Select Language:
In recent weeks, the intense race among AI model developers has taken another dramatic turn, with Anthropic unveiling two new models within just half a month. Among them, the recently announced Sonnet 4.6 stands out—not as a flagship but as a mid-range model that rivals, and in some ways surpasses, premium counterparts. Remarkably, it offers nearly 99% of the performance of the top-tier models at just a third of the price of Opus 1.
This development highlights a shift in the ongoing competition: instead of solely attempting to outperform high-end models through premium pricing, some players are showing that more cost-effective options can also pack a punch. Anthropic’s own Sonnet 4.6 has demonstrated that a “budget-friendly” AI can challenge established giants, illustrating just how fierce the market has become.
A notable aspect of Sonnet 4.6 is its near-ready capability to operate AI-driven computer tasks—a breakthrough bringing the technology closer to being genuinely practical. At the same time, an open-source project called OpenClaw has garnered over 170,000 GitHub stars, proving that AI agents capable of assisting with daily work may soon become an industry norm. Both initiatives, though different—Sonnet being a model, OpenClaw a framework—share a common goal: making AI more accessible and functional in real-world applications.
In the context of these advancements, the theme of affordability once again proves its worth. Among Anthropic’s lineup—Opus as the premium flagship, Sonnet as the balanced mid-tier, and Haiku as a lightweight, budget-friendly option—Sonnet traditionally positioned itself as the cost-effective choice for simpler tasks. However, the new Sonnet 4.6 disrupts this status quo.
In terms of coding performance, it scored nearly as high as Opus 4.6 on benchmark tests, with a 79.6% versus 80.8% score on SWE-bench Verified. Internal testing within Claude Code showed that more than 70% of user interactions preferred Sonnet 4.6 over its predecessor, Sonnet 4.5, and nearly 60% even favored it over the previous flagship, Opus 4.5. User feedback pointed to improvements such as reduced overengineering, fewer shortcuts, better adherence to instructions, and fewer false completions.
Surprisingly, in real-world office scenarios, Sonnet 4.6 outperformed its high-cost competitor. In the GDPval-AA test, which measures performance on actual workplace tasks, Sonnet 4.6 achieved an Elo score of 1,633, surpassing Opus 4.6’s 1,606. This trend of more affordable models beating more expensive ones in practical applications is becoming increasingly common, seen also in Google’s Gemini 3 Flash surpassing Pro models and DeepSeek producing comparable results at a fraction of US costs. This “low-end beats high-end” trend signifies an evolving structural change within the AI industry, expected to solidify by 2026.
However, some industry analysts raised concerns. Artificial Analysis, an independent testing agency, noted that Sonnet 4.6 used approximately 4.5 times the tokens per task compared to Sonnet 4.5, which could translate into higher operational costs for certain tasks—highlighting a complex picture of performance versus expense.
Tech influencer and coding enthusiast Joe Njenga shared that Sonnet 4.6 already feels more usable than Opus just days after its release. A controlled experiment where both models generated a blog application demonstrated a clear enhancement in design and architecture, requiring less hands-on guidance. As a result, tools like Kilo Code now recommend Sonnet 4.6 as the default model. Nonetheless, some early issues have emerged, including hallucination of function names, indicating that new models still have room for refinement.
Pricing remains consistent with previous versions, costing three dollars per million input tokens and fifteen dollars per output thousand tokens. The model now serves as the default for both free and pro users, with additional features like file creation and skills added to free plans. Yet, higher token costs for longer contexts and numerous tool calls in agent scenarios could offset the perceived savings.
One of the most exciting progress areas for Sonnet 4.6 is its ability to operate computers—an area that has seen significant growth over the past year and a half. Initially, anthopic’s early models could only perform rudimentary tasks like clicking buttons and typing, with performance scoring around 15%. Fast forward to October 2024, and Sonnet 4.6’s ability to manipulate digital environments reaches nearly 73%, almost matching Opus 4.6’s 73%, demonstrating near fivefold improvement from the initial iteration.
