3 points and takes 45.8s longer. Claude Opus 5.5 Max takes 142.3s to reach 88.9%, still below GPT-6 Astra High.

Token appetite

Recall · tokens per alert

GPT-6 Astra Medium is the most token-efficient top scorer, 87.4% on 24,600 tokens per alert.

Gemini 3.8 Flash High burns 68,400 tokens for 76.8%, nearly three times as many for 10.6 points less. GPT-6 Astra Medium uses 4,300 fewer tokens than production's 28,900.

Turns against recall

Recall · turns per alert

GPT-6 Astra Medium needs the fewest turns of any run above 85%, catching 87.4% in 23 turns.

That is 4 fewer turns than production for 9.0 points more recall. Claude Fable 5.1 Max needs 104 turns for 86.1%.

How this was scored

Cases: 640; from your alert queue, CVE backlog, and attack replays
Runs: avg@3; 3 runs per case, scores averaged
Grading: Ground truth; analyst labels, your test suites, and replayed attacks
Next run: Next release; OpenAI, Anthropic, and Google releases watched

Release history

Date	Release	Champion	Result
Oct 7	GPT-6 Astra	GPT-6 Astra	Better on alert triage and vulnerability remediation
Sep 23	GPT-6.1 Sol	GPT-5.6 Sol	Same on alert triage at 1.1× the cost
Sep 9	GPT-5.6 Sol	GPT-5.6 Sol	Better on alert triage and vulnerability remediation
Aug 12	Claude Opus 5.5	Claude Opus 5.5	Better on detection engineering
Jul 21	Gemini 3.8 Flash	Claude Opus 5.1	Slower on detection engineering
Jun 24	Claude Fable 5.1	Claude Opus 5.1	Same on detection engineering at 2× the cost