Claude Sonnet 5 is the new default: the migration checklist I run before swapping claude-sonnet-4-6 in production

Introduction

While Fable 5 dominated headlines in June, Sonnet 5 quietly became the model most production agents will actually run on in July. Anthropic positioned it as the most agentic Sonnet yet: multi-step planning, browser and terminal use, and autonomous work at a level that used to require larger models. For builders still pinned to claude-sonnet-4-6, the swap is not a drop-in rename. Three API behavior changes and a tokenizer shift will move your cost and eval curves.

I wrote Gemini 3.5 Flash vs Sonnet 4.6 routing when Sonnet 4.6 was the mid-tier default. This post is the Sonnet 5 migration pass: what breaks, what improves, and what to measure before you flip routing configs while Fable 5 returns to production.

What Sonnet 5 changes on the API (June 30 release notes)

• Model ID: claude-sonnet-5 on Platform API, Claude Code, Cowork, and chat defaults.
• Intro pricing $2/$10 per million through August 31, 2026; standard $3/$15 after.
• 1M token context window and 128k max output tokens.
• Adaptive thinking on by default (no manual extended thinking budgets).
• Non-default temperature, top_p, top_k return HTTP 400.
• New tokenizer: ~30% more tokens for the same text vs pre-Opus 4.7 tokenizers.
• Priority Tier not available on Sonnet 5 (plan latency expectations accordingly).

The migration checklist I run on every engagement

1. 01
   Grep and replace model IDs in config, not prompts

   Search env vars, Helm values, Agent SDK frontmatter, and CI secrets for claude-sonnet-4-6. Pin SONNET_MODEL and SONNET_MODEL_FALLBACK explicitly. Same discipline as the June 15 retirement checklist.
2. 02
   Remove manual extended thinking parameters

   Any request sending thinking type enabled with budget_tokens will 400. If you relied on budget caps to limit reasoning spend, replace with max_tokens and cost-per-step telemetry.
3. 03
   Strip non-default sampling parameters

   temperature, top_p, top_k must be default or omitted. Teams that tuned temperature for JSON reliability need to fix that in prompts and schemas instead.
4. 04
   Re-count tokens and re-benchmark cost per completed task

   Use the token counting API with model claude-sonnet-5 on representative prompts. A 30% tokenizer inflation invalidates last month's routing spreadsheet.
5. 05
   Re-run tool-call and eval suites

   Sonnet 5 scores higher on agentic benchmarks, but your registry descriptions drive selection accuracy. Run the eval datasets beyond happy path suite before promotion.

Where Sonnet 5 fits in your routing layer after Fable returns

With Fable 5 back on the Platform from July 1, most teams now run a three-tier table: Haiku for routing, Sonnet 5 for production agent steps, Fable or Opus for the hardest 5% of tasks. Sonnet 5 at intro pricing may beat Opus 4.8 fast mode on cost per completed task for mid-horizon coding even when Fable is available on subscription inclusion through July 7.

Do not promote Sonnet 5 because the benchmark blog said so. Promote it when your eval suite and cost-per-task metric beat Sonnet 4.6 on the workflows you actually ship. The stop paying frontier prices note still applies: route by task, not by leaderboard.

Managed Agents and platform features to retest

June 30 also shipped Managed Agents session overrides (agent_with_overrides for per-session model swaps), event deltas on session streams, and vault injection_location for egress credentials. If you run Managed Agents on Sonnet 4.6 today, test overrides with Sonnet 5 in a staging deployment before changing the default agent version.

Wire model_id as a required dimension in the agent observability stack so you can compare Sonnet 4.6 and Sonnet 5 traces side by side during the migration window.

Common Sonnet 5 migration mistakes

• Renaming the model ID without removing thinking budget params (immediate 400s).
• Using last month's token budget spreadsheets without re-counting.
• Assuming Sonnet 5 intro pricing lasts past August 31 without a fallback plan.
• Skipping tool evals because "it's still Sonnet."
• Routing everything to Fable because it is back, ignoring Sonnet 5 economics on short-horizon tasks.

Conclusion

Sonnet 5 is the default mid-tier model for 2026 H2. The capability bump is real. The breaking changes are also real: adaptive thinking, stricter sampling, tokenizer inflation. Grep configs, re-count tokens, re-run evals, then promote. Treat Sonnet 4.6 retirement the same way you treated June 15 model IDs: a scheduled migration, not a fire drill.

Sources: Anthropic Sonnet 5 announcement at https://www.anthropic.com/news/claude-sonnet-5; Platform release notes June 30, 2026 at https://platform.claude.com/docs/en/release-notes/overview.