Claude Sonnet 5: Near-Opus Agentic Coding

Claude Sonnet 5 shipped on June 30, 2026 as the most agentic model in the Sonnet line — a model Anthropic describes as able to make plans, use tools like browsers and terminals, and run autonomously at a level that, only a few months ago, took a larger and more expensive model. The headline is not a single benchmark. It is that a Sonnet-tier model now does most agentic work within a few points of the Opus 4.8 flagship, at a fraction of the price.

That matters because the constraint on putting agents into production is rarely raw capability — it is cost at volume. An agentic loop that drives a screen or a terminal burns tokens on every step, so the model you can afford to run thousands of times a month is usually the one that ships. Sonnet 5 is Anthropic’s bet that near-flagship quality at Sonnet prices is what moves agentic automation from demo to deployment.

This guide covers exactly what launched and where to use it: an honest read of Anthropic’s benchmark chart (every figure is first-party), the specific gaps where Opus 4.8 still leads, the pricing against both Sonnet 4.6 and Opus 4.8, the API changes that affect how you build, and a routing framework for deciding which model each step of your pipeline should call.

Sonnet 5 is available immediately and broadly. It is the default model on Anthropic’s Free and Pro plans, available to Max, Team, and Enterprise users, and live across the Claude apps, Claude Code, and the Claude Platform. On the API you call it with the model ID claude-sonnet-5 — a clean string with no date suffix — and it carries a 1M-token context window with up to 128K output tokens.

The framing Anthropic chose is “most agentic Sonnet yet,” and the substance behind that phrase is what the model does between the question and the answer. It makes a plan, calls tools — a browser, a terminal, a search — and runs the loop autonomously instead of handing control back after one step. Two behaviors its early-access testers singled out are the tell: it finishes complex tasks where previous Sonnet models would stop short, and it checks its own output without being explicitly asked. Self-verification is the difference between an agent you can leave running and one you have to babysit.

Read against the rest of the Sonnet line, this is a step change rather than a point release. Our Claude Sonnet 4.6 guide covered a capable but clearly mid-tier model; Sonnet 5 moves the tier’s ceiling up to brush against the flagship. And for anyone who followed the speculation, this is the model behind the “Fennec” leak we analyzed in January — now shipped, named, and benchmarked.

Anthropic published Sonnet 5 against Sonnet 4.6, with Opus 4.8 in the chart as a reference. One caveat governs every figure below, so state it first: these are first-party numbers from Anthropic’s release, not results from an independent evaluator. Read the table as vendor claims arranged on a common axis — useful for direction, not a substitute for evaluating the model on your own tasks.

The shape of the table is the story. The jump over Sonnet 4.6 is large everywhere — and largest where it counts for agents. Terminal-Bench, which measures real terminal work, climbs from 67.0% to 80.4%. Humanity’s Last Exam with tools rises from 46.8% to 57.4%. And on knowledge work — GDPval-AA v2, Anthropic’s economic-value benchmark — Sonnet 5 leaps 223 Elo points, from 1395 to 1618, which is where it crosses into Opus territory outright.

“Close to Opus 4.8” is the honest summary, but the gap is not uniform — and where it opens up tells you exactly when to reach for the flagship. The pattern is clean: when tools are in the loop, Sonnet 5 is within a point or two of Opus 4.8; when the task is pure reasoning with nothing to lean on, Opus pulls ahead by roughly six points. The bars below put Sonnet 5’s five percentage benchmarks against the Opus 4.8 reference.

The most striking pair is Humanity’s Last Exam with and without tools. With tools, Sonnet 5 (57.4%) is half a point behind Opus 4.8 (57.9%) — a statistical tie. Without tools, the gap widens to 6.6 points (43.2% vs 49.8%). That is the whole thesis of the release in one comparison: hand Sonnet 5 the ability to search, run code, and check itself, and it reasons like the flagship; take the tools away and the flagship’s deeper unaided reasoning shows.

SWE-bench Pro tells the same story from the coding side. The 6-point gap (63.2% vs 69.2%) is real, and on the hardest end-to-end software tasks Opus 4.8 is still the model that clears bars Sonnet 5 does not. But Terminal-Bench — closer to how an agentic coding tool actually operates, step by step in a real shell — is a near tie. For most production agents, which work with tools rather than in a single unaided pass, Sonnet 5 captures the part of the flagship that the workload actually uses. Opus 4.8 earns its premium on the tasks at the very top of the difficulty range, not on the average one.

With the benchmarks close, price is the lever — and it is where Sonnet 5 makes its case. The list rate is $3 per million input tokens and $15 per million output, with introductory pricing of $2 / $10 through August 31, 2026. The standard rate matches Sonnet 4.6 exactly, so once the intro window closes you are paying the same list price as the predecessor for a large step up in capability. The comparison that matters, though, is against Opus 4.8.

One honest hedge keeps the cost comparison clean. Sonnet 5 uses a new tokenizer that turns the same text into roughly 30% more tokens than Sonnet 4.6 did, so a per-token price that looks identical does not translate to an identical bill on the same workload — re-measure with real token counts before you forecast. Even after that adjustment, the structural point holds: a model that performs within a couple of points of the flagship at 40-60% of its price changes which automations clear the bar of being worth building. That arithmetic, not a hero benchmark, is what tends to decide a model-routing decision.

If you are integrating Sonnet 5 rather than chatting with it, three things behave differently from Sonnet 4.6. None is hard, but each can quietly change cost or output if you carry a 4.6 configuration over unchanged.

The effort dial is the lever most worth tuning. A rough cross-model mapping from Anthropic’s own guidance: Sonnet 5 at medium is comparable in intelligence to Sonnet 4.6 at high, and Sonnet 5 at high is comparable to Sonnet 4.6 at max. In practice that means you can often hold quality steady and step effort down a notch to cut latency and tokens, or hold effort and bank the capability gain. For the hardest agentic coding, xhigh is the new ceiling. Set it deliberately rather than inheriting the 4.6 default.

The benchmark gaps above translate into a simple routing rule: default to Sonnet 5 for the volume of agentic work, and escalate to Opus 4.8 only at the top of the difficulty range. The cleanest pipelines run a Sonnet 5 floor with an Opus 4.8 ceiling, switching models per step rather than per project — the shared tool surface and 1M context make those hand-offs cheap.

For a marketing or operations team, the practical opening is the work that was previously too token-expensive to automate at a frontier model’s price. Agent-written reporting, content pipelines that draft and self-check, lead-routing and CRM hygiene agents, browser automations that read dashboards and flag anomalies — these are high-volume, repetitive tasks where a near-flagship model at 40-60% of the cost changes the build-versus-skip calculation.

The sequencing we use with clients is the same regardless of model: prove value on a single well-scoped, low-risk workflow — usually read-only at first — add human-confirmed writes once the audit trail earns trust, and only then widen autonomy. Sonnet 5’s self-checking and tool fluency make that ramp faster, but the discipline is unchanged: scope tight, log everything, and route the occasional hard step to Opus 4.8 rather than over-paying for the whole pipeline. That scoping — which workflows, which guardrails, which model per step — is exactly where our agentic AI transformation engagements begin, before any model commitment. For teams rolling agents out at scale, our enterprise deployment playbook covers the governance layer.