Claude Sonnet 5: Cheaper Agents, Near-Opus Power (2026)

The headline everyone ran with was "a cheaper way to run agents."¹ The more useful question for anyone actually paying an API bill is: cheaper than what, and is it cheaper at all after August?

In one line: Claude Sonnet 5, released June 30, 2026, is Anthropic's most agentic Sonnet yet — performance close to the flagship Opus 4.8 at a lower price, launching at $2 per million input tokens and $10 per million output tokens through August 31, 2026.²

TL;DR

• What launched: Claude Sonnet 5, the successor to Sonnet 4.6, on June 30, 2026. Anthropic calls it "the most agentic Sonnet model yet," with performance "close to that of Opus 4.8, but at lower prices."²
• Pricing: Introductory $2 / $10 per million input/output tokens through August 31, 2026, then standard $3 / $15. Opus 4.8, for reference, is $5 / $25.²
• The catch: Sonnet 5's standard $3 / $15 is the same rate as Sonnet 4.6 — but a new tokenizer turns the same input into roughly 1.0–1.35× as many tokens, so it is not automatically cheaper than its predecessor once the introductory window closes.²
• How to use it: API model id claude-sonnet-5; default model on Free and Pro; available on Max, Team, Enterprise, in Claude Code, and on the Claude Platform.²³
• The knob that matters: an effort setting lets you dial cost against accuracy, so Sonnet 5 and Opus 4.8 now sit on one continuous cost-performance curve.²
• The under-covered story: the system card shows Sonnet 5 refuses malicious requests far better than 4.6 (92.4% vs 76.6%) — but also over-refuses more benign ones.⁴

What You'll Learn

• What Claude Sonnet 5 is and where it sits in Anthropic's lineup
• Exactly what Sonnet 5 costs, and how that compares to Opus 4.8 and Sonnet 4.6
• Why the new tokenizer means "same price" is not the same as "same bill"
• How Anthropic frames Sonnet 5 vs Opus 4.8 as a single cost-performance curve
• What the system card reveals about agentic safety and cyber capability
• Who should switch to Sonnet 5 today — and who should stay on Opus 4.8

What is Claude Sonnet 5?

Claude Sonnet 5 is Anthropic's mid-tier ("Sonnet"-class) large language model, released on June 30, 2026 as the successor to Sonnet 4.6. It is built for agentic work — planning, tool use, and running browser and terminal workflows autonomously — and Anthropic positions its intelligence as close to the flagship Opus 4.8 at a lower price.²

That positioning is the whole point. For much of the last year, the clearest jumps in agentic ability came from Anthropic's Opus-class models, leaving Sonnet as the value option; by Anthropic's own account, the previous best Sonnet model "fell well short of Opus 4.8." Sonnet 5 is Anthropic's attempt to close that gap: on its own agentic search (BrowseComp) and computer-use (OSWorld-Verified) evaluations, Anthropic reports Sonnet 5 is a "strict improvement" over Sonnet 4.6 at every effort level, and can match Opus 4.8 on some tasks at higher effort.² The full benchmark table lives in the Claude Sonnet 5 system card.⁴

Claude Sonnet 5 pricing: is it actually cheaper?

Here is the pricing, straight from the launch, next to its neighbors:²⁵

Against Opus 4.8, the savings are real and easy to compute: at standard rates Sonnet 5 is 40% cheaper on both input and output ($3 vs $5, $15 vs $25), and during the introductory period it is 60% cheaper ($2 vs $5, $10 vs $25). An output-heavy workload sitting on Opus is a natural candidate to test on Sonnet 5.

Against Sonnet 4.6, the story is subtler — and this is where most launch-day coverage glossed over the fine print.

The tokenizer catch every team should measure

Sonnet 5's standard price ($3 / $15) is identical to Sonnet 4.6's. But identical rates do not guarantee identical bills, because Sonnet 5 ships with an updated tokenizer — the same change Anthropic introduced with Claude Opus 4.7. Anthropic states the same input can map to roughly 1.0–1.35× as many tokens depending on the content, and says it set the introductory pricing so that the switch is "roughly cost-neutral" versus Sonnet 4.6.²

Read that carefully. "Roughly cost-neutral" is a claim about the introductory $2 / $10 window offsetting the extra tokens. Once standard pricing kicks in on September 1, you are paying the same per-token rate as 4.6 on potentially up to a third more tokens. For short prompts the effect is negligible; for token-heavy agents — long contexts, big diffs, verbose tool output — it can quietly erase the savings.

The practical move is boring but important: before you assume Sonnet 5 lowers your Sonnet 4.6 bill, run a representative sample of your real prompts through both and compare dollar totals, not rate cards. The rate card looks flat; your workload decides whether it actually is.

