Anthropic launched Claude Sonnet 5 on June 30, 2026 — and it immediately became Kunya's default model for Anthropic-powered conversations. If you've opened a new chat recently and noticed things feel a little sharper, more deliberate, that's why. But before you assume Sonnet 5 is simply "better than everything that came before," there's an important nuance worth understanding: Sonnet 5 is a specialized leap, not a universal one. Claude Sonnet 4.6 is still available on Kunya, and for a meaningful set of tasks, it's still the right choice.
This article breaks down what's actually new in Claude Sonnet 5, where it genuinely outperforms its predecessor, and how to think about choosing between the two models on Kunya.
What Is Claude Sonnet 5?
Claude Sonnet 5 is Anthropic's latest mid-tier model — sitting in the same product line as Sonnet 4.6 but rebuilt with a different architecture focus. Where previous Sonnet iterations balanced general-purpose capability with speed, Sonnet 5 is explicitly optimized for agentic performance: complex, multi-step tasks that require planning, tool use, code execution, and autonomous decision-making over extended sessions.
It ships with three headline changes:
Selectable reasoning effort — you can dial reasoning from
lowtoxhigh, giving you control over speed vs. depthA new tokenizer — improving efficiency on code-heavy and structured outputs
Substantially improved agentic benchmarks — more on those numbers below
Pricing remains identical to Sonnet 4.6: $3 per million input tokens and $15 per million output tokens. The context window stays at 1 million tokens. So the upgrade path, at least on the cost side, is frictionless.
Where Claude Sonnet 5 Actually Pulls Ahead
Let's get specific. "Better" is a lazy word. Here's what the numbers actually show — and why they translate into real workflow improvements for Kunya users.
On MMLU (Massive Multitask Language Understanding — the broad academic knowledge benchmark spanning 57 subjects from medicine to law to abstract algebra), Sonnet 5 scores approximately 91.5%, up from Sonnet 4.6's ~88%. That gap reflects genuine improvements in domain reasoning, not just surface-level fluency. It's the difference between a model that retrieves facts and one that actually reasons within a domain. On graduate-level reasoning tests like GPQA Diamond — problems designed to stump PhD-level researchers in biology, chemistry, and physics — Sonnet 5 hits around 84%. That's a significant jump over Sonnet 4.6's ~78%, and it shows up most clearly in tasks requiring multi-step inference rather than pattern recognition. When the path to an answer isn't obvious, Sonnet 5 finds it more often.
The headline number, though, is SWE-bench.
Agentic and Coding Tasks
SWE-bench is the industry-standard evaluation for AI performance on real-world software engineering tasks — not toy problems, but actual GitHub issues pulled from production codebases. The model is handed a broken repo and a bug report, and it has to fix the issue autonomously. No hints. No hand-holding. Sonnet 5 scores 72.7% versus Sonnet 4.6's 65%. That 7.7-point gap is not marginal. In software engineering terms, it represents entire categories of bugs correctly diagnosed, refactors completed without regressions, and features shipped without a human cleaning up after the model. At 65%, you're constantly reviewing the model's work. At 72.7%, you're reviewing it occasionally.
Even more telling is the Terminal-Bench score of 80.4% — a benchmark designed specifically around autonomous terminal work: running commands, managing file systems, navigating complex directory structures, and recovering gracefully from errors without being explicitly told to try again. Sonnet 4.6 would stall or loop when it hit an unexpected state. Sonnet 5 recovers. This is where the model's architectural intent shows most clearly: it was built to operate in real environments, not controlled demos.
In practice, Sonnet 5 handles tasks that would have required significant human supervision on 4.6, including:
Identifying a race condition across three interdependent files and proposing a fix that doesn't break the existing test suite — without being told which files are involved
Executing a multi-step browser automation workflow, hitting an unexpected redirect mid-sequence, and self-correcting without the user noticing anything went wrong
Chaining 12+ tool calls across a research session while maintaining consistent variable naming, logic, and output format across the full context window
Refactoring a legacy API endpoint — rewriting the handler, updating the schema, flagging downstream dependencies, and leaving inline comments explaining what changed and why — in a single agentic pass
Debugging a failing CI pipeline by reading the error log, tracing the failure back to a dependency version conflict introduced three commits ago, and outputting a precise fix with rollback instructions
The selectable reasoning effort feature amplifies all of this. Set it to xhigh for a gnarly architectural decision; drop it to low for a quick regex or boilerplate script. You're not paying the reasoning overhead on tasks that don't need it — and on the tasks that do, you're getting the full engine. That's a meaningful efficiency lever, and it's one of the reasons Sonnet 5 makes sense as a default rather than a specialty tool you reach for occasionally.
Complex, Multi-Turn Reasoning
Beyond raw coding benchmarks, Sonnet 5's improvements in sustained logical coherence are where Kunya users doing serious knowledge work will feel the difference most. Think about tasks that feel more like a project than a question — synthesizing contradictory findings across 30 research papers, building a structured decision framework with 8 competing constraints, pressure-testing a legal argument across multiple rounds of adversarial follow-up, or maintaining a consistent analytical thread across a conversation that's been running for an hour and covered significant ground.
Sonnet 4.6 would start to drift on these. It would quietly drop a constraint it established earlier, or give you an answer in round 12 that technically contradicted something it said in round 4. Not obviously — it would still sound confident and coherent. But the internal consistency would erode. If you weren't tracking carefully, you'd miss it. Sonnet 5 holds the thread. The reasoning architecture improvements mean the model actively tracks its own prior commitments rather than treating each response as a fresh generation. It remembers what it ruled out and why. It flags when new information creates tension with an earlier position rather than silently resolving the conflict in whichever direction is easier.
