Claude Sonnet 4.6 vs Sonnet 5: What Changed and Which to Use
If you are weighing Sonnet 4.6 against its replacement, the short version is simple: Claude Sonnet 5 beats the previous Sonnet on every axis that matters for real work, and it launched at a lower introductory price. Anthropic released Sonnet 5 on June 30, 2026 as the direct successor to Sonnet 4.6.
The real story is not a small quality bump. It is a jump in agentic capability — planning, tool use, and autonomy — paired with pricing that undercuts Anthropic’s own flagship. For most teams, moving from Sonnet 4.6 to the new Sonnet is a straightforward upgrade.
The short answer: Sonnet 5 replaces Sonnet 4.6
Sonnet 5 is not a side-grade or a niche variant. On launch day it became the default model across Claude’s consumer plans, taking the seat Sonnet 4.6 held before it. Sonnet 4.6 is now the previous generation.
Release and availability
Sonnet 5 launched June 30, 2026 and immediately became the default model on Claude’s Free and Pro plans. It is also available to Max, Team, and Enterprise customers, inside Claude Code, and through the Claude Platform and API under the model id claude-sonnet-5. If you use Claude anywhere that Sonnet 4.6 was the default, you are almost certainly already talking to the new Sonnet.
That broad rollout matters for the comparison: Sonnet 4.6 is being retired from the front of the lineup rather than kept alongside as a cheaper tier. The decision most people face is not “which do I pick” but “am I ready to move.”
Benchmarks: how much better is Sonnet 5?
Across coding, reasoning, tool use, and knowledge work, Anthropic describes Sonnet 5 as a substantial improvement over Sonnet 4.6. The headline number is agentic coding, where the gap is large enough to change what the model can realistically automate.
Agentic coding
On the agentic coding benchmark, Sonnet 5 scores 63.2% against Sonnet 4.6’s 58.1% — roughly a five-point gain. That lands it much closer to Anthropic’s flagship, Opus 4.8, which leads at 69.2%. In practice this narrows the old rule of thumb that you needed Opus for serious multi-step coding agents.
| Model | Agentic coding | API price (input / output per 1M) |
|---|---|---|
| Sonnet 4.6 | 58.1% | $3 / $15 |
| Sonnet 5 | 63.2% | $2 / $10 intro, then $3 / $15 |
| Opus 4.8 | 69.2% | $5 / $25 |
Reasoning, tool use, and knowledge work
Beyond coding, Sonnet 5 improves on reasoning and tool use, and on some knowledge-work benchmarks it slightly surpasses Opus 4.8. Anthropic is candid that this does not make Sonnet the universal winner. As the company notes, “Opus 4.8 is still the model of choice for higher accuracy on these tasks” — so the flagship keeps a role at the top of demanding, accuracy-critical workflows.
The takeaway for the 4.6-versus-5 question is that there is no benchmark category where Sonnet 4.6 comes out ahead. The new Sonnet is a strict step up.
Pricing: the introductory discount
Pricing is where Sonnet 5 gets genuinely interesting, because Anthropic launched it below the standard Sonnet rate for a limited window.
What Sonnet 5 costs
Through August 31, 2026, Sonnet 5 runs at an introductory $2 per million input tokens and $10 per million output tokens. After that window, it settles at $3/$15 — the same standard rate Sonnet 4.6 charged. So during the intro period you get more capable output for less money, and after it, you get more capability at the price you were already paying.
Opus 4.8, by contrast, costs $5/$25 per million tokens. That means Sonnet 5 delivers performance approaching Opus on many tasks at roughly half the price. Anthropic’s pricing documentation is the source of truth for the post-intro rate; some early press coverage listed a different post-intro output price, so check the official docs before you budget.
- Intro (through Aug 31, 2026): $2 input / $10 output per 1M tokens
- Standard (after Aug 31, 2026): $3 input / $15 output per 1M tokens
- Opus 4.8: $5 input / $25 output per 1M tokens
- Sonnet 4.6 (previous): $3 input / $15 output per 1M tokens
New capabilities in Sonnet 5
The benchmark gains come from real changes in how the model works, not just tuning. Two stand out: it is built to act, and it checks itself.
More agentic by design
Sonnet 5 is described as the most agentic Sonnet yet. It makes plans, uses tools like browsers and terminals, coordinates subagents, and runs autonomously for longer stretches without hand-holding. Anthropic frames the leap this way:
Claude Sonnet 5 is built to be the most agentic Sonnet model yet. It can make plans, use tools like browsers and terminals, and run autonomously.
Anthropic
For Sonnet 4.6 users, this is the most tangible difference day to day. Tasks that used to need close supervision — multi-file refactors, browser research loops, chained tool calls — hold together better on the new Sonnet.
Self-correction and safety
Sonnet 5 reviews its own output and fixes errors before you see them, a behavior early users noticed it does even without being asked. That self-checking loop is a big part of why its agentic scores climbed.
On safety, Anthropic reports lower rates of undesirable behavior than Sonnet 4.6, including better refusal of malicious requests and stronger resistance to prompt-injection attacks. Both new Sonnet models scored 0.0% on the Firefox 147 exploit benchmark — an evaluation run with Mozilla, whose findings were addressed in Firefox 148.
Sonnet 5 vs Opus 4.8 and competitors
The 4.6-to-5 upgrade also shifts where Sonnet sits in the broader market, so it is worth knowing the neighbors.
Against Opus 4.8. Sonnet 5 is positioned as the everyday agentic workhorse, with Opus 4.8 reserved for peak-accuracy jobs. On raw agentic coding, Opus still leads (69.2% vs 63.2%), but Sonnet 5’s much lower price makes it the default choice for high-volume automation.
Against OpenAI and Google. Anthropic pitches Sonnet 5 as a lower-cost alternative to OpenAI’s GPT-5.6 Sol and Google’s Gemini 3.5 Flash for routine agentic tasks. It is cheaper than Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro, while remaining more expensive than the very cheap Gemini 3.5 Flash. As TechCrunch reported at launch, the pitch is explicitly a cheaper way to run agents — near-flagship quality at a mid-tier price to win the bulk of production traffic.
Should you upgrade from Sonnet 4.6?
For nearly every use case, yes. Sonnet 5 is a strict improvement over Sonnet 4.6 across effort levels, and after the intro period it costs exactly what 4.6 did. Here is a quick way to confirm the switch is safe for your workload:
- Point one non-critical job at
claude-sonnet-5and keep 4.6 running in parallel. - Compare output quality and total token cost over a representative sample.
- Check that adaptive thinking (on by default in Sonnet 5) behaves as you expect for your prompts.
- Confirm any tool-heavy or agentic flows complete more reliably end to end.
- Roll the new Sonnet out as your default and retire the 4.6 path.
Most teams find the harder question is not whether to move, but how quickly they can take advantage of the new agentic headroom. If your product leans on autonomous coding or research loops, that gain is where the value is. You can try the new model directly through the free browser extension at sonnet5.pro without keys or setup.