← Back to feed

Anthropic shipped Claude Sonnet 5 on June 30, positioning it as the most capable mid-tier model it has ever released and immediately deploying it as the default for Free and Pro plan users. The move compresses the performance gap between the Sonnet and Opus tiers that had widened considerably over the past year as Opus 4.7 and Opus 4.8 pulled ahead on agentic capability.

The pricing structure rewards early adopters. Introductory API pricing runs at $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which it steps up to $3/$15 at standard pricing. Opus 4.8 carries $15/$75 pricing. The effective agentic cost advantage at launch is approximately 7.5x on output compared to the flagship — if the performance differential is as narrow as Anthropic claims.

What Changed

Anthropic’s core claim is that Sonnet 5 finishes tasks that Sonnet 4.6 stopped halfway through. That sounds like marketing language, but the partner feedback in the launch post is specific enough to be credible. Lovable reported that a two-part workflow — updating Salesforce account tiers and sending an enterprise launch announcement — completed end-to-end without intervention. A previous Sonnet would stall midway. Eve, which uses Claude for plaintiff-side legal research, described Sonnet 5 as sitting on the Pareto frontier for its task mix.

On agentic search (BrowseComp) and computer use (OSWorld-Verified), Sonnet 5 scores above Sonnet 4.6 at every effort level and approaches Opus 4.8 curves on the same charts. Anthropic publishes these as visual comparisons rather than a single number, so direct benchmarking will have to come from third parties.

The self-correction behavior is documented. Partners describe Sonnet 5 writing a reproducing test, implementing a fix, stashing the fix to verify the bug returns, then committing — all within a single unprompted pass. That kind of chained verification is the behavior that distinguishes production-ready coding agents from demo-ready ones.

Safety Position

The system card draws a clear line between Sonnet 5 and the Opus tier on cyber capability. Both Sonnet 5 and its predecessor scored 0% on successful exploit development — Sonnet 5 showed slightly higher partial success than Sonnet 4.6, but neither crossed into the functional range that triggers Anthropic’s high-risk designation. Mythos 5 and Opus 4.8 are in a different category. The Sonnet tier is deliberately kept well below that threshold.

Automated behavioral audit results show an overall lower rate of undesirable behaviors in Sonnet 5 compared to Sonnet 4.6 — Anthropic’s pre-deployment red-teaming found the model safer to deploy in agentic contexts.

Availability

Claude Sonnet 5 is available across all Claude plans, in Claude Code, and on the Claude API via claude-sonnet-5. Rate limits have been increased across Chat, Cowork, Claude Code, and the API to accommodate the higher token volumes that come with extended effort levels.

Competitive Context

Sonnet 5 arrives as the mid-tier market gets contested. GPT-5.5 Instant is OpenAI’s current default, priced at $5 per million input tokens — 2.5x the introductory Sonnet 5 rate. Gemini 3.5 Flash holds the speed tier. Anthropic’s bet is that agentic follow-through and coding reliability at the mid-tier price point is where enterprise developers are actually making switching decisions. Sonnet 4.6 had the reliability reputation; Sonnet 5 is meant to extend that to full-task completion.

For developers currently paying for Opus 4.8 to run routine agentic workflows, Sonnet 5 at $2/$10 introduces a meaningful re-evaluation point.