Claude Sonnet 5: Anthropic collapses the pricing pyramid and brings autonomous agency to the mainstream model

By Zendoric · June 30, 2026.

Anthropic has just launched Claude Sonnet 5, and the official headline —'performance almost equal to Opus at a lower price'— hides something more structural: the company is deliberately compressing its own pyramid of models. According to the published data, Sonnet 5 matches Opus 4.8 in performance, surpasses Sonnet 4.6 in reasoning, tool use, code and knowledge work, and is capable of finishing complex tasks the previous generation did not complete. The launch price is 2 dollars per million input tokens and 10 per million output, with a scheduled increase on September 1 to 3 and 15 dollars respectively. It is available on all plans and becomes the default model for the Free and Pro tiers.

What most deserves attention is not the price but the functional description: Sonnet 5 is, according to Anthropic, the most agentic Sonnet model to date. It can make plans, handle external tools such as browsers and terminals, run autonomously and —here is the novelty that matters— review its own output without anyone prompting it. It also shows lower rates of hallucination and sycophancy, and is better at refusing malicious requests. That is not just 'more capable'; it is a qualitative leap in the risk and reliability profile when deployed in workflows without constant human supervision.

**Our reading: agency is no longer a high-end luxury.**

Until recently, sophisticated agentic capabilities —real autonomy, self-correction, chained tool use— lived in the most expensive models, with prices that reserved them for enterprise use cases with high tickets. Sonnet 5 lowers that threshold to the free plan. That means any developer, startup or advanced user has access to an agent that can work on its own on complex tasks, without needing to supervise every step.

The move has a clear competitive logic: OpenAI, Google and the Chinese open-weight ecosystem (Qwen, DeepSeek, Kimi) are pushing hard in the segment of high-performance 'mid-range' models. Anthropic responds not only with power, but with a differentiating safety argument: less sycophancy, better refusal of malicious requests, active self-correction. These are exactly the vectors where competitors' models have drawn the most criticism in real deployments.

The detail of the introductory pricing also deserves a paragraph of its own. The current 2/10 dollars will rise to 3/15 on September 1, a 50% increase. It is a classic accelerated-adoption strategy: hook developers' workflows at a low price and then normalize the real price. It is not deception —it is announced months in advance— but the cost architectures of those who integrate it today should factor in that increase from day one.

**The underlying pattern we are tracking.**

In our model comparison, Fable 5 (also from Anthropic, launched on June 9) clearly leads Zendoric's Quality index with 90 points. Sonnet 5, hovering around the performance of Opus 4.8, would sit at an upper-intermediate level, well above the current open-weight ecosystem but below the ceiling set by Fable. What is relevant is that the gap between that ceiling and the model anyone can use for free is closing at a pace that two years ago would have seemed optimistic.

In the short term, there is an uncomfortable transition that should not be glossed over: when a free model can autonomously execute complex knowledge tasks, the pressure on intermediate roles —analysis, process management, technical back office— ratchets up another notch. This is not doom-mongering; it is the dynamic we were already seeing in banking, law or business administration. The question is whether educational systems, organizations and labor policies are moving at the same speed as the technology stack. The honest answer is that, for now, they are not.

In the long term, this compression of the pyramid is exactly the mechanism by which AI ends up being widespread infrastructure: when quality agentic capability costs the end user nothing, the workflows that once required hiring specialized labor become accessible to anyone with an idea and time. That is the beginning of the abundance we defend as a horizon —but the road between 'today' and 'that horizon' is not a smooth slide.