Embed badge
Show
Style
[](https://versuz.dev/skills/seb155-atlas-plugin-skills-refs-model-benchmarks-2026-04)Free SKILL.md scraped from GitHub. Clone the repo or copy the file directly into your Claude Code skills directory.
npx versuz@latest install seb155-atlas-plugin-skills-refs-model-benchmarks-2026-04git clone https://github.com/seb155/atlas-plugin.gitcp atlas-plugin/SKILL.MD ~/.claude/skills/seb155-atlas-plugin-skills-refs-model-benchmarks-2026-04/SKILL.md--- name: model-benchmarks-2026-04 description: "Claude Opus 4.7 vs Sonnet 4.6 benchmark comparison, decision framework, pricing, and prompt caching strategy. Reference for ATLAS model allocation." mode: [engineering] --- # Claude Model Benchmarks — April 2026 > Reference for ATLAS model allocation decisions. > SSoT for task-to-model routing: `skills/execution-strategy/model-rules.yaml` > Last updated: 2026-04-23 (P1 doc pass — [1m] variant mandate clarified) ## 🔴 CRITICAL — Opus [1m] Variant Mandate **Always use `claude-opus-4-7[1m]` for Opus allocations, never shortcut `opus`.** The shortcut `model: opus` in AGENT.md / skill frontmatter **silently resolves to the 256K context variant** (not the 1M variant) in some Claude Code configurations. For long sessions, multi-plan work, or anything requiring > 200K context, this is a silent capability regression. **Correct**: ```yaml # AGENT.md / SKILL.md frontmatter model: claude-opus-4-7[1m] # ✅ explicit 1M context variant effort: max | xhigh ``` **Incorrect (silent 256K fallback risk)**: ```yaml model: opus # ❌ may resolve to 256K variant ``` **Model IDs canonical list**: | Intended | Variant | Correct ID | Context | |----------|---------|-----------|---------| | Opus (deep reasoning) | 1M | `claude-opus-4-7[1m]` | 1M tokens | | Opus (default) | 256K | `claude-opus-4-7` | 256K tokens (NOT for ATLAS) | | Sonnet | — | `claude-sonnet-4-6` | 200K tokens | | Haiku | — | `claude-haiku-4-5-20251001` | 200K tokens | **Enforcement layers** (v6.0.0-alpha.5+): - L1 (type): `.atlas/models.yaml` registry (planned v6.1) - L3 (semgrep): `.semgrep/atlas-opus-1m-only.yaml` (planned v6.1) - L5 (CI gate): blocking if non-[1m] opus detected (planned v6.1) - L7 (doc): THIS DOCUMENT + `~/.claude/CLAUDE.md` critical reminder **History**: Sprint 0 P0 fixes (2026-04-23 / v6.0.0-alpha.5) fixed 3 AGENT.md that had `model: opus` shortcut: - `agents/code-reviewer/AGENT.md` - `agents/infra-expert/AGENT.md` - `agents/plan-architect/AGENT.md` All now explicitly use `claude-opus-4-7[1m]`. --- ## Model Comparison > **Note**: SWE-bench, GPQA, and GDPval figures below are **as of Opus 4.6** (April 2026 snapshot). > Opus 4.7 launched 2026-04-16 with same pricing; official 4.7 benchmarks pending dedicated update. > Tokenizer change: Opus 4.7 may produce up to +35% tokens for same text vs 4.6 (effective cost impact). | Metric | Opus 4.7 | Sonnet 4.6 | Haiku 4.5 | Notes | |--------|----------|-----------|-----------|-------| | **SWE-bench Verified** | 80.8% (4.6 data) | 79.6% | ~50% | Gap: 1.2pts (smallest in Claude history, 4.6 data) | | **GPQA Diamond** | 91.3% (4.6 data) | 74.1% | ~40% | Gap: 17.2pts (PhD-level science, 4.6 data) | | **GDPval-AA** | 1606 Elo (4.6 data) | **1633 Elo** | N/A | Sonnet LEADS on practical tasks (4.6 era) | | **Context window** | 1M tokens | 1M tokens | 200K | Both top models = 1M | | **Max output** | 128K tokens | 64K tokens | 8K | Opus 2x for long plans | | **Speed** | ~20-30 tok/s | ~40-60 tok/s | ~80-100 tok/s | Sonnet 2.7x faster | | **Input price** | $5/MTok | $3/MTok | $0.25/MTok | Opus 1.7x Sonnet | | **Output price** | $25/MTok | $15/MTok | $1.25/MTok | Opus 1.7x Sonnet | | **Cache read** | $0.50/MTok | $0.30/MTok | $0.025/MTok | 90% discount on cached | | **Effort levels** | low/medium/high/xhigh/max | low/medium/high | low/medium/high | xhigh = Opus 4.7 exclusive (v2.1.111+) | ## Decision Framework ``` TASK REQUIRES... → MODEL ───────────────────────────────────────────────── Architecture, multi-file design → Opus 4.7 Adaptive thinking, ultrathink → Opus 4.7 (xhigh/max effort) Planning (>50h effort scope) → Opus 4.7 Complex debugging (cross-system) → Opus 4.7 (xhigh) Irreversible decisions → Opus 4.7 ───────────────────────────────────────────────── Implementation (clear spec) → Sonnet 4.6 Code review, security audit → Sonnet 4.6 DB migrations → Sonnet 4.6 Test writing → Sonnet 4.6 Frontend components → Sonnet 4.6 Bug fixes (isolated scope) → Sonnet 4.6 Refactor (<5 files) → Sonnet 4.6 ───────────────────────────────────────────────── Checklist validation → Haiku 4.5 Simple search/grep → Haiku 4.5 Status checks → Haiku 4.5 ───────────────────────────────────────────────── Lint, format, type-check → DET (bash) ``` ## Orchestrator Pattern (Cursor-Validated) Based on Cursor's production experience (30% PRs from cloud agents, April 2026): 1. **Main session = Opus 4.7 [1m]** — holds full project context, makes orchestration decisions 2. **Subagents = Sonnet 4.6** — scoped tasks with distilled context from orchestrator 3. **Validators = Haiku 4.5** — cheapest for pass/fail checks 4. **Deterministic ops = DET** — bash commands, zero AI tokens **Why Opus for orchestration (not just context)**: - Both models support 1M tokens — context is NOT the differentiator - Opus's reasoning advantage means better task decomposition decisions - Fewer compactions = higher fidelity orchestration over long sessions - Orchestrator distills relevant context → subagents get focused, smaller prompts ## Prompt Caching Strategy Structure subagent prompts for maximum cache hits: ``` ┌──────────────────────────────────────────┐ │ BLOC 1: Project Context (cacheable) │ ~5K tokens │ - Stack, conventions, rules │ Shared across ALL subagents │ - CLAUDE.md essentials │ Cache: $0.30/MTok (Sonnet) ├──────────────────────────────────────────┤ │ BLOC 2: Skill Instructions (cacheable) │ ~2K tokens │ - Task type specific guidelines │ Shared per skill type │ - Quality criteria │ Cache: $0.30/MTok (Sonnet) ├──────────────────────────────────────────┤ │ BLOC 3: Task Context (unique) │ Variable │ - Specific files to modify │ Unique per task │ - Exact requirements │ Full price: $3/MTok (Sonnet) └──────────────────────────────────────────┘ ``` 5 subagents sharing 7K cacheable preamble = 35K tokens at cache price ($0.01) vs standard ($0.11). **1-hour TTL cache** (Opus 4.7 era, CC 2.1.108+): Set `ENABLE_PROMPT_CACHING_1H=true` env var to extend cache from 5min → 1h. Useful for long-running sessions. ## Cost Scenarios (Opus 4.7 pricing, +35% tokenizer adjustment) | Scenario | Per Session | Per Month (2/day) | Annual | |----------|------------|-------------------|--------| | All Opus 4.7 (naive) | $3-6 | $180-360 | $2.2K-4.3K | | Current ATLAS (mixed) | $1-3 | $60-180 | $720-2.2K | | Optimized (aggressive Sonnet) | $1-2 | $60-120 | $720-1.4K | | + SP-DEDUP + caching | $0.50-1.50 | $30-90 | $360-1.1K | **Savings**: Opus 4.7 pricing ($5/$25) vs 4.6 ($15/$75) = **~66% reduction** at same token volume. Partial offset by tokenizer +35% → **net ~55% cost reduction** expected. ## Sources - Anthropic official: [Claude Opus 4.7 announcement](https://www.anthropic.com/news/claude-opus-4-7), [pricing docs](https://platform.claude.com/docs/en/about-claude/pricing) - [BenchLM pricing April 2026](https://benchlm.ai/blog/posts/claude-api-pricing) - SWE-bench Verified leaderboard (4.6 snapshot) - Cursor internal data (Aman talk, April 2026) - ATLAS execution-strategy production data - CC 2.1.111 changelog (Opus 4.7 xhigh effort level) ## When to Update This Doc - New model release (Opus 4.8, Sonnet 5.0, etc.) - Significant pricing changes (>20% delta) - New benchmarks that change the decision framework - Official Opus 4.7 benchmarks published by Anthropic (follow-up item) - Every 3 months as standard refresh