Anthropic on Tuesday launched Claude Sonnet 4.6, a mannequin that quantities to a seismic repricing occasion for the AI {industry}. It delivers near-flagship intelligence at mid-tier value, and it lands squarely in the course of an unprecedented company rush to deploy AI brokers and automatic coding instruments.
The mannequin is a full improve throughout coding, pc use, long-context reasoning, agent planning, data work, and design. It incorporates a 1M token context window in beta. It’s now the default mannequin in claude.ai and Claude Cowork, and pricing holds regular at $3/$15 per million tokens — the identical as its predecessor, Sonnet 4.5.
That pricing element is the headline that issues most. Anthropic's flagship Opus fashions value $15/$75 per million tokens — 5 occasions the Sonnet value. But efficiency that may have beforehand required reaching for an Opus-class mannequin — together with on real-world, economically useful workplace duties — is now obtainable with Sonnet 4.6. For the hundreds of enterprises now deploying AI brokers that make thousands and thousands of API calls per day, that math adjustments every thing.
Why the price of working AI brokers at scale simply dropped dramatically
To know the importance of this launch, it’s worthwhile to perceive the second it arrives in. The previous 12 months has been dominated by the dual phenomena of "vibe coding" and agentic AI. Claude Code — Anthropic's developer-facing terminal instrument — has develop into a cultural pressure in Silicon Valley, with engineers constructing total functions by way of natural-language dialog. The New York Occasions profiled its meteoric rise in January. The Verge just lately declared that Claude Code is having a real "second." OpenAI, in the meantime, has been waging its personal offensive with Codex desktop functions and sooner inference chips.
The result’s an {industry} the place AI fashions are not evaluated in isolation. They’re evaluated because the engines inside autonomous brokers — programs that run for hours, make hundreds of instrument calls, write and execute code, navigate browsers, and work together with enterprise software program. Each greenback spent per million tokens will get multiplied throughout these hundreds of calls. At scale, the distinction between $15 and $3 per million enter tokens will not be incremental. It’s transformational.
The benchmark desk Anthropic launched paints a putting image. On SWE-bench Verified, the industry-standard take a look at for real-world software program coding, Sonnet 4.6 scored 79.6% — almost matching Opus 4.6's 80.8%. On agentic pc use (OSWorld-Verified), Sonnet 4.6 scored 72.5%, primarily tied with Opus 4.6's 72.7%. On workplace duties (GDPval-AA Elo), Sonnet 4.6 really scored 1633, surpassing Opus 4.6's 1606. On agentic monetary evaluation, Sonnet 4.6 hit 63.3%, beating each mannequin within the comparability, together with Opus 4.6 at 60.1%.
These aren’t marginal variations. In lots of the classes enterprises care about most, Sonnet 4.6 matches or beats fashions that value 5 occasions as a lot to run. An enterprise working an AI agent that processes 10 million tokens per day was beforehand compelled to decide on between inferior outcomes at decrease value or superior outcomes at quickly scaling expense. Sonnet 4.6 largely eliminates that trade-off.
In Claude Code, early testing discovered that customers most popular Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time. Customers even most popular Sonnet 4.6 to Opus 4.5, Anthropic's frontier mannequin from November, 59% of the time. They rated Sonnet 4.6 as considerably much less susceptible to over-engineering and "laziness," and meaningfully higher at instruction following. They reported fewer false claims of success, fewer hallucinations, and extra constant follow-through on multi-step duties.
How Claude's pc use talents went from 'experimental' to near-human in 16 months
Some of the dramatic storylines within the launch is Anthropic's progress on pc use — the flexibility of an AI to function a pc the way in which a human does, clicking a mouse, typing on a keyboard, and navigating software program that lacks fashionable APIs.
When Anthropic first launched this functionality in October 2024, the corporate acknowledged it was "nonetheless experimental — at occasions cumbersome and error-prone." The numbers since then inform a exceptional story: on OSWorld, Claude Sonnet 3.5 scored 14.9% in October 2024. Sonnet 3.7 reached 28.0% in February 2025. Sonnet 4 hit 42.2% by June. Sonnet 4.5 climbed to 61.4% in October. Now Sonnet 4.6 has reached 72.5% — almost a fivefold enchancment in 16 months.
This issues as a result of pc use is the aptitude that unlocks the broadest set of enterprise functions for AI brokers. Virtually each group has legacy software program — insurance coverage portals, authorities databases, ERP programs, hospital scheduling instruments — that was constructed earlier than APIs existed. A mannequin that may merely take a look at a display and work together with it opens all of those to automation with out constructing bespoke connectors.
Jamie Cuffe, CEO of Tempo, stated Sonnet 4.6 hit 94% on their complicated insurance coverage pc use benchmark, the best of any Claude mannequin examined. "It causes by way of failures and self-corrects in methods we haven't seen earlier than," Cuffe stated in an announcement despatched to VentureBeat. Will Harvey, co-founder of Convey, known as it "a transparent enchancment over anything we've examined in our evals."
The protection dimension of pc use additionally obtained consideration. Anthropic famous that pc use poses immediate injection dangers — malicious actors hiding directions on web sites to hijack the mannequin — and stated its evaluations present Sonnet 4.6 is a significant enchancment over Sonnet 4.5 in resisting such assaults. For enterprises deploying brokers that browse the online and work together with exterior programs, that hardening will not be non-compulsory.
Enterprise clients say the mannequin closes the hole between Sonnet and Opus pricing tiers
The client response has been unusually particular about cost-performance dynamics. A number of early testers explicitly described Sonnet 4.6 as eliminating the necessity to attain for the dearer Opus tier.