Claude Sonnet 5 vs Opus 4.8: one cost-performance curve

The most useful mental model Anthropic offers is that Sonnet 5 and Opus 4.8 no longer occupy separate tiers so much as different points on a single curve, mediated by an effort parameter. Sonnet 5 "provides substantially improved cost efficiency at medium effort," and "its higher-effort performance can match Opus 4.8 on some tasks."² In other words, you tune effort up when a task needs Opus-grade accuracy and down when throughput and cost matter more.

This mirrors the tiering pattern across the frontier — the same "one flagship, several cheaper tiers" shape as OpenAI's GPT-5.6 family and Anthropic's own Opus 4.8 launch. What Sonnet 5 changes is that the cheaper point on the curve is now genuinely close to the top on agentic tasks, rather than a clear step down.

Anthropic's own numbers give a sense of the gap. On one agentic-coding benchmark, Sonnet 5 scores 63.2%, behind Opus 4.8's 69.2% but well ahead of Sonnet 4.6's 58.1%; on a knowledge-work benchmark, Sonnet 5 slightly outperforms Opus 4.8.²¹ That is the shape of the whole release — near-Opus on the things agents actually do, occasionally ahead, at a lower price.

One caveat worth stating plainly: launch-day third-party leaderboards disagree on other figures — notably the exact SWE-bench Verified and Terminal-Bench scores, reported anywhere from the low-80s to high-80s — so for the authoritative table, treat the system card as the source of record.⁴ As a trajectory reference, Anthropic re-reported Sonnet 4.6 at 78.5% on OSWorld-Verified and 34.6% / 46.8% on Humanity's Last Exam (without / with tools) under updated methodology — the baseline Sonnet 5 improves on.²

What the system card reveals about agentic safety

The part of this launch that got the least attention is the most interesting for anyone deploying autonomous agents. Anthropic's system card reports a genuine safety improvement over Sonnet 4.6 — with one honest trade-off.⁴

On Anthropic's Claude Code evaluation, Sonnet 5 responded safely to 92.37% of malicious requests, a large jump from Sonnet 4.6's 76.60%.⁴ That is the good news. The trade-off: on dual-use and benign requests that the model should help with, Sonnet 5's success rate was 91.55%, below Sonnet 4.6's 97.33% — meaning Sonnet 5 refuses more legitimate security-adjacent work (network reconnaissance, penetration-test analysis) than its predecessor did.⁴ If your team runs legitimate offensive-security tooling through Claude, expect to feel that.

Two more findings matter for agentic deployments:

• Computer use is a wash on safety. On malicious computer-use tasks, Sonnet 5 refused 84.68% of the time — essentially identical to Sonnet 4.6's 84.82%.⁴
• Cyber capability is limited by design. Anthropic says it did not deliberately train Sonnet 5 on cybersecurity tasks. On a vulnerability-reproduction test (CyberGym) it succeeded on a single try 52.7% of the time versus Sonnet 4.6's 65.2%, and on a Firefox 147 exploit-development eval (built with Mozilla; the bugs are patched in Firefox 148) neither Sonnet model produced a full working exploit — both scored 0.0%.²⁴

Because Sonnet 5 is slightly stronger here than 4.6, Anthropic launched it with real-time cyber safeguards on by default, the same ones used for Opus 4.7 and 4.8 — though less restrictive than the broader blocks that shipped with the Mythos-class Fable 5.² For work that needs reduced guardrails, Anthropic still points to Opus 4.8.

Claude Sonnet 5 for agents: who should switch

• Running agents on Opus 4.8 for cost reasons? Test Sonnet 5 now. At 40–60% lower cost and near-Opus agentic performance at higher effort, output-heavy production agents are the obvious win.²
• On Sonnet 4.6 today? Upgrading is low-risk on quality, but measure your token bill — the tokenizer change means "same rate" may not mean "same spend" after August 31.²
• Building coding agents? Sonnet 5's gains are concentrated exactly where agents live: sustained multi-step coding, tool use, and self-checking. If you already wire up an agentic tool-use loop like the Sonnet 4.6 pattern, swapping the model id to claude-sonnet-5 is the whole migration.²
• Doing legitimate security work? Watch the higher benign-refusal rate, and keep Opus 4.8 in reach for tasks that need looser guardrails.⁴
• Tune effort, don't just pick a model. The cost-performance win comes from matching effort to the task, not from defaulting everything to high.²

The Bottom Line

Claude Sonnet 5 does the thing the market has been waiting for: it pushes near-flagship agentic capability down into the mid-tier, at 40–60% less than Opus 4.8. For teams running production agents on the flagship purely to get the quality, it is an easy test and probably an easy win. The asterisk is the tokenizer — against Sonnet 4.6, "same rate" plus "more tokens" is not the same as "cheaper," and the honest answer only shows up when you run your own workload through both. And the quieter story, in the system card, is a model that is measurably safer against misuse but a little quicker to say no to legitimate work. Both are worth knowing before you flip the model id.