Claude Sonnet 5 is the first model to test before paying Claude Opus API prices for routine coding and agent work. Anthropic launched it on June 30, 2026 with the API model ID claude-sonnet-5, a 1M-token context window, up to 128K output tokens, and introductory API pricing of $2 input / $10 output per 1M tokens through August 31, 2026. Standard pricing becomes $3 / $15 on September 1.

Anthropic says Sonnet 5 is “close to” Opus 4.8 and can match it on some tasks at higher effort. That is vendor positioning, not an independent general ranking. Keep Opus 4.8 for the highest-accuracy review, architecture, and final-arbitration work until your own task results justify moving that boundary.

July API value shift

Low-cost supported-tool/API lane

GLM-5.2

$1.40 in / $4.40 out

1M context

Vendor specs + independent Artificial Analysis signal

Claude cost/performance test

Sonnet 5

$2 / $10 through Aug 31

$3 / $15 after · 1M context

Anthropic vendor evidence; independent normalized result pending

Premium accuracy and arbitration lane

Opus 4.8

$5 in / $25 out

1M context

Vendor specs + independent Artificial Analysis signal

USD per 1 million tokens. Prices, context limits, and evidence provenance are separate decision inputs; this is not a benchmark ranking.

Quick facts

SpecClaude Sonnet 5
Launch dateJune 30, 2026
API model IDclaude-sonnet-5
Context window1M tokens
Maximum output128K tokens
ThinkingAdaptive thinking on by default; control depth with effort
Introductory API price$2 input / $10 output per 1M through August 31, 2026
Standard API price$3 input / $15 output per 1M from September 1, 2026
Best first useCoding agents, tool use, knowledge work, and routine Claude production
Evidence boundaryLaunch comparisons are Anthropic-reported; independent normalized results are pending

The tokenizer changes the real cost

Sonnet 5 uses a new tokenizer. Anthropic’s migration guide says the same text produces approximately 30% more tokens than Sonnet 4.6; the launch footnote gives a content-dependent range of roughly 1.0–1.35×. That means the 1M-token window holds less equivalent text, a prior max_tokens value may truncate an equivalent response, and per-request cost can rise even when the listed per-token rate is unchanged.

The introductory $2/$10 rate is intended to make migration roughly cost-neutral. Do not assume it makes every workload cheaper. Re-run token counting and measure complete agent loops—including thinking, tool results, retries, and output—before changing production budgets.

API migration checklist

Sonnet 5 is close to a drop-in upgrade from Sonnet 4.6, but two changes can return HTTP 400 errors:

  1. Replace manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) with adaptive thinking and the effort parameter. Adaptive thinking is on by default; explicitly disable it when needed.
  2. Remove non-default temperature, top_p, and top_k values.

Also update the model ID, re-count tokens, review max_tokens, and handle cybersecurity refusals returned as HTTP 200 responses with stop_reason: "refusal". Priority Tier is not available for Sonnet 5 at launch.

Where it is available

SurfaceLaunch statusPrimary evidence
Claude.aiDefault for Free and Pro; available to Max, Team, and EnterpriseAnthropic launch
Claude CodeAvailable at launchAnthropic launch
Claude APIAvailable to all customers as claude-sonnet-5Anthropic platform docs
GitHub CopilotGenerally available from June 30GitHub changelog
AWSAmazon Bedrock and Claude Platform on AWSAWS launch post

AWS notes that regional availability varies. Anthropic’s platform documentation also distinguishes the current Claude-in-Bedrock Messages endpoint from legacy Bedrock InvokeModel and Converse paths, so verify the exact integration rather than assuming every older endpoint supports Sonnet 5.

How to read the launch benchmarks

Treat every launch comparison as vendor evidence. Anthropic says Sonnet 5 substantially improves over Sonnet 4.6 in reasoning, coding, tool use, and knowledge work, and that higher-effort Sonnet 5 can match Opus 4.8 on some tasks. Those claims justify a test, not a universal replacement decision.

Anthropic also corrected the BrowseComp cost-performance chart on launch day because the original used a simpler methodology than its standard agentic-search evaluation. Prefer the updated launch chart and Sonnet 5 system card over copied launch-day numbers. Do not combine BrowseComp, OSWorld-Verified, SWE-bench, or unrelated independent indexes into one synthetic score.

Buying decision

  • Start with Sonnet 5 when you want Claude quality without defaulting to Opus pricing.
  • Keep GLM-5.2 as the lower-cost supported-tool/API lane when its tool restrictions and quota mechanics fit.
  • Use Opus 4.8 when a premium second pass changes architecture, correctness, or risk decisions.
  • Compare subscription seats separately from API tokens in the Smart Spend guide.

Sources


Last verified: July 1, 2026. Introductory pricing ends August 31, 2026; availability, limits, and independent evidence can change sooner.