Claude Sonnet 5

https://www.anthropic.com/news/claude-sonnet-5

Free for Free/Pro users; $2/$10 per million tokens (API, intro through Aug 31); $3/$15 per million tokens thereafter

Agentic Architecture — Significantly improved planning, tool use, and autonomous multi-step execution compared to Sonnet 4.6; closer to Opus 4.8 on agentic benchmarks
1M Token Context Window — Massive context for long documents, large codebases, and extended reasoning sessions
128K Max Output Tokens — Generous per-turn ceiling for sustained generation tasks
Claude Code Integration — Available directly in Anthropic's CLI coding agent
Claude Platform API — $2/$10 per million tokens intro pricing through Aug 31, 2026; $3/$15 standard thereafter
Improved Safety Evaluations — Lower hallucination and sycophancy rates vs Sonnet 4.6; better refusal of malicious requests
Prompt Injection Resistance — Strengthened defenses against prompt injection attacks in agentic deployments
Browser and Terminal Tool Use — Sustained autonomous use of web browsing and terminal tools across complex workflows
No Temperature/top_p/top_k — Simplified parameter set reflects training approach changes

Most agentic Sonnet ever — meaningful jump over Sonnet 4.6 on planning, tool use, and multi-step reasoning
Opus-level performance at Sonnet prices — narrows the cost-capability gap significantly
Now default for Free and Pro users — broadest possible reach
1M token context window and 128k max output — excellent for long-form agentic sessions
Improved safety profile vs Sonnet 4.6 — lower hallucination, lower sycophancy, better refusal behavior
Same API pricing as Sonnet 4.6 through August 31 (introductory rate)
Available in Claude Code — direct integration for developers
Better at brownfield coding tasks — race conditions, hidden tests, messy production code

New tokenizer produces ~30% more tokens for same text — cost equivalence with Sonnet 4.6 not exact
Temperature/top_p/top_k no longer supported — deliberate simplification some users may find limiting
Safety profile trails Opus 4.8 on behavioral audit — still a consideration for highest-stakes deployments
No specialized fine-tuned variant yet — general-purpose agentic model rather than domain-specific