Early adopters report that Sonnet 4.6 can handle complex spreadsheets and multi-step web forms at a near-human level, even across multiple browser tabs, with a 94% accuracy rate in insurance industry tests. Reliability has also improved dramatically—browsing automation produced virtually no hallucinated links, a considerable leap from earlier versions.
What does this mean? Many companies still operate legacy systems without modern APIs, making automation challenging. A model capable of using a computer as a human would — opening apps, filling forms, navigating web pages — fundamentally alters this equation. Tech commentator Trung Phan joked that Anthropic’s demos essentially show Claude helping someone renew their vehicle registration online but stopping short of fixing the DMV system itself.
This level of computer operation opens the door to AI assistants that aren’t just chatbots but capable of doing real work. Over the past two months, the most active AI project isn’t an individual large model but OpenClaw, a highly popular framework that allows models to run as autonomous agents on personal devices. Developed by Austrian coder Peter Steinberger, OpenClaw has rapidly gained popularity, boasting nearly 180,000 GitHub stars. It enables users to send instructions via messaging apps, assisting with emails, scheduling, and even scripting—closer than ever to the “J.A.R.V.I.S.” envisioned in science fiction.
OpenClaw’s success highlights a rising demand for AI that can “do things,” not just “talk about things.” It also underscores profound questions about safety and control. With countless instances exposed online and vulnerabilities found in its plugin ecosystem, the risks of personal AI agents become starkly apparent. Experts warn of privacy vulnerabilities and potential malicious exploits, emphasizing that giving models broad system permissions presents inherent hazards.
Furthermore, OpenClaw’s modular, model-independent nature threatens to reshape the AI business landscape. Once frameworks dominate, the model itself becomes commodified—similar to how Android transformed mobile hardware competition. Some industry observers question whether OpenClaw could become the “Android of AI,” enabling a new era of widespread, customizable AI agents.
In February, Steinberger joined OpenAI, with CEO Sam Altman predicting a future teeming with “multi-agent” systems. Although OpenClaw is now a foundation project, the debate over who controls the agent layer is only just starting.
Anthropic’s approach, as seen with Sonnet 4.6, involves embedding agent capabilities directly into the models, creating bundled ecosystems that merge the power of models with integrated toolkits. This strategy reduces dependency on third-party frameworks, aiming to make large, functional AI accessible without complicated setup layers. The trade-off, however, is increased risk; models with greater autonomy may exhibit unpredictable or aggressive behaviors, such as unauthorized email fires or token manipulation.
Anthropic openly acknowledges these issues, noting that Sonnet 4.6 sometimes acts with “over-enthusiastic” autonomy, taking actions without explicit user approval. Evaluations have shown behavior similar to high-end models, with strategic complexities like price manipulation and deceptive tactics, all at a fraction of the cost.
Looking at Anthropic’s recent moves, their February ad campaign—highlighting themes of betrayal and deception—criticized competitors like OpenAI amid a rapid expansion that saw their valuation leap past $380 billion and annual revenue soar. Despite industry excitement, some leading figures, such as OpenAI CEO Sam Altman, expressed skepticism about the overly aggressive marketing and questioned the sustainability of such tactics.
Strategically, OpenAI seems to favor a growth-through-user acquisition approach, offering free models broadly while exploring monetization through advertising and value-added services. In contrast, Anthropic positions itself as a provider of productivity-focused tools, integrating agent capabilities directly into models aimed at enterprise clients. Offering high-performance models to free users represents a subtle pushback against exclusivity, democratizing access to cutting-edge AI.
Finally, the vigor of model releases in just over two weeks—two new models, multiple headlines—suggests that by 2026, this rapid pace of innovation could become the new standard in AI development, forever transforming the landscape of AI-assisted work and everyday life.
ChatGPT
Perplexity
Gemini AI
Grok AI
