Anthropic's frontier Sonnet-tier model — top-tier agentic coding and reasoning, 1M context, 128K output.
Claude Sonnet 5, released June 30 2026, is Anthropic's next Sonnet-tier agent model and the default practical upgrade from Sonnet 4.6. It keeps the Sonnet economic lane while adding a native 1M-token context window, 128k synchronous max output, adaptive thinking by default, and stronger coding, tool-use, computer-use, and knowledge-work behavior that Anthropic positions close to Opus 4.8 at a lower price.
For routing teams, Sonnet 5 is not a blind alias swap. The migration has three production-impacting constraints: manual extended-thinking budgets now return 400, non-default sampling parameters return 400, and the new tokenizer can produce about 1.0-1.35x as many tokens for the same text. TheRouter should therefore treat Sonnet 5 as a high-throughput agent route with explicit token recounting, effort controls, and fallback rules rather than a silent replacement for every Sonnet 4.6 workload.
| Type | Rate |
|---|---|
| Input | $2.16 / 1M tokens |
| Output | $10.80 / 1M tokens |
| Release date | 2026-06-30platform.claude.com ↗ | verified |
| Claude API model id | claude-sonnet-5platform.claude.com ↗ | verified |
| Context window | 1,000,000 tokensplatform.claude.com ↗ | verified |
| Max output tokens | 128,000 synchronous; up to 300,000 on Message Batches with output-300k betaplatform.claude.com ↗ | verified |
| Pricing | $2/$10 per MTok through 2026-08-31; $3/$15 per MTok starting 2026-09-01platform.claude.com ↗ | verified |
| Prompt caching | 5m writes $2.50/MTok, 1h writes $4/MTok, cache reads $0.20/MTok during introductory pricing; standard rates rise with $3/$15 pricingplatform.claude.com ↗ | verified |
| Training data cutoff | January 2026platform.claude.com ↗ | verified |
| Reliable knowledge cutoff | January 2026platform.claude.com ↗ | verified |
| Reasoning controls | Adaptive thinking on by default; steer with effort; manual budget_tokens removedplatform.claude.com ↗ | verified |
| Unsupported controls | Non-default temperature, top_p, and top_k return HTTP 400; assistant prefill remains unsupportedplatform.claude.com ↗ | verified |
| Priority Tier | Not available on Claude Sonnet 5 at launchplatform.claude.com ↗ | verified |
| Tokenizer migration | New tokenizer; same text can produce about 1.0-1.35x tokens versus Sonnet 4.6www.anthropic.com ↗ | verified |
| Deployment | Claude API, Claude Code, Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI, Microsoft Foundryplatform.claude.com ↗ | verified |
| License | Anthropic Usage Policy (proprietary, API-only) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
BrowseComp Anthropic reports Sonnet 5 as a strict improvement over Sonnet 4.6 and says higher-effort Sonnet 5 can match Opus 4.8 on some agentic-search tasks; the chart values are published graphically in the announcement/system card rather than as machine-readable numbers in the extracted source. | Not numerically disclosed in extracted launch text | www.anthropic.com ↗ | |
OSWorld-Verified Anthropic compares Sonnet 5 with Sonnet 4.6 and Opus 4.8 on OSWorld-Verified and frames Sonnet 5 as a wider cost-performance frontier for computer-use workflows. Use route-local evals before committing automation traffic. | Strictly improved over Sonnet 4.6 in Anthropic launch comparison | www.anthropic.com ↗ | |
Mozilla Firefox exploit evaluation Anthropic reports neither Sonnet 4.6 nor Sonnet 5 developed a full working exploit in the Firefox 147 evaluation; Sonnet 5 had a slightly higher partial-success rate, and cyber safeguards are enabled by default. | 0.0% full working exploit success% | www.anthropic.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "anthropic/claude-sonnet-5",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Use TheRouter's OpenAI-compatible endpoint for normal text, coding, and document prompts. Omit sampling parameters unless you know they resolve to Anthropic defaults.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"anthropic/claude-sonnet-5","messages":[{"role":"user","content":"Audit this migration plan for hidden production risks."}],"max_tokens":2000}'Recent TheRouter coverage flags the missing Priority Tier lane as the main production routing caveat: use Sonnet 5 for agentic throughput, but keep latency-sensitive SLAs on a model with explicit priority support until measured route data says otherwise.
