Claude Sonnet 5 Makes Near-Flagship AI the New Default

What Happened

On June 30, Anthropic released Claude Sonnet 5, describing it as its "most agentic Sonnet model yet." The new model plans multi-step work, drives tools like browsers and terminals, and runs autonomously at a level that, only a few months ago, required larger and more expensive models. And unusually for a capability jump of this size, it arrives not as a premium tier but as the everyday default: Sonnet 5 is now the standard model for Free and Pro plans on Claude, is available to Max, Team, and Enterprise users, and ships in Claude Code and on the Claude Platform under the API name claude-sonnet-5.

The headline claim is that Sonnet 5's performance is close to that of Opus 4.8, Anthropic's flagship, at meaningfully lower prices. Against its direct predecessor, Sonnet 4.6, the company reports substantial improvements across reasoning, tool use, coding, and knowledge work. On SWE-bench Pro, a demanding agentic coding benchmark, Sonnet 5 scores 63.2 percent versus 58.1 percent for Sonnet 4.6 — still behind Opus 4.8 at 69.2 percent, but a clear step toward it, as IT Pro noted in its launch coverage.

The commercial terms are aggressive. Sonnet 5 launches with introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which it moves to $3 and $15. Opus 4.8, for comparison, is priced at $5 and $25. Anthropic also raised rate limits across its chat, Cowork, Claude Code, and platform products to absorb the heavier token usage that comes with higher effort settings.

Why It Matters

Sonnet is Anthropic's volume tier — the model that the overwhelming majority of its users actually touch. Moving near-flagship agentic capability into that slot quietly resets the baseline for what a free chat tab can do: not just answer questions, but plan a task, operate a browser, run code, and check its own output before handing it back. In a market where rivals have recently gated their newest systems behind partner programs, making this class of model the default is itself a strategic statement.

There is also a neat historical symmetry. For many developers the agentic era effectively began with Sonnet-class models — Claude Sonnet 3.5, 3.6, and 3.7 were the first widely used models to show convincing skills in coding and tool use. In the past year, though, the clearest agentic gains migrated up to the Opus class, leaving Sonnet as the sensible-but-limited middle child. Sonnet 5 is Anthropic's attempt to pull that frontier back down into the workhorse tier.

The mechanism for doing so is the effort dial. On agentic search (BrowseComp) and computer-use (OSWorld-Verified) evaluations, Anthropic's cost-performance curves show Sonnet 5 as a strict improvement over Sonnet 4.6 at every setting, with substantially better cost efficiency at medium effort and, at the highest settings, performance that can match Opus 4.8 on some tasks. In effect, one model now spans a range of price-performance points that previously required choosing between two different models.

One caveat belongs on every budget owner's radar: Sonnet 5 uses an updated tokenizer that can map the same input to roughly 1.0 to 1.35 times as many tokens depending on content. The introductory pricing is calibrated so the transition is roughly cost-neutral — which also means the September move to standard pricing is where real-world costs will be felt.

How Developers Are Reacting

Feedback from early-access partners has clustered around a consistent theme: the model finishes jobs that earlier Sonnets abandoned halfway. Testers describe multi-step workflows — updating CRM records and then sending the follow-up announcement, or tracing a bug, writing a reproducing test, and verifying the fix — completed end to end without prodding, and self-review that happens unprompted.

The launch quotes span an unusually wide range of industries. App-building platform Lovable highlighted that the model refuses unsafe requests cleanly while serving millions of builders; database company ClickHouse said tighter reasoning steps get its users to answers noticeably faster; legal AI startup Eve called it the new price-performance frontier for plaintiff-law research; and insurance automation firm Pace reported reliable computer-use runs across intake and claims workflows. Coding-tool makers echoed the pattern, with agents staying on plan through messy, multi-step changes at a cost that makes migration easy to justify.

The measured notes are worth keeping alongside the enthusiasm. Early third-party analysis, including DataCamp's review, places Sonnet 5 clearly above Sonnet 4.6 but still generally below Opus 4.8 on the hardest accuracy-sensitive tasks. Anthropic's own behavioral audit likewise shows Sonnet 5 improving on its predecessor while still registering somewhat higher rates of misaligned behavior than the flagship models — better, not yet equal.