Caitlin Colgrove, CTO of Hex Applied sciences, stated the corporate is shifting nearly all of its visitors to Sonnet 4.6, noting that with adaptive considering and excessive effort, "we see Opus-level efficiency on all however our hardest analytical duties with a extra environment friendly and versatile profile. At Sonnet pricing, it's a simple name for our workloads."
Ben Kus, CTO of Field, stated the mannequin outperformed Sonnet 4.5 in heavy reasoning Q&A by 15 share factors throughout actual enterprise paperwork. Michele Catasta, President of Replit, known as the performance-to-cost ratio "extraordinary." Ryan Wiggins of Mercury Banking put it extra bluntly: "Claude Sonnet 4.6 is quicker, cheaper, and extra prone to nail issues on the primary attempt. That mixture was a stunning mixture of enhancements, and we didn't count on to see it at this value level."
The coding enhancements resonate notably given Claude Code's dominance within the developer instruments market. David Loker, VP of AI at CodeRabbit, stated the mannequin "punches means above its weight class for the overwhelming majority of real-world PRs." Leo Tchourakov of Manufacturing facility AI stated the group is "transitioning our Sonnet visitors over to this mannequin." GitHub's VP of Product, Joe Binder, confirmed the mannequin is "already excelling at complicated code fixes, particularly when looking out throughout massive codebases is important."
Brendan Falk, Founder and CEO of Hercules, went additional: "Claude Sonnet 4.6 is the very best mannequin we have now seen to this point. It has Opus 4.6 degree accuracy, instruction following, and UI, all for a meaningfully decrease value."
A simulated enterprise competitors reveals how AI brokers plan over months, not minutes
Buried within the technical particulars is a functionality that hints at the place autonomous AI brokers are heading. Sonnet 4.6's 1M token context window can maintain total codebases, prolonged contracts, or dozens of analysis papers in a single request. Anthropic says the mannequin causes successfully throughout all that context — a declare the corporate demonstrated by way of an uncommon analysis.
The Merchandising-Bench Area assessments how properly a mannequin can run a simulated enterprise over time, with completely different AI fashions competing in opposition to one another for the largest earnings. With out human prompting, Sonnet 4.6 developed a novel technique: it invested closely in capability for the primary ten simulated months, spending considerably greater than its rivals, after which pivoted sharply to concentrate on profitability within the ultimate stretch. The mannequin ended its 365-day simulation at roughly $5,700 in steadiness, in comparison with Sonnet 4.5's roughly $2,100.
This type of multi-month strategic planning, executed autonomously, represents a qualitatively completely different functionality than answering questions or producing code snippets. It’s the kind of long-horizon reasoning that makes AI brokers viable for actual enterprise operations — and it helps clarify why Anthropic is positioning Sonnet 4.6 not simply as a chatbot improve, however because the engine for a brand new era of autonomous programs.
Anthropic's Sonnet 4.6 arrives as the corporate expands into enterprise markets and protection
This launch doesn’t arrive in a vacuum. Anthropic is in the course of probably the most consequential stretch in its historical past, and the aggressive panorama is intensifying on each entrance.
On the identical day as this launch, TechCrunch reported that Indian IT big Infosys introduced a partnership with Anthropic to construct enterprise-grade AI brokers, integrating Claude fashions into Infosys's Topaz AI platform for banking, telecoms, and manufacturing. Anthropic CEO Dario Amodei instructed TechCrunch there may be "an enormous hole between an AI mannequin that works in a demo and one which works in a regulated {industry}," and that Infosys helps bridge it. TechCrunch additionally reported that Anthropic opened its first India workplace in Bengaluru, and that India now accounts for about 6% of worldwide Claude utilization, second solely to the U.S. The corporate, which CNBC reported is valued at $183 billion, has been increasing its enterprise footprint quickly.
In the meantime, Anthropic president Daniela Amodei instructed ABC Information final week that AI would make humanities majors "extra vital than ever," arguing that essential considering expertise would develop into extra useful as massive language fashions grasp technical work. It’s the form of assertion an organization makes when it believes its know-how is about to reshape total classes of white-collar employment.
The aggressive image for Sonnet 4.6 can be notable. The mannequin outperforms Google's Gemini 3 Professional and OpenAI's GPT-5.2 on a number of benchmarks. GPT-5.2 trails on agentic pc use (38.2% vs. 72.5%), agentic search (77.9% vs. 74.7% for Sonnet 4.6's non-Professional rating), and agentic monetary evaluation (59.0% vs. 63.3%). Gemini 3 Professional reveals aggressive efficiency on visible reasoning and multilingual benchmarks, however falls behind on the agentic classes the place enterprise funding is surging.
The broader takeaway might not be about any single mannequin. It’s about what occurs when Opus-class intelligence turns into obtainable for just a few {dollars} per million tokens relatively than just a few tens of {dollars}. Corporations that had been cautiously piloting AI brokers with small deployments now face a essentially completely different value calculus. The brokers that had been too costly to run repeatedly in January are all of the sudden inexpensive in February.
Claude Sonnet 4.6 is offered now on all Claude plans, Claude Cowork, Claude Code, the API, and all main cloud platforms. Anthropic has additionally upgraded its free tier to Sonnet 4.6 by default. Builders can entry it instantly utilizing claude-sonnet-4-6 through the Claude API.
.jpg "Gear Information of the Week: Intel’s New Chips Arrive, and Apple Could Debut iPads and MacBooks This Month")