On 30 June 2026, Anthropic released Claude Sonnet 5, a model that approaches the performance of Opus 4.8 (Anthropic's top tier) while costing up to 60% less until 31 August. It is now the default model for all free and Pro accounts. In practice: the power that cost a fortune six months ago is becoming the standard option.
Just a year ago, getting an AI assistant capable of carrying out a complex task end-to-end, without stopping halfway, meant paying for the most expensive and slowest model in the catalogue. On 30 June 2026, Anthropic changed the game with Claude Sonnet 5, presented as "the most agentic Sonnet ever released".
The word agentic has been everywhere in AI news for the past year. It refers to an AI's ability not just to answer a question, but to plan several steps, use tools like a browser or a terminal, and carry a task through to completion autonomously, without being prompted at every stage. That is exactly the ground where Sonnet 5 is shaking up the hierarchy.
What changes in practice
Sonnet 5 becomes, from launch, the default model for the Free and Pro tiers of Claude.ai. It is also available on the Max, Team and Enterprise tiers, as well as in Claude Code (the tool that lets you delegate programming tasks to Claude) and on the API platform for developers.
On the technical side, Anthropic reports marked progress on reasoning, tool use, programming and knowledge work, compared with its predecessor Sonnet 4.6 released in February. Early-access testers report a notable shift in behaviour: the model completes complex tasks where previous versions stopped halfway, and it checks its own work without being asked to.
A senior engineer at Zapier says he gave Sonnet 5 a two-part task: update customer account tiers in Salesforce, then send a launch announcement to business contacts. The model completed both steps end-to-end, without intervention. "It always used to stall halfway before," he sums up. For everyday automation, that is exactly the kind of progress that changes a work routine.
The numbers that matter
Here are the announced prices per million tokens, with the promotional period running until 31 August 2026.
| Model | Promo price (until 31/08) | Standard price (after) |
|---|---|---|
| Sonnet 5 | $2 / $10 | $3 / $15 |
| Opus 4.8 | $5 / $25 (fixed, no promo) | |
On an agentic coding benchmark, Sonnet 5 scores 63.2%, compared with 69.2% for Opus 4.8 and 58.1% for the older Sonnet 4.6. The gap with the top tier is narrowing sharply, for a much lighter bill. On a knowledge work benchmark, Sonnet 5 even edges out Opus 4.8 slightly, which surprised even people inside Anthropic.
Why Anthropic is doing this now
The context gives this release its meaning. Anthropic confidentially filed for its IPO on 1 June 2026, and is heading towards a valuation that could approach $1 trillion. In that climate, proving you can make agentic power accessible at a controlled cost is a strong argument with investors, as much as with companies looking to cut their AI infrastructure costs.
There is also a clear product-line logic. Sonnet remains the fast, affordable everyday model, while Opus keeps the edge on the most precision-demanding tasks, very long contexts, or working with multiple agents in parallel. Anthropic puts it bluntly: "Opus 4.8 remains the model of choice for the highest precision on these tasks, but Sonnet 5 offers developers cheaper options of far higher quality than what existed before."
The safety point, without the filter
Anthropic says Sonnet 5 shows a lower overall rate of undesirable behaviours than Sonnet 4.6 (cooperating with malicious requests, deception), with better ability to spot manipulation attempts via prompt injection. The rate of hallucination and excessive sycophancy (saying what the user wants to hear rather than the truth) also drops.
Important nuance: on offensive cybersecurity capabilities, Sonnet 5 remains well behind Opus 4.8, and far behind Claude Mythos, Anthropic's cybersecurity-specialised model. On a vulnerability exploitation test developed with Mozilla, Sonnet 5 gets a 0% full success rate, with only 13.2% partial success, compared with 68.8% for Opus 4.8. Anthropic therefore did not need to apply the same strict guardrails to Sonnet 5 as it did to Fable 5, the model that has just gone through a turbulent episode with the US government.
What to take away
Sonnet 5 illustrates a broader trend of 2026: the capability that cost a fortune a year ago is gradually becoming the accessible norm. It is not the most powerful model in Anthropic's catalogue, but it is probably the one that will change everyday usage the most, by making serious agentic work available at a price that no longer makes you think twice before launching a task.
For everyday use, personal or professional, it is now the most sensible default option. Keep Opus in reserve for tasks that demand maximum precision, and let Sonnet 5 handle the bulk of the daily work.