What Comes Next

The first date to watch is August 31, when the introductory pricing lapses and Sonnet 5 moves to $3 per million input tokens and $15 per million output. That is the point at which the cost-performance story faces its real test: teams that migrated during the launch window will decide whether the gains justify the standard rate, especially with the new tokenizer inflating token counts on some content types.

The competitive backdrop makes the timing pointed. OpenAI has so far limited general access to its newest GPT-5.6 tier, keeping the Sol, Terra, and Luna models with a group of roughly twenty enterprise partners, with wider availability expected over the summer. Anthropic pushing near-flagship capability to every free user is close to the opposite posture, and the contrast will not be lost on developers deciding where to build.

Google, meanwhile, shipped two new Gemini image-generation models on the same day Sonnet 5 launched, while general availability of Gemini 3.5 Pro has slipped toward mid-July amid enterprise feedback about token consumption in long agentic tasks. All three labs are converging on the same battleground: not peak capability, but the cost of sustained, multi-step autonomous work.

For working teams, the practical question is shifting from which model is smartest to which effort level is worth paying for on which task — a procurement question that barely existed a year ago and will likely define the mid-tier market through the rest of 2026.

Closing Thoughts

Model launches arrive weekly, but defaults change rarely, and defaults are where the industry's center of gravity actually lives. When the standard model handed to every new user can plan, browse, execute, and verify, "chatbot" stops being the right mental model for what a free AI account is. Sonnet 5's real significance may be less any single benchmark than this quiet redefinition of the starting point.

The safety posture is part of the product story too. Anthropic says it deliberately did not train Sonnet 5 on offensive cybersecurity tasks, launched it with real-time cyber safeguards switched on by default, and measured lower hallucination and sycophancy rates than its predecessor alongside stronger resistance to prompt-injection attacks. Whatever one makes of any individual claim, the pattern is notable: safety characteristics are being presented as spec-sheet features of a mass-market default, not as footnotes. As autonomous capability becomes the baseline rather than the upgrade, that is exactly where the scrutiny belongs. Full details are in Anthropic's announcement and the accompanying system card.

한글 요약

앤트로픽이 6월 30일 새 AI 모델 '클로드 소네트 5(Claude Sonnet 5)'를 출시했습니다. 회사가 "가장 에이전트적인 소네트"라고 소개한 이 모델은 브라우저와 터미널 같은 도구를 스스로 조작하며 여러 단계의 작업을 자율적으로 수행하고, 무료·프로 요금제의 기본 모델로 즉시 적용됐습니다. 에이전트 코딩 벤치마크 SWE-bench Pro에서 63.2%를 기록해 전작 소네트 4.6(58.1%)을 크게 앞섰고, 플래그십인 오푸스 4.8(69.2%)과의 격차도 좁혔습니다.

가격 정책이 특히 공격적입니다. 8월 31일까지 API 기준 입력 100만 토큰당 2달러, 출력 10달러의 런칭 가격이 적용되며 이후 3달러·15달러로 전환됩니다. '노력(effort) 수준'을 조절해 한 모델 안에서 비용과 성능의 균형점을 고를 수 있는 것이 핵심 특징으로, 중간 설정에서는 비용 효율이 크게 개선되고 최고 설정에서는 일부 작업에서 오푸스 4.8급 성능을 냅니다. 다만 새 토크나이저로 같은 입력이 최대 1.35배 많은 토큰으로 계산될 수 있어, 9월 정가 전환 시점의 실질 비용은 지켜볼 필요가 있습니다.

업계 맥락도 흥미롭습니다. OpenAI가 최신 GPT-5.6 상위 모델을 약 20개 파트너에게만 제한적으로 제공하는 것과 달리, 앤트로픽은 준플래그십급 성능을 모든 무료 사용자에게 기본으로 열었습니다. 구글도 같은 날 새 이미지 생성 모델을 내놓는 등 중간 가격대 모델 시장의 경쟁이 한층 뜨거워지는 모습입니다. 사이버 보안 작업 능력을 의도적으로 제한하고 실시간 안전장치를 기본 적용하는 등, 안전 설계를 제품 사양처럼 전면에 내세운 점도 이번 출시의 특징입니다.