gpt-5-5 — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited gpt-5-5 (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
What it is: GPT-5.5 is OpenAI's current GPT frontier model generation (~1M context). This skill is the routing decision: when a task's shape favors GPT-5.5's strengths enough to send it there instead of a Claude tier.
Mental model: Treat frontier models as specialists with overlapping but non-identical edges. GPT-5.5 and Claude Opus are within a couple of points on aggregate coding benchmarks, but they diverge by task type: GPT-5.5 leads on terminal/CLI workflows and precise tool-use; Claude Opus leads on architecture and complex multi-file resolution. Routing is matching the task's dominant axis to the model that owns that axis — not crowning one model "best."
Why it exists: "Just use the strongest model" is the wrong default — it ignores that strength is task-shaped, and it ignores cost. An explicit per-task routing rule captures the real per-benchmark splits so concrete-implementation and review work lands on GPT-5.5 while architecture work stays on Opus, instead of being decided by habit.
What it is NOT: It is not the Codex CLI (that is the harness that runs GPT — a separate decision), not the OpenAI SDK/API, and not a claim that GPT-5.5 beats Claude everywhere. It is a model-lane selector, not a leaderboard.
Adjacent concepts: the Codex CLI (the harness for GPT), the Claude tier skills (claude-opus, claude-sonnet, claude-haiku), gemini-pro (the third frontier lane), and skill-router (request-time dispatch among local skills, a different layer).
One-line analogy: Routing GPT-5.5 vs Opus is choosing a surgeon by procedure, not by overall reputation — the more decorated surgeon is still the wrong pick for the operation the other one specializes in.
Common misconception: That the higher SWE-bench Verified score makes GPT-5.5 the universal coding choice. The aggregate margin is within noise and benchmark-contamination caveats; on complex multi-file resolution (SWE-bench Pro) Claude Opus leads, so the verified-bench headline does not generalize to architecture-heavy work.
gemini-pro (the third frontier lane), and skill-router (request-time local-skill dispatch)Routing is matching, not ranking. The tempting shortcut — "use the strongest model and stop thinking" — is wrong twice over: it treats model strength as a single scalar when it is actually a profile of task-shaped edges, and it ignores cost entirely. This skill refuses both errors. It insists the routing primitive is the task's dominant axis, not a model's aggregate benchmark rank, because two near-equal models split by task type and the per-task winner can flip even when one model leads on paper. It also refuses to let headline price be the primitive: on a flat subscription the per-token USD is notional, and even on API the capability-per-task delta usually dominates the price delta. The discipline is to name the task's dominant axis first, route to the model that owns that axis, and treat both "always strongest" and "always cheapest" as reflexes to override.
Send a task to GPT-5.5 when its dominant axis is one GPT-5.5 owns:
| Task axis | Route | Why |
|---|---|---|
| Infrastructure scripts, CI pipelines | GPT-5.5 | Strong on concrete, well-specified implementation |
| CLI / terminal-heavy iteration (planning, tool coordination in a shell) | GPT-5.5 | Leads Terminal-Bench 2.0 (82.7%) — its most decisive win |
| Analytical code review of a diff; security review | GPT-5.5 | Strong analytical diff-reading and threat-surface work |
| Precise tool use + file navigation | GPT-5.5 | Holds a slight edge here |
| Broad architectural reasoning across a large codebase | Claude Opus | Opus leads architecture work |
| Complex multi-file GitHub issue resolution | Claude Opus | Opus leads SWE-bench Pro (64.3% vs 58.6%) |
| Multi-tool agent coordination | Claude | Claude's tool-coordination edge (MCP Atlas) |
| Cost-sensitive but still-capable implementation | Claude Sonnet | Cheaper than both frontier tiers — see claude-sonnet |
| Fact | GPT-5.5 | For comparison |
|---|---|---|
| Context window | ~1,050,000 tokens | Claude Opus 4.8 / Sonnet 4.6: 1M |
| Standard input / output | $5 / $30 per 1M | Opus 4.8: $5 / $25 · Sonnet 4.6: $3 / $15 |
| Pro tier | $30 / $180 per 1M | — |
| Cached input | $0.50 / 1M | — |
| Long-context surcharge | > 272K input → 2× in / 1.5× out (full session) | — |
| SWE-bench Verified | #1 (88.7%) | Opus 4.7: 87.6% (margin within noise + contamination caveats) |
| SWE-bench Pro | 58.6% | Opus 4.7 leads (64.3%) |
| Terminal-Bench 2.0 | 82.7% (leads) | — |
On subscription plans (not per-token API), the per-token USD above is notional, not the billed cost — do not route on headline price alone when billing is a flat subscription.
Strengths
Weaknesses / watch-outs
Use this checklist to confirm a GPT-vs-Claude routing decision is correct and current.
Instead of gpt-5-5 | Use | Why |
|---|---|---|
| Running GPT through its CLI agent (resume sessions, exec) | codex | That is a harness decision, not a model-routing one |
| Choosing among the Claude tiers (Opus vs Sonnet vs Haiku) | claude-opus / claude-sonnet / claude-haiku | Intra-Claude routing, not GPT-vs-Claude |
| Considering the Google frontier lane | gemini-pro | A different model family's routing |
| Dispatching among local skills at request time | skill-router | A different layer entirely |
references/model-facts.md — verified GPT-5.5 capability, pricing, and benchmark facts (2026-06-08) with sourcesclaude-opus, claude-sonnet, gemini-pro — the other frontier lanes this routing decision weighs against; codex — the harness that runs GPT~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.