Best LLM for threat hunting (November 2026)
Claude Opus 5.5 leads BlueBench's threat hunting track at 76.1. Rankings for 26 models with cost per task, reliability, and the best value and open-weight picks.
Summary
Claude Opus 5.5 is the best LLM for threat hunting as of November 2026, based on BlueBench's threat hunting track, where it scored 76.1 against 74.5 for Gemini 4 Argon and 72.6 for Grok 4.6. The best value is DeepSeek V4.1 Flash, an open-weight model that scored 68.5 at $0.17 per task, against $3.54 for Claude Opus 5.5. Hunting ranks models differently from the overall leaderboard: GPT-6 Astra, second overall, placed 10th. This ranking is refreshed monthly.
- Top threat hunting score, Claude Opus 5.5
- 76.1
- Cost per task for DeepSeek V4.1 Flash, which scored 68.5 on hunting
- $0.17
- GPT-6 Astra's threat hunting rank; it is 2nd on the overall leaderboard
- 10th
- Open-weight models in the threat hunting top nine
- 4 of top 9
Key findings
- Claude Opus 5.5 led BlueBench's threat hunting track at 76.1, ahead of Gemini 4 Argon (74.5) and Grok 4.6 (72.6), at $3.54 per task against $6.76 for Gemini 4 Argon.
- DeepSeek V4.1 Flash, an open-weight model, scored 68.5 at $0.17 per task: 7.6 points below the leader for about a twentieth of the cost.
- GPT-6 Astra, second on the overall leaderboard, ranked 10th on threat hunting at 62.8. No OpenAI model made the track's top nine.
- Open-ended simulated hunts carry four-sevenths of the track and separate models most: GPT-6 Astra scored 99.3% on the macOS hunt questions and 39.0% on Simulation-001's hunts.
- Reliability is measured across the whole suite, not per track: Claude Opus 5.5 ranked 2nd at 56.2 and Gemini 4 Argon 12th at 38.0.
This ranking is refreshed monthly. The November 2026 edition uses the October 2026 results of BlueBench, Cotool's benchmark for defensive security work, and ranks all 26 tested models on its threat hunting track alone. For the ranking across all four tracks, see which AI models are best for security work.
Threat hunting leaderboard, November 2026
The top ten models on the threat hunting track of the capability leaderboard:
| Hunt rank | Model | Weights | Threat hunting | Overall rank | Cost per task | Reliability |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Closed | 76.1 | 1 | $3.54 | 56.2 |
| 2 | Gemini 4 Argon | Closed | 74.5 | 3 | $6.76 | 38.0 |
| 3 | Grok 4.6 | Closed | 72.6 | 4 | $3.64 | 47.4 |
| 4 | Claude Opus 4.8 | Closed | 68.6 | 8 | $1.56 | 17.3 |
| 5 | DeepSeek V4.1 Flash | Open | 68.5 | 9 | $0.17 | 38.5 |
| 6 | GLM-5.3 | Open | 68.2 | 11 | $3.27 | 37.1 |
| 7 | Claude Sonnet 5.5 | Closed | 66.9 | 5 | $1.01 | 50.0 |
| 8 | Kimi K3 | Open | 64.3 | 16 | $2.62 | 36.3 |
| 9 | GLM-5.3 Flash | Open | 63.2 | 12 | $0.11 | 37.5 |
| 10 | GPT-6 Astra | Closed | 62.8 | 2 | $9.60 | 64.1 |
Two columns cover the whole suite rather than hunting. Cost per task is each model's average across all 74 tasks, and BlueBench keeps it benchmark-wide in track views (methodology). Hunts cost more than that average: on Simulation-001, where three of the five tasks are hunts, Claude Opus 5.5 cost $5.61 per task. Reliability, from the reliability leaderboard, is 100 minus the average gap between a model's best and worst scores in each benchmark. It measures consistency across tasks, not repeat runs of one hunt.
Best overall, best value, and best open-weight
Best overall: Claude Opus 5.5. It led the track at 76.1, 1.6 points ahead of Gemini 4 Argon, ranked second of 26 on reliability, and led the open-ended hunts in Simulation-001 at 68.3%. Gemini 4 Argon is the closest on score and led the hunts in Simulation-002 (86.1%) and the AWS intrusion (87.5%). It also costs $6.76 per task, ranked 12th on reliability, and recorded two refusals in Simulation-001 that stay in its score.
