Claude Sonnet 5: Agentic Pricing Guide [2026]
Bottom Line
Sonnet 5 is the new default Free/Pro model and is positioned as nearly Opus-class for many agentic tasks at lower price — but Microsoft’s agent-eval work shows cheaper rate cards can still raise total cost if token volume explodes.
Key Takeaways
- ›Sonnet 5 launched June 30, 2026 as the default model for Free and Pro; available on Max/Team/Enterprise, Claude Code, and Claude Platform.
- ›Introductory API pricing: $2/M input and $10/M output through August 31, 2026; then $3/$15.
- ›Anthropic claims Sonnet 5 approaches Opus 4.8 on key agentic evals while reducing undesirable behaviors vs Sonnet 4.6.
- ›Sonnet 5 has much lower cybersecurity task ability than current Opus models — choose model by threat surface, not only coding benchmarks.
- ›Always measure cost per successful task, not only $/M tokens.
Anthropic’s “Introducing Claude Sonnet 5” (June 30, 2026) positions Sonnet 5 as the most agentic Sonnet yet — planning, browsers, terminals, and autonomous runs that recently needed larger models.
Capability positioning
- Substantial improvement over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work
- Performance described as close to Opus 4.8 on important agentic dimensions at lower price
- Safety assessments: overall lower rate of undesirable behaviors than Sonnet 4.6 in agentic contexts
- Much lower ability to perform cybersecurity tasks than current Opus models
Pricing you should hard-code into budgets
| Window | Input / 1M | Output / 1M |
|---|---|---|
| Intro through Aug 31, 2026 | $2 | $10 |
| After Aug 31, 2026 | $3 | $15 |
Availability: Free/Pro default; Max/Team/Enterprise; Claude Code; Claude Platform.
The upgrade trap (read this before flipping defaults)
Microsoft’s agent-experience team ran multi-scenario Copilot Chat evals comparing Claude Sonnet 4.6 vs Sonnet 5 and found that lower per-token prices can still raise total bill when the newer model consumes far more tokens (double-digit multipliers on some architecture tasks). Pair Anthropic’s rate card with your own harness metrics:
- p50/p95 tokens per successful task
- Tool-call count and retry rate
- Human intervention rate
- Latency under your real repo tools
Rollout recipe
- Keep Sonnet 4.6 (or previous default) as a canary control for one week.
- Route 10–20% of agent traffic to Sonnet 5 with identical prompts/tools.
- Promote only if successful-task cost drops or quality rises enough to justify spend.
- Reserve Opus for long-horizon jobs and security-sensitive investigations.