cd /news/ai-research/claude-opus-5-5-on-iac-bench-2026-10 · home › topics › ai-research › article
[ARTICLE · art-173574] src=massdriver.cloud ↗ pub= topic=ai-research verified=true sentiment=↑ positive

Claude Opus 5.5 on iac-bench 2026-10

Anthropic's Claude Opus 5.5 scored an overall composite of 88.9 (95% CI 87.5–91.5) on the iac-bench 2026-10 infrastructure-as-code benchmark, ranking 2nd of 16 models behind Claude Fable 5.2, according to Massdriver's LLM rankings. Across 9 attempts on 3 fixtures (S3, Cloud SQL, App Service), Claude Opus 5.5 failed 0 mandatory security gates and produced 0 insecure attempts, with a worst attempt of 85.0 and an estimated total cost of $3.58 at list prices of $4.00/$20.00 per million input/output tokens. Its lead over GPT-6.1 Sol, Claude Fable 5.1 and GPT-6 Astra fell within the three-sample margin (bootstrap win probabilities 0.74, 0.84 and 0.86), while its modules hard-coded security settings and merged enforced database flags last so inputs could not disable them.

recommendation

Evaluate IaC-generating models on multiple fixtures with 3 samples each and report confidence intervals; Claude Opus 5.5's 88.9 composite carried a 95% CI of 87.5–91.5, so single-run rankings can mislead.

more in playbook: Evals and testing for LLM apps
by read4 min views5 publishedNov 4, 2026
Claude Opus 5.5 on iac-bench 2026-10
Image: Massdriver (auto-discovered)

New: LLM Rankings are live!

See how the top LLMs rank on real infrastructure tasks

How Claude Opus 5.5 did on iac-bench 2026-10: an overall composite of 88.9, scores per fixture, strengths, weaknesses and the patterns in its OpenTofu modules.

Every model evaluated on the 2026-10 fixtures side by side, with a sortable leaderboard, picks by use case and the cost of each: LLM rankings.

| Developer | Anthropic | | API identifier | claude-opus-5-5 | | Access | Anthropic API | | Weights | Closed | | List price (USD / M tokens, in / out) | 4.00 / 20.00 (OpenRouter, 2026-10-23) | | Evaluated | 2026-10-22 (generation) · 2026-10-23 (judge) | | Attempts | 9 (3 fixtures × 3 samples) | | Overall composite | 88.9 (95% CI 87.5–91.5) | | Rank | 2 of 16 (95% rank interval 2–5) | | Mechanical only / judge only | 86.4 (rank 2) / 94.7 (rank 2) | | Failed gates · insecure attempts | 0/9 · 0/9 | | Estimated cost, all 9 attempts | $3.58 |

Claude Opus 5.5 has the second-highest overall score in the 2026-10 release, behind Claude Fable 5.2. It is the only model besides Claude Fable 5.2 whose 95% rank interval stays within the top five. All 9 of its modules planned, and none failed a mandatory security item. Its weakest attempt (85.0) is the strongest worst case in the release after Claude Fable 5.2's (88.6). Its overall lead over GPT-6.1 Sol, Claude Fable 5.1 and GPT-6 Astra is within the margin of three samples (bootstrap win probabilities 0.74, 0.84 and 0.86). The clearer distinction is consistency: no fixture produced a weak attempt.

Its modules fix the security settings in code and do not offer them as inputs. They derive values from a small set of concept inputs, and they include most of the service-specific practices the benchmark checks for: 7/7 on S3, 8/9 on Cloud SQL and 5/5 on App Service, in every sample.

Fixture Composite median (range) Rubric median Expertise Judge median ✗ gate ⚠ security Median inputs Median output tokens Median time Median cost
G001 · S3 87.9 (87.9–90.2) 0.826 7/7 97.4 0/3 0/3 10 11,128 243 s $0.328
G002 · Cloud SQL 88.6 (85.0–93.5) 0.832 8/9 100.0 0/3 0/3 29 8,733 112 s $0.245
G003 · App Service 90.2 (89.5–93.2) 0.911 5/5 86.8 0/3 0/3 28 21,622 468 s $0.635

Rubric items not at full credit, counted over attempts that planned: G001: A5 3/3, B1 3/3, H9 2/3, H13 3/3; G002: A8 3/3, B1 2/3, H3 1/3, H4 3/3, H5 3/3, H11 3/3; G003: A7 3/3. Definitions are in Scoring criteria.

ipv4_enabled = false and ssl_mode = "ENCRYPTED_ONLY" are literals.database_flags input. The module merges its own required flags (IAM authentication, pgaudit, connection logging) last, so no input can turn them off (excerpt below).force_destroy is refused in production, and on Cloud SQL the Enterprise Plus log retention is checked against the edition.A8 is lost in 3/3.A7 is lost in 3/3.deletion_policy (H11 lost in 3/3).H4 and H5 earn partial credit.H3). H13, 3/3), and two of three offer no access logging at all ( H9).>= 5.0, < 7.0, or unbounded >= 5.0, in 5/9 attempts (B1). ignore_changes = [worker_count]; environment (apps-${var.environment}), and so does high availability. main.tf by default. G002/claude-opus-5-5-e3/main.tf lines 7–17

  enforced_flags = {
    "cloudsql.iam_authentication" = "on"
    "cloudsql.enable_pgaudit"     = "on"
    "pgaudit.log"                 = var.pgaudit_log_classes
    "log_connections"             = "on"
    "log_disconnections"          = "on"
  }

  database_flags = merge(var.database_flags, local.enforced_flags)

The DBA team keeps an open-ended flags input, but the module's own flags are merged last, so no input can switch off IAM authentication or connection auditing.

The judge is Claude Opus 5.5, so this page grades the judge's own model. Its grades for this model are 97.4 (S3) and 100 (Cloud SQL). Without the judge, ranking on the mechanical score alone, Claude Opus 5.5 still ranks second (86.4), behind Claude Fable 5.2, so the top of the ranking does not depend on the judge. See Methodology §5.3.

How Claude Fable 5.1 did on iac-bench 2026-10: an overall composite of 86.9, scores per fixture, strengths, weaknesses and the patterns in its OpenTofu modules.

How Claude Fable 5.2 did on iac-bench 2026-10: an overall composite of 93.9, scores per fixture, strengths, weaknesses and the patterns in its OpenTofu modules.

How Claude Haiku 5.5 did on iac-bench 2026-10: an overall composite of 68.0, scores per fixture, strengths, weaknesses and the patterns in its OpenTofu modules.

── more in #ai-research 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/claude-opus-5-5-on-i…] indexed:0 read:4min 2026-11-04 · —