Claude Haiku 5.5 security benchmark rundown
Claude Haiku 5.5 (Anthropic) ranks 15th of 26 on BlueBench with 68.5 overall at $0.59 per task. Track scores, value against peers, strengths, weaknesses, and when to use it.
Verdict
Claude Haiku 5.5 ranks 15th of 26 on BlueBench with an overall score of 68.5, in the lower half of the leaderboard and 14.7 points behind Claude Opus 5.5. At $0.59 per task it is the 7th cheapest of 26, and DeepSeek V4.1 Flash scores higher (71.5) for less ($0.17). Against the 6 models priced within 2x of its cost per task, it scores 4.5 points above the peer median overall; it is strongest against them in detection engineering. Malware Analysis, weighted 10%, is a secondary weak spot (−8.3 vs the peer median). Its results vary more than most across the benchmarks (reliability 27.9, 20th of 26).
- BlueBench score
- 68.5
- #15 of 26 models
- Cost per task
- $0.59
- #7 cheapest of 26
- Reliability
- 27.9
- #20 of 26 models
- Overall vs cost peers
- +4.5
- Off the cost–score frontier
Capability by track
Incident response, threat hunting, and detection engineering each carry 30% of the BlueBench score; malware analysis carries 10%.
Claude Haiku 5.5 scores 68.5 overall, 15th of 26 on the capability leaderboard. Its best rank is in detection engineering (80.5, 9th of 26) and its worst is in malware analysis (67.8, 20th of 26).
| Track | Weight | Score | Rank | Peer median | vs peers |
|---|---|---|---|---|---|
| Incident Response | 30% | 65.5 | #16 of 26 | 62.3 | +3.2 |
| Threat Hunting | 30% | 59.9 | #14 of 26 | 53.1 | +6.8 |
| Detection Engineering | 30% | 80.5 | #9 of 26 | 71.7 | +8.8 |
| Malware Analysis | 10% | 67.8 | #20 of 26 | 76.1 | −8.3 |
Cost, reliability, and value
Cost is USD per evaluated task including repeated attempts. Reliability is 100 minus the mean score range across the six benchmarks.
Claude Haiku 5.5 averages $0.59 per task, including repeated attempts: the 7th cheapest of 26 models with a cost figure. It is off the cost–score frontier: DeepSeek V4.1 Flash scores 71.5 at $0.17, and 2 other models are also both cheaper and at least as good. Its peer group is the 6 models priced within 2x of its cost per task. Their median overall score is 64.0, so Claude Haiku 5.5 is 4.5 points ahead of its price range. Reliability, 100 minus its mean score range across the benchmarks, is 27.9 (20th of 26).
Cost peers: the 6 models priced within 2x of its cost per task
Strengths and weaknesses
Each track compared with the median of the 6 models priced within 2x of its cost per task. A gap of 5 points or more counts.
Against the 6 models priced within 2x of its cost per task, Claude Haiku 5.5 is strongest in detection engineering (+8.8) and threat hunting (+6.8). It lags them in malware analysis (−8.3). Differences are versus the peer median score in each track; ±5 points is the threshold.
Beats its cost peers
- Detection Engineering80.5 (+8.8)
- Threat Hunting59.9 (+6.8)
Lags its cost peers
- Malware Analysis67.8 (−8.3)
Nearest neighbors on the leaderboard
The models ranked just above and below it overall, with the largest track gaps.
- GPT-6.1 Terra sits 2 places above, 2.0 points higher overall (70.5), at 28% of the cost per task ($0.16); it is 12.7 points ahead in incident response and 8.0 points behind in detection engineering.
- GPT-5.6 Terra sits one place above, 1.4 points higher overall (69.9), at 1.7x the cost per task ($1.01); it is 5.3 points ahead in malware analysis.
- Kimi K3 sits one place below, 1.2 points lower overall (67.3), at 4.4x the cost per task ($2.62); it is 9.4 points ahead in malware analysis and 18.5 points behind in detection engineering.
- GLM-5.2 sits 2 places below, 3.0 points lower overall (65.5), at 1.7x the cost per task ($1.03); it is 14.1 points ahead in malware analysis and 7.5 points behind in threat hunting.
| Rank | Model | Score | $/task |
|---|---|---|---|
| 13 | GPT-6.1 Terra | 70.5 | $0.16 |
| 14 | GPT-5.6 Terra | 69.9 | $1.01 |
| 15 | Claude Haiku 5.5 | 68.5 | $0.59 |
| 16 | Kimi K3 | 67.3 | $2.62 |
| 17 | GLM-5.2 | 65.5 | $1.03 |
Best for
Alternatives are limited to models costing at most 1.5x as much, except the overall leader and the cheaper-model pick.
- Use itUse it for detection engineering (80.5, +8.8) and threat hunting (59.9, +6.8) at its price: those tracks beat the median of the 6 models priced within 2x of its cost per task by 5 points or more.
- InsteadUse DeepSeek V4.1 Flash instead in most cases: it scores 71.5 overall (+3.0) at $0.17 per task (29% of the cost per task).
- InsteadUse GPT-6.1 Terra instead if consistent results matter: its reliability is 45.1 (vs 27.9) with an overall score of 70.5 at $0.16 per task.
- InsteadUse Claude Opus 5.5 instead if the highest overall score matters more than cost: it scores 83.2 (14.7 points higher) at $3.54 per task.
Notable benchmark results
Each BlueBench benchmark is a separate intrusion, real or simulated, with its own tasks and grading.
- Its best standing is on Simulation-002 (Identity & Active Directory Attacks): 72.5 (10th of 26).
- Its scores range from 54.3 on Simulation-001 to 86.9 on Intrusion-001, a spread of 32.6 points.
- 2 recorded refusals across the benchmarks; refused attempts stay in the score.
| Benchmark | Score | Rank |
|---|---|---|
| Simulation-003Impact & Exfiltration | 64.4 | #15 of 26 |
| Simulation-002Identity & Active Directory Attacks | 72.5 | #10 of 26 |
| Simulation-001Initial Access & Command-and-Control | 54.3 | #11 of 26 |
| Intrusion-003AWS Cloud Intrusion | 73.0 | #16 of 26 |
| Intrusion-002Windows Enterprise Intrusion | 66.7 | #18 of 26 |
| Intrusion-001macOS Threat Investigation | 86.9 | #18 of 26 |
About this rundown
- Provider
- Anthropic · closed weight
- Tested
- Oct 14, 2026
- Suite
- 74 tasks across 6 benchmarks, evaluated in a common harness
- Data updated
- Oct 27, 2026
- How it is written
- Generated from the BlueBench result files with fixed rules: cost peers are models priced within 2x per task (or the nearest by cost when fewer than three are), a track is a strength or weakness at 5+ points from the peer median, and ranks share a place on ties. The text is regenerated whenever the results change.
AI for the blue team.
Run Cotool's harness in your environment to get real security work done