That matters enormously when you're doing the kind of deep, iterative work Kunya is built for. A financial model built across 15 conversation turns needs the assumptions from turn 3 to still be live in turn 15. A content strategy developed through back-and-forth refinement needs the constraints you set early to actually constrain the final output. Sonnet 5 delivers that. Sonnet 4.6 delivered it sometimes.
If your use case involves legal analysis, financial modeling, complex content strategy, competitive research synthesis, or any domain where internal consistency isn't optional — Sonnet 5 is the version you want running.
Where Claude Sonnet 4.6 Remains the Better Choice
Here's the part of the story that often gets lost in model launch coverage: Sonnet 5's architecture is optimized for planning and agentic overhead — and that optimization comes with tradeoffs.
Claude Sonnet 4.6 was tuned for a different kind of excellence. It produces prose that feels more natural and less mechanical, handles creative constraints with more flexibility, and responds to conversational prompts without introducing the planning-layer overhead that Sonnet 5 brings to every interaction.
Concretely, Sonnet 4.6 still outperforms — or at minimum matches — Sonnet 5 in:
Creative writing: Fiction, poetry, marketing copy, brand voice work — tasks where "thinking more carefully" can actually produce stiffer, more formulaic output
Concise conversation: When you want a quick, sharp answer, Sonnet 4.6 doesn't introduce reasoning latency you didn't ask for
Tone-sensitive communication: Emails, social content, customer-facing copy — writing that needs to feel human rather than exhaustively reasoned
Simple Q&A and lookups: There's no reason to pay the cognitive overhead of agentic architecture for a task that doesn't need it
Think of it this way: asking Sonnet 5 to write a warm birthday message is like hiring a systems architect to paint your living room. Technically capable. Probably overkill. Sonnet 4.6 was built for that kind of task and does it more naturally.
Comparison at a Glance
⚡ Claude Sonnet 5 vs. Sonnet 4.6: Quick Reference
Capability | Sonnet 5 | Sonnet 4.6 |
|---|---|---|
SWE-bench (coding) | ✅ 63.2% | 58.1% |
Terminal-Bench (agents) | ✅ 80.4% | — |
Selectable reasoning effort | ✅ Yes | No |
Creative & conversational writing | Good | ✅ Better |
Concise, low-overhead responses | Good | ✅ Better |
Context window | 1M tokens | 1M tokens |
Pricing | $3 / $15 | $3 / $15 |
Which Model Should You Pick? A Practical Guide
Rather than thinking about Sonnet 5 as an upgrade and Sonnet 4.6 as a legacy option, think of them as two tools with different strengths. Here's how to decide in the moment:
Choose Claude Sonnet 5 When...
You're building or running an AI agent that needs to use tools, browse, or execute code autonomously
You have a complex coding task: debugging a gnarly issue, refactoring a large codebase, or generating a multi-file feature implementation
You need deep reasoning over long context: synthesizing a 200-page document, analyzing a sprawling dataset, or multi-step planning
You want to tune reasoning depth per task using the low-to-xhigh effort selector
You're working in a terminal or browser automation context
Choose Claude Sonnet 4.6 When...
You're doing creative writing: stories, scripts, poetry, brand copy, or anything where voice and feel matter more than logical completeness
You want fast, conversational responses without planning overhead
Your task is simple and well-defined: summarizing a short document, drafting a quick email, answering a factual question
You're producing tone-sensitive content for human audiences where warmth and naturalness are the priority
You just want a writing partner, not a reasoning engine
Both models are available on Kunya right now. Switching between them takes seconds — it's just a model selector in your chat settings. You can start a task in Sonnet 5, realize you need something more conversational, and continue in Sonnet 4.6 without losing your context. The 1M token window means nothing gets left behind.
Why Kunya Made Sonnet 5 the Default
Setting Claude Sonnet 5 as Kunya's default Anthropic model reflects where the platform is heading: more agentic workflows, more complex multi-step tasks, more users building rather than just asking. Sonnet 5's strengths align with what Kunya users increasingly come to the platform to do.
That said, "default" doesn't mean "only option" — and it definitely doesn't mean "right for everything." Kunya keeps Sonnet 4.6 fully available precisely because good tooling respects task context. Defaulting to Sonnet 5 gives new users the most capable starting point for complex work, while experienced users can — and should — reach for Sonnet 4.6 when the task calls for it.
You can read more about how Kunya approaches model selection in our guide to choosing the right AI model for your task, or explore how we support agentic workflows on Kunya.
Try Both and See the Difference Yourself
The most useful thing you can do right now is run the same task through both models and compare. Paste a creative brief into Sonnet 4.6, then try a coding problem in Sonnet 5. The difference in character — one fluid and intuitive, one precise and structured — becomes immediately clear once you feel it firsthand.
Both Claude Sonnet 5 and Claude Sonnet 4.6 are available on Kunya today, at the same price, with the same context window. No upgrade required, no tradeoff forced on you. You get both tools in the same platform, and the choice of which to reach for is yours.
Open Kunya and try Claude Sonnet 5 now — or switch to Sonnet 4.6 in your model settings and see which one fits your workflow better. The right answer probably isn't the same for every task you work on.
