Claude Sonnet 5 vs Sonnet 4.6 vs Opus 4.8: Benchmarks and Pricing Compared
Claude Sonnet 5 launches with 63.2% SWE-bench Pro and intro pricing of $2/$10 per MTok until Aug 31. Here's how it stacks up against Sonnet 4.6 and Opus 4.8.
Anthropic launched Claude Sonnet 5 on June 30, 2026, positioning it as its most capable mid-tier model for agentic work: long task chains, browser and terminal control, and autonomous multi-step execution. It beats Sonnet 4.6 on every published benchmark and undercuts Opus 4.8 on price. But a new tokenizer and variable effort levels mean the real cost depends heavily on how you run it.
What happened
| Detail | Value |
|---|---|
| Launch date | June 30, 2026 |
| SWE-bench Pro score | 63.2% |
| OSWorld-Verified score | 81.2% |
| HLE score | 57.4% |
| Intro pricing (input / output per MTok) | $2 / $10 (until Aug 31, 2026) |
| Standard pricing after Aug 31 | $3 / $15 |
| Opus 4.8 pricing | $5 / $25 |
| Tokenizer multiplier vs older models | 1.0x to 1.35x more tokens |
Sonnet 5 is the default model for Anthropic’s Free and Pro plans as of launch. Max, Team, and Enterprise users can select it manually. It is also available inside Claude Code and across the Claude Platform API.
Sonnet 5 replaces Sonnet 4.6, which launched in February 2026. Anthropic framed this release around agentic reliability rather than a single headline number. That means better self-correction when a tool call fails, longer task chains without losing context, and steadier behavior across extended sessions in tools like Claude Code.
How does Sonnet 5 compare to Sonnet 4.6 and Opus 4.8?
| Model | SWE-bench Pro | OSWorld-Verified | HLE | Input (per MTok) | Output (per MTok) |
|---|---|---|---|---|---|
| Sonnet 4.6 | – | – | – | $3 | $15 |
| Sonnet 5 (intro) | 63.2% | 81.2% | 57.4% | $2 | $10 |
| Sonnet 5 (standard) | 63.2% | 81.2% | 57.4% | $3 | $15 |
| Opus 4.8 | – | – | – | $5 | $25 |
Sonnet 5 closes a meaningful part of the gap to Opus 4.8, according to Anthropic’s published figures. For accuracy-critical work, Anthropic still positions Opus as the safer pick. Sonnet 5 was also designed with deliberately low cyber capability, which is a deliberate safety choice rather than an oversight.
The tokenizer change matters for cost estimates
Sonnet 5 uses the same updated tokenizer that shipped with Opus 4.7. The same input text can produce 1.0 to 1.35 times more tokens compared to older Sonnet models. If you’re estimating API costs based on previous Sonnet usage, apply that multiplier or your projections will be low.
Effort levels add another variable
The model exposes four effort settings: low, medium, high, and xhigh (extra high). Higher effort uses more tokens on internal reasoning steps, which raises both output quality and cost. At xhigh effort, Sonnet 5 can exceed what you’d pay for the same task on Opus 4.8. Low and medium effort is where the cost-performance advantage is clearest.
Why it matters
For teams running agentic coding pipelines or browser automation, Sonnet 5 is a meaningful upgrade over 4.6 at the same or lower price point (at least through August 31). The OSWorld-Verified score of 81.2% is particularly relevant for anyone using the model to drive terminals or desktop interfaces automatically.
The pricing window is real but short. From September onward, Sonnet 5 and Sonnet 4.6 sit at identical standard rates of $3/$15 per MTok. The decision between them becomes purely about capability, where Sonnet 5 wins on every published metric. For teams already exploring AI integration into their workflows, this is a good time to test Sonnet 5 on real tasks before the intro pricing closes.
If you follow agentic AI model releases closely, it’s worth noting that GPT-5.6 Sol recently posted a 54% efficiency gain on agentic coding as well. The competition in this specific category is moving fast.
Our take
The benchmarks are strong, but the tokenizer change is the detail most people will skip and then wonder why their costs came in higher than expected. A 1.35x token multiplier on a heavy agentic workload is not trivial. Run a small test batch with real prompts before scaling anything.
The effort levels are also worth treating as a tuning dial rather than a set-and-forget option. For tasks where speed matters more than exhaustive reasoning, low or medium effort keeps Sonnet 5 comfortably cheaper than Opus 4.8. At xhigh, that advantage disappears. Set your defaults carefully, especially in automated pipelines where nobody is watching per-request costs accumulate.
Sonnet 5 being the default on Free and Pro plans means a lot of users will switch without noticing. If you have customer-facing tools built on Anthropic’s API, double-check your output quality and token budgets now rather than after an unexpected invoice.
What to do about it
- Apply the 1.0 to 1.35x tokenizer multiplier to your current Sonnet usage numbers to get a realistic cost projection for Sonnet 5.
- Run a test batch on your most common task types at low and medium effort to establish a baseline before touching high or xhigh.
- Lock in workloads you can migrate before August 31 to capture the $2/$10 intro pricing.
- Keep Opus 4.8 in your stack for tasks where accuracy is non-negotiable, particularly anything flagged as safety or compliance-critical.
- Review any customer-facing tools that default to a Sonnet model, as Anthropic has already switched Free and Pro plan defaults to Sonnet 5.
The intro pricing window closes August 31. That’s the most time-sensitive reason to test now rather than later.
Frequently asked questions
What is Claude Sonnet 5's SWE-bench Pro score?
Claude Sonnet 5 scores 63.2% on SWE-bench Pro, according to Anthropic's figures from the June 30, 2026 launch.
How much does Claude Sonnet 5 cost via the API?
Sonnet 5 has intro pricing of $2 per million input tokens and $10 per million output tokens until August 31, 2026. After that, standard pricing applies at $3/$15 per million tokens, the same as Sonnet 4.6.
What is the difference between Claude Sonnet 5 and Opus 4.8?
Opus 4.8 is Anthropic's flagship model priced at $5/$25 per million tokens and is recommended for accuracy-critical work. Sonnet 5 is a mid-tier model that closes much of the performance gap at a lower price point, but Anthropic still positions Opus 4.8 as the pick where precision and safety are paramount.
Does Claude Sonnet 5 use a different tokenizer?
Yes. Sonnet 5 uses the same updated tokenizer introduced with Opus 4.7. The same text can map to 1.0 to 1.35 times more tokens than with older Sonnet models, which affects actual API costs.
