Benchmark Almanac

dfbench

Claude Mythos 5 69% leads since August 2026
Not in model cards

Agents do defensive security work on held-out code: find vulnerabilities, judge whether reported findings are real, and explain what a diff changed. Each part is scored separately.

Released August 19th, 2026 as dfbench v1. Detection is scored by recall and plotted against cost per task. 12 models in the Fireworks index.

Made by
depthfirst
Released
August 2026
Kind
Finding and exploiting real bugs
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
Kept private
Strongest model tested
Claude Opus 5.5 #4 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026
20
Entries on its own leaderboard, 18 models we track

Claude Mythos 5 leads with 69% among models released by September 2026

Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

20 entries. Scroll sideways for the rest.

  1. 69.0 Claude Mythos 5
  2. 65.7 GPT-5.6 Sol
  3. 62.2 dfs-large1
  4. 59.3 GPT-6 Sol
  5. 59.0 Grok 4.7
  6. 57.3 DeepSeek V4.1 Flash
  7. 57.2 Grok 4.5
  8. 55.0 MiMo-V2.6 Pro
  9. 54.9 Claude Opus 5.5
  10. 53.5 MiMo-V2.6 Flash
  11. 52.4 GPT-5.6 Luna
  12. 48.0 Kimi K3
  13. 47.8 Claude Opus 5
  14. 47.4 Muse Spark 1.3
  15. 45.7 GPT-6 Luna
  16. 44.2 Gemini 3.8 Flash
  17. 40.8 Qwen 3.8
  18. 40.7 GLM-5.2
  19. 31.8 DeepSeek-V4-Flash
  20. 20.9 Gemini 3.6 Flash
  1. 1 Claude Mythos 5 69%
  2. 2 GPT-5.6 Sol xhigh 65.7%
  3. 3 dfs-large1 62.2%
  4. 4 GPT-6 Sol xhigh 59.3%
  5. 5 Grok 4.7 xhigh 59%
  6. 6 DeepSeek V4.1 Flash high 57.3%
  7. 7 Grok 4.5 high 57.2%
  8. 8 MiMo-V2.6 Pro 55%
  9. 9 Claude Opus 5.5 high 54.9%
  10. 10 MiMo-V2.6 Flash 53.5%
  11. 11 GPT-5.6 Luna xhigh 52.4%
  12. 12 Kimi K3 max 48%
  13. 13 Claude Opus 5 med 47.8%
  14. 14 Muse Spark 1.3 xhigh 47.4%
  15. 15 GPT-6 Luna xhigh 45.7%
  16. 16 Gemini 3.8 Flash high 44.2%
  17. 17 Qwen 3.8 max 40.8%
  18. 18 GLM-5.2 xhigh 40.7%
  19. 19 DeepSeek-V4-Flash max 31.8%
  20. 20 Gemini 3.6 Flash high 20.9%

Source: Official dfbench leaderboard

#EntryScoreSettingDateSource
1 Claude Mythos 5 69% Jun 2026 Official leaderboard
2 GPT-5.6 Sol 65.7% xhigh Jul 2026 Official leaderboard
3 dfs-large1 62.2% Official leaderboard
4 GPT-6 Sol 59.3% xhigh Sep 2026 Official leaderboard
5 Grok 4.7 59% xhigh Sep 2026 Official leaderboard
6 DeepSeek V4.1 Flash 57.3% high Sep 2026 Official leaderboard
7 Grok 4.5 57.2% high Jul 2026 Official leaderboard
8 MiMo-V2.6 Pro 55% Sep 2026 Official leaderboard
9 Claude Opus 5.5 54.9% high Sep 2026 Official leaderboard
10 MiMo-V2.6 Flash 53.5% Sep 2026 Official leaderboard
11 GPT-5.6 Luna 52.4% xhigh Jul 2026 Official leaderboard
12 Kimi K3 48% max Jul 2026 Official leaderboard
13 Claude Opus 5 47.8% med Jul 2026 Official leaderboard
14 Muse Spark 1.3 47.4% xhigh Sep 2026 Official leaderboard
15 GPT-6 Luna 45.7% xhigh Sep 2026 Official leaderboard
16 Gemini 3.8 Flash 44.2% high Sep 2026 Official leaderboard
17 Qwen 3.8 40.8% max Official leaderboard
18 GLM-5.2 40.7% xhigh Jun 2026 Official leaderboard
19 DeepSeek-V4-Flash 31.8% max Apr 2026 Official leaderboard
20 Gemini 3.6 Flash 20.9% high Jul 2026 Official leaderboard

