Claude Sonnet 5 · API · High effort (default)
A polished code smith with expensive habits — still the sensible daily driver.
The default professional coding partner — excellent taste and reliability without always needing flagship spend.
Overall mean of attribute public scores after peer min-max (40–95 band within scored launch peer set).
- Endpoint
- claude-sonnet-5
- Product
- Anthropic API
- Price
- $2 / $10 per MTok
- Context
- 1000k tokens
Claude Sonnet 5 API, default high effort, 1M context, intro $2/$10 (→ $3/$15). Primary gold-sheet daily-driver coding Build.
Six stats · peer public
Peer public (main number / bar) ranks this Build inside the current scored peer set (≈40–95). Raw is the absolute evidence-fusion aggregate before min-max. A low peer score can still be a strong absolute model — see methodology.
When to choose
- Day-to-day software engineering in real repositories
- Greenfield apps with clear product direction
- Collaborative writing, docs, and instruction-heavy workflows
- Balancing quality and cost versus Opus-class flagships
- Cheapest possible high-volume automation (use a sprinter)
- Strict offline / private-local-only requirements
- Maximum autonomous long-horizon agenting at any cost (compare Opus / GPT-high)
- Latency-first chat where flash models dominate
You want one primary Build for coding and writing that stays pleasant under human supervision, holds constraints, and rarely needs the top-tier price.
Task proficiencies
Signature strengths and honest flaws
Holds multi-part requirements, style guides, and earlier decisions across long coding sessions better than many peer Builds — especially when instructions are explicit and structured.
- Multi-file coding sessions with written acceptance criteria
- Refactors with explicit "do not change X" constraints
Quickly builds a useful mental model of unfamiliar codebases — module boundaries, conventions, and likely change points — before editing.
- Greenfield exploration of mid-to-large repos
- Onboarding to legacy services with sparse docs
Adapts prose tone and code style to the surrounding project or brief without losing substance — strong default for collaborative writing and polished UI copy.
- Explicit style samples or "match this repo" instructions
- Product copy and documentation alongside code
Introduces extra abstractions, config layers, or "future-proof" structure when a smaller change would satisfy the brief — especially on underspecified tasks.
- Vague product requests without scope caps
- Small bugfixes framed as open-ended improvements
Workaround · State hard limits: max files to touch, "smallest correct change," forbid new dependencies, and require a one-paragraph plan before code.
Explains plans and options at length when a direct edit or shorter answer would move the work forward — raising cost and breaking flow on simple tasks.
- Fast everyday asks without "be brief" constraints
- Multi-turn sessions where prior context already stated the plan
Workaround · Prefer instructions like "code first, ≤5 bullets of notes" or "diff only, no preamble." Cap max output tokens for interactive chat.
Default thoroughness produces longer outputs than flash-class Builds, so token spend climbs quickly on high-volume or multi-file agent runs.
- Autonomous agent loops without output budgets
- Repeated full-file rewrites instead of patches
Workaround · Use for high-value sessions; route bulk/simple work to a sprinter Build. Prefer patch-style edits and explicit token/time budgets in agent configs.
Traceability
Every factual claim cites an evidence row. Tap a chip to see the benchmark, exact config, source, and caveats behind the number.
Engine pass meth_v0_1_1 (peer min-max vs scored launch peer set). Codex overall 50.8 (A). Attribute public scores map raw aggregates into a 40–95 band within the peer set (methodology min-max). INT uses hierarchical fusion (GPQA primary / HLE secondary / capped bleed).