--- title: Claude Sonnet 5 is near-Opus performance at half the... description: Anthropic released Claude Sonnet 5, positioning it as near-Opus 4.8 quality at Sonnet pricing ($2/$10 per million tokens introductory, rising to $3/$15 after... canonical: https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Claude Sonnet 5 is near-Opus performance at half the price, with a tokenizer catch | daily.dev og:description: Anthropic released Claude Sonnet 5, positioning it as near-Opus 4.8 quality at Sonnet pricing ($2/$10 per million tokens introductory, rising to $3/$15 after... og:url: https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2 og:image: https://api.daily.dev/og/posts/VpK3CYaz2.png og:image:alt: Claude Sonnet 5 is near-Opus performance at half the price, with a tokenizer catch og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Claude Sonnet 5 is near-Opus performance at half the price, with a tokenizer catch **[Trends](https://daily.dev/sources/trends)** · 5 min read · 5 upvotes · 1 comments ## Summary Anthropic released Claude Sonnet 5, positioning it as near-Opus 4.8 quality at Sonnet pricing ($2/$10 per million tokens introductory, rising to $3/$15 after August 2026). It scores 63.2% on SWE-Bench Pro versus Opus 4.8's 69.2%, and is already live on Claude, GitHub Copilot, Cursor, Devin, and other platforms. A key catch flagged by Simon Willison: a new tokenizer makes it ~1.4x more expensive per English character than the raw token price implies, narrowing the actual cost savings. The model is primarily positioned for agentic use cases — planning, spawning sub-agents, browser/terminal use — with adaptive thinking on by default. Community reception is mixed: some developers call it their new default, while detailed evaluations found it token-inefficient in practice, often costing more per completed task than Opus despite lower per-token rates. One benchmark comparison against GLM 5.2 scored it poorly on value. Safety improvements include better prompt-injection resistance and lower hallucination rates, though it trails Opus 4.8 on safety metrics. ## Content Anthropic dropped Claude Sonnet 5 on July 1, 2025, and the pitch is straightforward: near-Opus 4.8 performance at Sonnet pricing. Introductory rates are $2/M input and $10/M output through August 31, rising to $3/$15 after that. Opus 4.8 sits at $5/$25, so on paper this looks like a significant deal. The model is already everywhere. GitHub Copilot, Cursor, Devin, Perplexity, Lovable, Framer, GitLab, Kilo Code, Augment — basically every major coding tool flipped the switch within hours of launch. Cursor's internal benchmark shows a jump from 49% to 57% on their test suite. GitLab says it's the first model to complete 100% of their evaluation tasks, up from 93.8% for Sonnet 4.6. On SWE-bench Pro, it scores 63.2% versus Sonnet 4.6's 57.6% — real improvement, though still behind Opus 4.8's 69.2%. The headline capability is agentic orchestration: Sonnet 5 can now spontaneously spin up sub-agents, use browsers and terminals, and break multi-step tasks into pieces without being prompted to. That behavior was previously Opus-only. Anthropic is also pushing an "advisor pattern" where Opus handles planning and Sonnet 5 does execution — which tells you something about where they think the model actually sits. ## The tokenizer problem nobody put in the press release Here's where it gets complicated. Sonnet 5 ships with a new tokenizer, and the same text produces meaningfully more tokens than it did on Sonnet 4.6 or GPT's o200k tokenizer. Simon Willison flagged it: English text runs about 1.4x more expensive, Spanish about 1.33x more, while Simplified Mandarin comes out roughly the same. A separate analysis of TypeScript specifically found that the same file costs 1.73x more tokens on Claude's new tokenizer than on GPT's o200k — and about 31% more than Claude's own previous tokenizer, with no change to the rate card. For coding agents whose context windows are mostly TypeScript, JavaScript, and JSON, the effective price of Opus 4.8 at $5/Mtok works out closer to $7.50/Mtok once you account for tokenizer inflation. The same math applies to Sonnet 5. During the introductory period, the token inflation still yields roughly 13% savings over Sonnet 4.6. After September 1, the same workload costs about 30% more. That's not a price cut — that's a price hike with a discount window attached. ## The token efficiency problem is separate and worse Beyond the tokenizer, Sonnet 5 appears to be genuinely inefficient at hard tasks. Multiple independent evaluations are landing on the same finding: the model grinds. Kilo Code's writeup notes that on harder problems, Sonnet 5 burns up to 2x more tokens than Opus because it loops inefficiently rather than solving cleanly. A 150-task benchmark comparing Sonnet 4.6 and Sonnet 5 found that on