The launch analysis connects Sonnet 5's agentic gains with its temporary $2/$10 pricing. The operational takeaway is to benchmark during the discount window but model September pricing and tokenizer inflation before promising durable savings.
The migration note maps the three breaking behaviors to router checks: remove manual thinking budgets, remove non-default sampling, and recount tokens under the new tokenizer before shifting production traffic.
It is a model-id migration, but not a no-audit migration. Tool schemas and response shapes are broadly compatible, yet Sonnet 5 changes thinking behavior, rejects non-default sampling parameters, removes manual budget_tokens, and uses a tokenizer that can increase token counts for the same text.
The upstream introductory price is $2 per million input tokens and $10 per million output tokens through August 31, 2026, then $3/$15 from September 1. TheRouter's operational pricing is sourced from standard-models.yaml on the live page; this curated module adds the billing caveat that tokenizer inflation can offset part of the apparent discount.
No. Anthropic's launch release notes state that Sonnet 5 has the same tools and platform features as Sonnet 4.6 except Priority Tier. Latency-sensitive fleets should keep a separate route or fallback until measured TheRouter latency is good enough for the SLA.
No for non-default values. Anthropic says non-default temperature, top_p, and top_k return HTTP 400 on Sonnet 5. Use system instructions, tool schemas, structured outputs, and effort rather than sampling knobs.
Re-run token counting against real production prompts before moving traffic. Anthropic says the same text can map to roughly 1.0-1.35x as many tokens depending on content type, so a workload that looks cheaper per token can still cost more per task.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | platform.claude.com ↗ | 2026-07-28 | verified |
| Claude API model id | platform.claude.com ↗ | 2026-07-28 | verified |
| Context window | platform.claude.com ↗ | 2026-07-28 | verified |
| Max output tokens | platform.claude.com ↗ | 2026-07-28 | verified |
| Pricing | platform.claude.com ↗ | 2026-07-28 | verified |
| Prompt caching | platform.claude.com ↗ | 2026-07-28 | verified |
| Training data cutoff | platform.claude.com ↗ | 2026-07-28 | verified |
| Reliable knowledge cutoff | platform.claude.com ↗ | 2026-07-28 | verified |
| Reasoning controls | platform.claude.com ↗ | 2026-07-28 | verified |
| Unsupported controls | platform.claude.com ↗ | 2026-07-28 | verified |
| Priority Tier | platform.claude.com ↗ | 2026-07-28 | verified |
| Tokenizer migration | www.anthropic.com ↗ | 2026-07-28 | verified |
| Deployment | platform.claude.com ↗ | 2026-07-28 | verified |
| License | — | — | verified |
| BrowseComp | www.anthropic.com ↗ | 2026-07-28 | to verify |
| OSWorld-Verified | www.anthropic.com ↗ | 2026-07-28 | to verify |
| Mozilla Firefox exploit evaluation | www.anthropic.com ↗ | 2026-07-28 | verified |
| Claude Sonnet 5 Skips Priority Tier: What the Service-Tier Gap Means for Latency-Sensitive Routing Teams | TheRouter news ↗ | 2026-07-28 | verified |
| Claude Sonnet 5 Launches as Anthropic's Most Agentic Sonnet: What the Introductory Pricing Window Means for Routing Teams | TheRouter news ↗ | 2026-07-28 | verified |
| Claude Sonnet 5's Three Breaking API Changes: What Every Operator Must Audit Before Migrating | TheRouter news ↗ | 2026-07-28 | verified |
| Is Claude Sonnet 5 a drop-in replacement for Claude Sonnet 4.6? | platform.claude.com ↗ | 2026-07-28 | to verify |
| What does Claude Sonnet 5 cost through TheRouter? | platform.claude.com ↗ | 2026-07-28 | to verify |
| Does Claude Sonnet 5 support Priority Tier? | platform.claude.com ↗ | 2026-07-28 | to verify |
| Should I set temperature on Claude Sonnet 5? | platform.claude.com ↗ | 2026-07-28 | to verify |
| How should routing teams test the new tokenizer? | www.anthropic.com ↗ | 2026-07-28 | to verify |