Best value: DeepSeek V4.1 Flash. It scored 68.5 at $0.17 per task, 7.6 points below Claude Opus 5.5 for about a twentieth of the cost. Claude Opus 4.8 scored 68.6, a tenth of a point higher, at $1.56, with the 24th-ranked reliability of 26. Four models sit on this track's cost frontier, where no model scores higher for less: GLM-5.3 Flash ($0.11), DeepSeek V4.1 Flash, Claude Opus 4.8, and Claude Opus 5.5.
Best open-weight: DeepSeek V4.1 Flash. GLM-5.3 is close at 68.2 but costs $3.27 per task; it led the hunt reports in the Windows intrusion at 89.7%. Four of the top nine hunting models are open-weight, against one in the overall top nine, which matters for teams that can't send hunt data to a closed-model provider.
What the threat hunting track tests
BlueBench defines a hunt as starting "from a weak lead," working "across raw multi-source telemetry without assuming compromise," and reporting what is actually happening. A single fixed LLM judge grades each hunt report against a hidden rubric. Five of the six benchmarks contribute hunts, in two shapes. For the practice itself, see what threat hunting is.
- Guided hunts on real intrusions. The three Intrusion benchmarks replay hands-on-keyboard intrusions captured in production telemetry, built with Threat Hunting Labs. The macOS case asks focused questions, such as which built-in utility decodes content just before the payload runs (base64 -d). The Windows case includes two hunt reports, and the AWS case scores a hunting rubric on the same report as incident response.
- Open-ended hunts in simulated environments. Simulation-001 has three hunts (a mail-and-endpoint compromise, a webshell on a Linux DMZ portal, and DGA botnet beaconing), and Simulation-002 has one (a password spray followed by account manipulation). Each starts from a single weak alert, in environments with benign lookalikes "that punish pattern-matching without verification."
The track weights each Intrusion benchmark once and the simulated hunts four times, so the open-ended hunts carry four-sevenths of the score.
Why the ranking differs from the overall leaderboard
The guided intrusion questions are close to saturated, and the simulated hunts are where models separate. On the macOS hunt questions, 22 of 26 models scored above 80%, and GPT-6 Astra and GPT-5.6 Sol tied at 99.3%. On Simulation-001's open-ended hunts, the leader scored 68.3% and GPT-6 Astra scored 39.0%; GPT-6 Astra scored 52.8% on the Simulation-002 hunt.
Because those hunts weigh most, models that answer guided questions well but stall on an open lead drop down the table: GPT-6 Astra is second overall and 10th on hunting, GPT-5.6 Sol sixth and 11th, GPT-6.1 Sol seventh and 13th. Models that hunt better than they write incident reports move up: Kimi K3 is 16th overall and eighth on hunting, GLM-5.3 11th and sixth.
Read close gaps loosely. The simulated hunts are four tasks with three trials each, so one case moves a score several points; Gemini 4 Argon's 86.1% on Simulation-002 is a single hunt. The benchmark pages state that "close scores do not establish a statistically significant ordering."
Other benchmarks also find open-ended hunting hard. The Cyber Defense Benchmark gives agents raw Windows event logs with no guided questions and asks for the exact timestamps of malicious events across 106 attack procedures; its best model flagged 3.8% of malicious events on average. Its scoring is stricter than a rubric, so the numbers don't compare directly.
When a cheaper model is good enough
Choose by the shape of the hunt, not by the overall rank. AI threat hunting covers how agents run these hunts in practice.
- Sweeps with a known question. Checking telemetry for a published indicator set or one technique resembles the guided intrusion questions, where most models score high. DeepSeek V4.1 Flash scored 87.8% on the macOS hunt questions and GLM-5.3 Flash 91.3%, at $0.17 and $0.11 per task. Cotool's intel-driven hunts, which check new threat intel against your environment for relevance and exposure, are closer to this shape.
- Open-ended hypothesis hunts. When the agent has to decide where to look, the leader's margin is real: on Simulation-001's hunts, Claude Opus 5.5 scored 68.3% and DeepSeek V4.1 Flash 55.6%. Put the strongest model here.
- A cheap first pass with escalation. Run scheduled hunts on a low-cost model and send anything it flags to a stronger one. That keeps volume near the cheap model's price, but the second model never sees what the first one missed. Grading each run against acceptance criteria, as Cotool's AI judge does, flags skipped steps such as an EDR that was never queried, though not a lead neither model followed.
- Hunts that end in a report. If the same agent writes the incident report, check that track too: DeepSeek V4.1 Flash scored 60.9 on incident response against 87.8 for Claude Opus 5.5.
- Refusals. Hunting sends attacker commands and payloads to a model. Claude Opus 5 and Claude Sonnet 5 refused a material share of BlueBench tasks on cybersecurity grounds and scored 36.1 and 23.8 on hunting. Test a model on your own hunt data before scheduling it; abliterated models covers one workaround and its risks.