The best score, 69%, costs $99 a task

Runs on the cost frontier score higher than any run that's cheaper. GPT-5.6 Sol at xhigh comes within 3.3 points of the best for $43 a task, about half the cost. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Anthropic OpenAI DeepSeek Xiaomi Corp Cost frontier One model at each effort level
0% 20% 40% 60% 80% $0.10 $1.00 $10 $100 Cost per task, log scale Claude Mythos 5 GPT-5.6 Sol, xhigh GPT-5.6 Sol, high GPT-6 Sol, high DeepSeek V4.1 Flash, high MiMo-V2.6 Flash GPT-6 Luna, xhigh 0% 20% 40% 60% 80% $0.10 $1.00 $10 $100 Cost per task, log scale Claude Mythos 5 GPT-5.6 Sol, high

Source: Official dfbench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Mythos 5 69% $99
GPT-5.6 Sol xhigh 65.7% $43
GPT-5.6 Sol high 59.6% $26
GPT-6 Sol xhigh 59.3% $13
Grok 4.7 xhigh 59% $20
GPT-6 Sol high 57.7% $6.45
DeepSeek V4.1 Flash high 57.3% $1.69
Grok 4.5 high 57.2% $7.70
Grok 4.7 high 56.6% $16
MiMo-V2.6 Pro 55% $1.38
Claude Opus 5.5 high 54.9% $8.42
MiMo-V2.6 Flash 53.5% $0.86
GPT-5.6 Luna xhigh 52.4% $2.53
GPT-6 Sol med 49.6% $3.43
GPT-5.6 Sol med 48.9% $9.51
Claude Opus 5.5 med 48.1% $7.66
Kimi K3 max 48% $9.10
Claude Opus 5 med 47.8% $22
Muse Spark 1.3 xhigh 47.4% $10
Kimi K3 high 46.7% $6.22
GPT-6 Luna xhigh 45.7% $0.43
Claude Opus 5 high 45.4% $23
Gemini 3.8 Flash high 44.2% $8.03
GLM-5.2 xhigh 40.7% $9.72
GPT-5.6 Luna high 39.6% $1.57
GLM-5.2 high 38.2% $4.09
GPT-6 Luna high 38.1% $0.28
GPT-6 Luna med 37.3% $0.15
DeepSeek-V4-Flash max 31.8% $1.23
DeepSeek-V4-Flash high 30.1% $1.03
GPT-5.6 Luna med 28% $0.69
Gemini 3.6 Flash high 20.9% $7.44

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Own leaderboard Not tested since Sep 2026
Claude Mythos 5 69%

Source: Official dfbench leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Mythos 5 DeepSeek-V4-Flash 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Mythos 5 DeepSeek-V4-Flash

Source: Official dfbench leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Cybersecurity

All 115 in Cybersecurity
  • ExploitGym In use 869 real-vulnerability exploit-development tasks (userspace, V8, Linux kernel). Success means capturing a flag through the intended vulnerability. GPT-6 Astra 42.4%
  • ExploitBench In use 41 recent V8 JavaScript-engine N-day vulnerabilities. Progress measured by 16 grader-verified capability flags across five tiers. GPT-6 Astra 100%
  • CyberBench v1.1 (Vals) In use Agents work on OSS-Fuzz memory-safety bugs from ARVO images in two tracks. They write a proof-of-concept input that crashes the target, or patch the source so the crash stops and tests still pass. Muse Wonder 80.4%
  • SRE-Bench In use Software reverse engineering without source: 262 binary instances from 19 private programs, each with six graded objectives. GPT-6 Astra 56.9%
  • CyScenarioBench In use Scenario-based long-horizon security challenges in sandboxes, organized in task, scenario and campaign tiers. Evaluation set kept private. Gemini 4 Argon 100%
  • ExploitBench - Internal Port In use Agents develop working exploits for recently disclosed vulnerabilities to limit contamination. Scored as arbitrary code execution success rate. GPT-6 Cyber 42%