Claude Sonnet 5

Anthropic released Claude Sonnet 5 at the end of June 2026, and the pitch is straightforward: take the coding and agentic capability that used to require the Opus tier and deliver most of it at Sonnet prices. Anthropic describes it as its most agentic Sonnet model yet, with meaningful gains in reasoning, tool use, and autonomous multi-step work. For the large population of developers and product teams who found Opus overkill for routine agent loops but Sonnet 4.6 not quite sharp enough, Sonnet 5 is aimed squarely at that gap, and on the evidence so far it lands in it.

The headline numbers are concrete. Sonnet 5 ships with a 1M-token context window and up to 128K tokens of output per request. Adaptive thinking — the model deciding for itself how much to reason before answering — is on by default, a change from Sonnet 4.6, where you had to opt in. It is also the first model in the Sonnet tier to support the “xhigh” effort setting, previously an Opus-only control, which lets you push the model to reason harder on the most demanding coding and agentic tasks. Pricing sits at the familiar Sonnet sticker of $3 per million input tokens and $15 per million output, with an introductory rate of $2/$10 running through August 31, 2026. Anthropic has made it the default model in Claude Code, which is a meaningful signal about where the company thinks its everyday coding workload belongs.

The capability story is real, but there is a cost wrinkle worth stating plainly because it is easy to miss. Sonnet 5 uses a new tokenizer that counts roughly 30% more tokens for the same text than Sonnet 4.6 did. The per-token price hasn’t changed, but the same prompt and the same answer now consume more tokens, so the actual bill for an equivalent request rises even though nothing on the pricing page moved. Teams migrating from Sonnet 4.6 should re-measure their token usage rather than assume flat costs — a workload budgeted against the old tokenizer can run over without any change in behaviour. It is not a gotcha so much as an accounting detail that the marketing understandably doesn’t lead with.

Set against the models PixlRun already tracks, Sonnet 5’s position is clarifying. The comparison most users will make is to the flagship chat models — GPT from OpenAI and Gemini from Google — and to Anthropic’s own Claude and Opus lines. Against the flagships, Sonnet 5 competes on a specific axis: agentic coding and tool use, where it now trades blows with far more expensive models rather than sitting a tier below them. Against Opus, it is the deliberate value play — you give up the last increment of raw reasoning on the hardest problems in exchange for a materially lower price, and for most agent loops, code-review passes, and multi-step tool workflows that trade is easy to accept. Against Sonnet 4.6, it is a clear upgrade on coding and autonomy, with the tokenizer caveat as the only real friction.

The vision improvements deserve a mention because they are the quiet kind of upgrade that matters more than it sounds. Sonnet 5 is the first Sonnet-tier model with high-resolution image support, accepting images up to 2576 pixels on the long edge, up from 1568 on Sonnet 4.6. For anyone doing screenshot understanding, document analysis, chart transcription, or computer-use style automation, that extra fidelity is the difference between a model that can read a dense dashboard and one that squints. Full-resolution images cost more image tokens, so the same cost-awareness applies, but the capability ceiling is genuinely higher.

Where does it fall short? Sonnet 5 is still a Sonnet. On the very hardest, longest-horizon reasoning — the multi-day autonomous runs and the most intricate problems — the Opus tier remains ahead, and Anthropic doesn’t pretend otherwise. The adaptive-thinking-by-default change is a double edge: it improves answers but adds latency and token spend, and teams that want fast, cheap responses need to set effort to a lower level or disable thinking explicitly rather than accepting the default. And the tokenizer shift, while not a capability problem, is a real budgeting one that will surprise anyone who ports a cost model over unchanged.

The verdict: Claude Sonnet 5 is the new sensible default for serious coding and agentic work that doesn’t justify Opus pricing, and Anthropic making it the Claude Code default is a fair endorsement of exactly that use case. It closes most of the gap to the top tier on the tasks that matter to builders — tool use, code, autonomous loops — while staying in the Sonnet price band. Deduct a little for the tokenizer-driven cost creep and the need to manage the new thinking defaults, and it still sits near the top of our scale. For most teams currently on Sonnet 4.6, or paying for Opus on workloads that don’t strictly need it, this is the model to evaluate first.