Anthropic shipped Claude Sonnet 5 on June 30, 2026, and made it the default model for Free and Pro users the same day. On the company's headline agentic-coding benchmark, Sonnet 5 scores 63.2%, up from Sonnet 4.6's 58.1% and closing most of the gap to Opus 4.8's 69.2%. On Terminal-bench 2.1's agentic coding suite, it jumps from 67% to 80.5%. By the numbers Anthropic wants you to read, the mid-tier is now doing Opus-adjacent work.

The pricing card looks reassuring at a glance. Standard rates hold at $3 per million input tokens and $15 per million output, unchanged from Sonnet 4.6, with an introductory promo of $2 in and $10 out running through August 31, 2026. Context window sits at 1M tokens with a 128k max output. That's the sticker.

The fine print is where it gets interesting.

Anthropic's platform docs disclose a new tokenizer that produces roughly 30% more tokens for the same input text. The 1M window, in practice, holds less text than it did on 4.6, and max_tokens budgets tuned to the older model can silently truncate equivalent output. Artificial Analysis measures the downstream effect: Sonnet 5 spends 40% more output tokens per Intelligence Index task at max effort and takes roughly 3x the agentic turns of Sonnet 4.6 on AA-Briefcase and GDPval-AA. Per-task cost at standard pricing lands at $2.29, about 2x Sonnet 4.6 and roughly 15% above Opus 4.8.

The intelligence-index number itself is 53, trailing GPT-5.5 (xhigh) and Opus 4.8 (max) by two to three points, though Sonnet 5 sits just ahead of Opus 4.8 on AA-Briefcase and GDPval-AA and trails only Fable 5, which isn't yet generally available. TechCrunch notes it slightly outperforms Opus 4.8 on Anthropic's own knowledge-work benchmark and frames agentic capability as table stakes at the mid-price tier now, alongside GPT-5.6 Sol and Gemini 3.5 Flash.

There are harder edges. Manual extended thinking returns a 400 error. Non-default temperature, top_p, or top_k values also 400. Refused requests come back as HTTP 200 with stop_reason: 'refusal'. Cyber safeguards that previously shipped only in Opus 4.7 and 4.8 are on by default. These are hard removals, not warnings, and pipelines built against 4.6 assumptions will need reworking before the promo pricing expires.

The differentiator is no longer whether a mid-tier model can run an agent. It's how many tokens the agent burns getting there.

That's the number to instrument. Teams running Sonnet 5 through orchestration layers like LemonLime, which surface per-task token accounting rather than per-call price, will spot the drift early. Everyone else will find it on the invoice.

Claude Sonnet 5 Lands Near Opus 4.8 on Agents — With a Tokenizer Catch

Sources