On 30 June 2026, Anthropic released Claude Sonnet 5, the new generation of its mid-tier model. The pitch is straightforward: performance approaching Opus 4.8, the lab's flagship model, at a fraction of the price. Sonnet 5 is now the default model on the Free and Pro plans, replacing Sonnet 4.6, which shipped in February.
Here is what stands out from Anthropic's announcement, cross-referenced with coverage from ZDNet, Le Clairvoyant Fnac and Les Numériques.
Contents
The context: a full-blown price war
Sonnet 5's launch fits into an AI race where the performance-to-price ratio has become a selling point in its own right. OpenAI (GPT-5.5, soon GPT-5.6) and Google (Gemini 3.1 Pro, then 3.5 Pro) keep shipping announcements on increasingly short cycles. Anthropic answers by betting on a mid-tier model capable of handling work that, until now, required much heavier and more expensive models.
What's new in Sonnet 5
Anthropic describes Sonnet 5 as "the most agentic Sonnet model ever built": it plans its own tasks, operates tools such as a web browser and a terminal, and carries out complex missions autonomously.
The real technical novelty lies in a five-level effort dial: low, medium, high, very high (xhigh) and maximum. The higher the effort, the longer the model thinks, the more tools it explores, and the more it refines its answer, at the cost of a steeper bill. At low or medium effort, Sonnet 5 already outperforms Sonnet 4.6 for a much lower cost. At high effort, it matches Opus 4.8 on several complex tasks.
Testers cited by Anthropic note that Sonnet 5 carries complex tasks through to completion where earlier versions stopped halfway, and that it spontaneously double-checks its own results without being asked. A developer at Zapier reports handing it a two-step task that ran through to completion entirely on its own.
Benchmarks: where does Sonnet 5 land?
On the SWE-bench Pro agentic coding benchmark, Sonnet 5 scores 63.2%, up from 58.1% for Sonnet 4.6, and behind Opus 4.8's 69.2%. On knowledge-work tasks (GDPval-AA v2), Sonnet 5 even edges slightly ahead of Opus 4.8.
| Benchmark | Sonnet 4.6 | Sonnet 5 | Opus 4.8 |
|---|---|---|---|
| SWE-bench Pro (agentic coding) | 58.1% | 63.2% | 69.2% |
| GDPval-AA v2 (knowledge work) | — | slightly ahead | baseline |
Anthropic is nonetheless clear about the limits: Opus 4.8 remains the recommended model for work that demands maximum precision, such as in-depth analysis or complex judgment calls.
Pricing: the headline argument
This is Anthropic's main selling point: Sonnet 5 costs less than Opus 4.8, GPT-5.5 and Gemini 3.1 Pro, while claiming comparable performance on many tasks.
| Model | Introductory price (in / out per million tokens) | Standard price (in / out) |
|---|---|---|
| Sonnet 5 | $2 / $10 | $3 / $15 |
| Opus 4.8 | — | $5 / $25 |
The introductory pricing holds through summer 2026 (ZDNet cites 31 July, Les Numériques cites 31 August). It's worth checking Anthropic's documentation directly when billing a project. After that date, the standard rate of $3 / $15 per million tokens applies.
Safety and alignment
Anthropic publishes a misaligned-behavior score (hallucinations, sycophancy, cooperation with malicious requests), where a lower score is better.
| Model | Misaligned-behavior score |
|---|---|
| Sonnet 4.6 | 2.89 |
| Sonnet 5 | 2.53 |
| Opus 4.8 | 2.10 |
| Claude Mythos Preview | 1.95 |
Sonnet 5 shows fewer hallucinations and better resistance to prompt injection attempts than Sonnet 4.6, and never produced a functional cybersecurity exploit during Anthropic's testing. It still trails Opus 4.8 and Mythos Preview on the most advanced alignment evaluations, and its ability to carry out dangerous cybersecurity tasks is markedly lower than that of the Opus models.
Anyone deploying autonomous agents with access to a browser or a terminal should test this with their own guardrails before going to production, regardless of the scores published by the vendor.
Availability
Claude Sonnet 5 is available today in Claude Code, on the Claude Platform, and via the API under the identifier claude-sonnet-5. It is also the default model on the Free and Pro plans, and remains accessible across every subscription tier.
My take as an AI integrator
What stands out most to me in this announcement isn't the handful of extra benchmark points, but the five-level effort dial. On client projects where I build agentic workflows (document extraction, coding agents, automations through Claude Code), being able to fine-tune compute cost based on how critical a task is changes the economics in a very concrete way: low effort for sorting or classification, high effort reserved for the steps that actually matter.
With pricing half that of Opus 4.8 for performance that comes close on agentic coding and knowledge work, Sonnet 5 becomes a serious default candidate for most agentic pipelines in production. Opus stays reserved for cases where a mistake is expensive, such as legal analysis, complex judgment calls, or high-stakes decisions.
Key takeaways
- Sonnet 5 replaces Sonnet 4.6 and is now the default model on the Free and Pro plans
- Five adjustable effort levels (low to maximum) to trade off cost and precision
- 63.2% on SWE-bench Pro, up from Sonnet 4.6 (58.1%), below Opus 4.8 (69.2%)
- Introductory pricing at $2 / $10 per million tokens (in/out), versus $5 / $25 for Opus 4.8
- Safety has improved but still trails Opus 4.8 on alignment
- Available in Claude Code, the Claude Platform and the API (
claude-sonnet-5)
The real test will be putting it to work on actual projects to see whether the announced performance gap with Opus holds up in practice. I'll likely follow up with a hands-on report.
Sources: ZDNet, Le Clairvoyant Fnac, Les Numériques.