architecture tasks, Sonnet 5 consumed up to 12x more tokens and produced lower quality output — 78% vs 90% on idiomatic checks — making it 3.7x more expensive despite the lower per-token rate. Token variance was also wild: identical prompts produced anywhere from 16K to 6.6M tokens. One reviewer compared it directly against GLM 5.2 on custom benchmarks (3D simulations, SVG generation, coding tasks) and scored GLM 5.2 at 81.43% for $4.88 versus Sonnet 5 at 55.71% for $9.40. That's a pretty brutal result if it holds up across more workloads. The Agent Arena leaderboard has Sonnet 5 at #6, below Opus 4.7 and just above GPT-5.5 high. Not a disaster, but not the "near-Opus" framing Anthropic is leading with. ## Where it actually works The picture isn't uniformly bad. The same 150-task benchmark that found Sonnet 5 struggling on architecture tasks found it genuinely better on instruction-following work — SPFx project upgrades went from 60% to 100% task completion. Cursor's numbers are real. GitLab's 100% task completion is real. Lovable says it catches risky requests better than Sonnet 4.6. Framer says it's a stronger partner for design work specifically. The emerging consensus from people who've actually run it: Sonnet 5 is good at well-defined, decomposable tasks where the sub-agent orchestration helps. It struggles on open-ended architecture work where it needs to reason about tradeoffs rather than execute steps. One analyst put it plainly — it's best used as a sub-agent called by a smarter orchestrator, not as a daily coding workhorse. Pierre Skirano pushed back on the benchmark skepticism: "I wouldn't weigh too heavily on the Sonnet 5 benchmarks not being that much better than Opus. It's a surprisingly sharp model, especially on long-horizon tasks." That's a reasonable counterpoint, but it's also the kind of thing that needs more than a few days of production data to verify. ## What to actually do If you're running coding agents at scale, measure your own token inflation before migrating. The headline $/Mtok comparison is not the number that matters — cost per completed task on your actual content is. During the introductory window, Sonnet 5 is probably worth testing on well-scoped agentic workflows. After September 1, the math gets harder to justify unless your workloads happen to be the ones where it's genuinely more efficient. For complex, open-ended tasks, Opus 4.8 still looks like the right call. For simple, high-volume work, Sonnet 4.6 may still be cheaper in practice. Sonnet 5 is most interesting in the middle — agentic pipelines with clear task boundaries where the sub-agent orchestration actually helps and the token burn stays predictable. ## Community discussion Top comments from developers on daily.dev. **@yaireo** · 3 upvotes > From my hours of experience with it on different effort levels it's not impressive at all, and the output can appear "pretty stupid". I am a full-time senior frontend dev with decades of experience and have been using many models since early AI so I have a good sense when a model is good. Opus seems to me much better (also not as good as me, by far, but at least the output is less "stupid" most of the times) --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic) [View this post on daily.dev](https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Claude Sonnet 5 is near-Opus performance at half the price, with a tokenizer catch","url":"https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2"},"datePublished":"2026-07-01T18:25:22.985Z","dateModified":"2026-07-13T19:42:54.995Z","description":"Anthropic released Claude Sonnet 5, positioning it as near-Opus 4.8 quality at Sonnet pricing ($2/$10 per million tokens introductory, rising to $3/$15 after...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"llm,ai-agents,claude,anthropic","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Claude Sonnet 5 is near-Opus performance at half the price, with a tokenizer catch"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/claude-sonnet-5-is-near-opus-performance-at-half-the-price-with-a-tokenizer-catch-vpk3cyaz2","comment":[{"@type":"Comment","text":"From my hours of experience with it on different effort levels it’s not impressive at all, and the output can appear “pretty stupid”. I am a full-time senior frontend dev with decades of experience and have been using many models since early AI so I have a good sense when a model is good. Opus seems to me much better (also not as good as me, by far, but at least the output is less “stupid” most of the times)","datePublished":"2026-07-02T06:38:05.852Z","url":"https://daily.dev/posts/VpK3CYaz2#c-L1Chr0rdd","author":{"@type":"Person","name":"Yair Even Or","url":"https://daily.dev/yaireo","image":"https://avatars.githubusercontent.com/u/845031?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}}]} ```