Claude Sonnet 5

Claude Sonnet 5 vs Sonnet 4.6: What Changed

Anthropic describes Sonnet 5 as a drop-in upgrade for tool definitions and response shapes, but its behavior changes can affect request validity, token budgets, output limits, and cost. Remove unsupported parameters and recount real prompts against the new tokenizer.

New Coding Capabilities

Anthropic positions Sonnet 5 as its most agentic Sonnet model at launch, with improvements in planning, coding, tool use, and long-running work. The model can operate with browsers, terminals, repository tools, and other agent infrastructure.

For a coding migration, test behaviors rather than general impressions:

• repository exploration before editing;
• multi-file interface changes;
• tool choice and argument validity;
• recovery from failed tests or commands;
• diff focus and avoidance of unrelated edits;
• long-task completion under compaction;
• prompt-injection resistance when tools retrieve external content.

The one-million-token context can hold large inputs, but repository retrieval still matters. Recounting is especially important because Anthropic says Sonnet 5's tokenizer produces more tokens for the same text than Sonnet 4.6.

Benchmark Comparison

Anthropic's launch materials show Sonnet 5 improving over Sonnet 4.6 on coding and agentic evaluations and approaching Opus-tier performance in some configurations. Those results use defined effort levels, tool budgets, and evaluation harnesses.

Avoid transferring a launch chart directly to production. Build a migration suite with:

1. a short deterministic bug fix;
2. a multi-file feature with tests;
3. a tool-heavy task with one forced failure;
4. a long-context repository question;
5. a prompt-injection test for retrieved content.

Compare accepted-task rate, token use under the new tokenizer, tool calls, wall time, and human repair effort. If Sonnet 5 produces more tokens but reduces retries, the total task can still be economical.

For latency-sensitive worker tasks, compare Claude Haiku 4.5 rather than routing every task to Sonnet.

Pricing & Access

Anthropic prices Sonnet 5 at $2 per million input tokens and $10 per million output tokens. On August 10, 2026, Anthropic made those launch rates permanent and withdrew the previously announced September increase. Record the billing date and source when preparing a production budget.

Sonnet 5 is available through the Claude API, Claude applications, Claude Code, and supported cloud platforms. The API alias is claude-sonnet-5; Anthropic also documents a dated snapshot for version stability.

Budgeting should include:

• higher token counts from the new tokenizer;
• adaptive-thinking output within max_tokens;
• prompt-cache write and read behavior;
• retries caused by unsupported parameters;
• tool and agent costs outside the base model.

Recheck official pricing before production budgeting because provider and platform rates can still change.

Using Claude Sonnet 5 in Verdent

Verdent's public model-pricing page does not currently list Claude Sonnet 5. Availability may differ by account, region, or rollout, so the active model selector is the product-level source of truth. Verdent supports Anthropic BYOK; Sonnet 5 can be evaluated only if it appears in the provider model list available to the account.

A controlled migration is:

1. add or validate an Anthropic BYOK key;
2. confirm claude-sonnet-5 appears in the model selector;
3. remove unsupported sampling and manual-thinking settings from the test request;
4. run the same plan, build, review, and test workflow used with Sonnet 4.6;
5. compare token counts, repair loops, and accepted-task cost;
6. change defaults only after the canary passes.

Verdent allows validated BYOK models to be assigned to presets and specialist agent roles. Smart Suggestions and automatic context compression remain unavailable for provider-key calls, a limitation that matters during long repository sessions.

See Anthropic's Sonnet 5 migration changes for the current request-compatibility checklist.