How Cotool runs hunts on these results
Cotool publishes BlueBench and runs hunts as agents. Its hunt agents threat-model the environment, form hypotheses about attacker behavior, and hunt continuously; when a hunt returns findings, the agent backtests and ships a new detection, the step covered in turning a threat hunt into a detection. The worked example on Cotool's hunt page starts from a threat-model hypothesis, a stolen GitHub token used to clone private repositories from an unmanaged device. The agent pulls 14 days of GitHub audit logs through the API, joins Okta sessions and device posture, finds one burst of 212 private repositories cloned in 6 minutes, and generates a detection for review. Each step is a choice about where to look next, which is the open-ended shape the simulated hunts score and where Claude Opus 5.5 and GPT-6 Astra were 29.3 points apart.
In Cotool's AI SOC, the model is set per agent, from Anthropic, OpenAI, Google, xAI, open-weight models, or your own endpoint, and Cotool recommends a starting model for each agent's job from BlueBench. The page's own guidance is "frontier models for hard cases, cheaper ones for volume," so a team could run hypothesis hunts on Claude Opus 5.5 and indicator sweeps on DeepSeek V4.1 Flash. An AI judge grades every run against your acceptance criteria, so after a model switch the eval scores on your own hunts show whether it helped, which BlueBench can't tell you. When a model refuses or a provider goes down, Cotool Router moves the work to another provider or an open-weight model without interrupting the agent. Picking a model automatically by task, such as a scheduled hunt versus a detection agent, is work Cotool describes as in progress. The detection side of the same data is in best LLM for detection engineering.
FAQ
- What is the best AI model for threat hunting?
- On BlueBench's threat hunting track, as of November 2026, the best AI model for threat hunting is Claude Opus 5.5 with a score of 76.1, followed by Gemini 4 Argon (74.5) and Grok 4.6 (72.6). Claude Opus 5.5 costs $3.54 per task across the suite and ranked second of 26 models on reliability. The ranking is refreshed monthly as new models are tested.
- What is the best open-weight LLM for threat hunting?
- The best open-weight LLM for threat hunting on BlueBench is DeepSeek V4.1 Flash, which scored 68.5 at $0.17 per task and ranked fifth of 26 on the track. GLM-5.3 scored 68.2 at $3.27, Kimi K3 64.3, and GLM-5.3 Flash 63.2 at $0.11. Four of the track's top nine models are open-weight.
- Is GPT-6 Astra good for threat hunting?
- GPT-6 Astra is second on BlueBench overall but ranked 10th on threat hunting, at 62.8, and its $9.60 per task was the highest cost of the 26 models. GPT-6 Astra tied for first on the guided macOS hunt questions at 99.3% but scored 39.0% on the open-ended hunts in Simulation-001, which weigh more in the track score.
- Can an LLM run threat hunts on its own?
- An LLM can run threat hunts, but no model has mastered open-ended hunting. On BlueBench's Simulation-001 hunts, which start from a single weak alert, the best model scored 68.3%. On the stricter Cyber Defense Benchmark, the best model flagged 3.8% of malicious events. Review hunt findings before acting on them.
- How does BlueBench score threat hunting?
- BlueBench scores threat hunting from free-form hunt reports, graded by a single fixed LLM judge against a hidden rubric. Each of the three real-intrusion benchmarks counts once and the simulated hunts count four times, so the simulated hunts make up four-sevenths of the threat hunting score.
Sources
- BlueBench — AI Benchmark for Defensive Security · accessed November 16, 2026
- BlueBench Capability Leaderboard · accessed November 16, 2026
- BlueBench Cost Efficiency · accessed November 16, 2026
- BlueBench Reliability Leaderboard · accessed November 16, 2026
- BlueBench Methodology · accessed November 16, 2026
- macOS Threat Investigation · accessed November 16, 2026
- Windows Enterprise Intrusion · accessed November 16, 2026
- AWS Cloud Intrusion · accessed November 16, 2026
- Initial Access & Command-and-Control · accessed November 16, 2026
- Identity & Active Directory Attacks · accessed November 16, 2026
- Cyber Defense Benchmark: Agentic Threat Hunting Evaluation for LLMs in SecOps · accessed November 16, 2026
- AI Threat Hunting and Actionable Threat Intel - Cotool · accessed November 16, 2026
- AI SOC Without the Black Box - Cotool · accessed November 16, 2026
- Introducing Cotool Router · accessed November 16, 2026





