Anthropic's new model is coming! Claude Sonnet 5, codenamed Fennec, is about to be released. Its performance beats all large programming models on the market, and its price is cut by 50%. It can also rival an entire human development team. It can be said to have reached the pinnacle of the programming field.Claude Sonnet 5, coming soon!
Multiple information shows that Claude Sonnet 5 (codenamed Fennec) already exists in Google's infrastructure, and is a full generation ahead of Google's "Snow Bunny" in performance.
It has a context window of 1 million tokens and is priced 50% cheaper than Opus 4.5. It will directly solve the core dissatisfaction of developers with Opus that it is "too slow and too expensive".
Moreover, this time it is not a simple model upgrade, but Anthropic directly plugged a development team into the model.
On SWE-Bench, its score exceeds 80.9%, far exceeding all large programming models currently on the market!
It can be said that this agent, which is cheaper, faster, has stronger coding capabilities and a multi-agent development model, is the key unlocking moment for Claude Code.
Beginning: a "404 error log"
This is how things started.
On February 2, a developer discovered a mysterious model ID when calling the API on Google Vertex AI.
claude-sonnet-5@20260203
Its access result is 404 Not Found.
Everyone understands it, which means that the model has been deployed in Google's infrastructure, but it has not been used by you yet.
Later, the well-known AI blogger Pankaj Kumar directly confirmed that Claude Sonnet 5 is codenamed Fennec and is already in Google infrastructure.
The release time is likely to be February 3, 2026, or the day after tomorrow at the earliest.
The AI programming war is about to begin!
And now, early reviews have leaked.
Testers found that Sonnet 5 showed extremely strong performance when processing structured visual generation tasks. Just with the prompt word "ASCII World Map", it generated the most complete and detailed results that testers had ever seen!
Sonnet 5 generates website
In addition, Sonnet 5 also performs well in UI rendering and complex code construction.
Sonnet 5 play games
Sonnet 5, performance surpasses Opus 4.5
Leaked information shows that the overall performance of Claude Sonnet 5 has surpassed Claude Opus 4.5 in multiple internal benchmark tests.
However, its positioning is Sonnet, not Opus. In other words, Anthropic is using mid-range models to serve as flagships.
The programming capabilities of Sonnet 5 are as strong as ever.
Internal test data shows that Claude Sonnet 5 has scored more than 80.9% on SWE-Bench.
The previous SWE-Bench had a highest score of 74.4%
Therefore, it can independently complete most of the problems that software engineers encounter daily, such as fixing bugs, refactoring, supplementing tests, running CI, finding regression problems, etc.
In a word, Claude Sonnet 5 is not just writing code, he is delivering code!
At the same time, here comes the part that makes friends lose sleep the most - its inference cost has dropped significantly, and the price is only 50% of that of Claude Opus 4.5!
The reason is simple: Sonnet 5 is trained and optimized directly on Google TPU. While others were still scrambling for H100, Anthropic had already found another way.
In addition to the price advantage, the context of Sonnet 5 is also up to 1 million tokens, and the running speed has been improved, with lower latency.
From then on, we can stuff the entire project code base into it. It is not just about "looking at the files", but about understanding the overall situation before starting.
For large projects, legacy systems, and mountains of code, this is a life-saving feature.
Automatic development team model, online!
Here comes the really scary part, Claude has also evolved an "automatic development team" model.
This new form is called Dev Team.
In Dev Team mode, Sonnet 5 can automatically generate multiple sub-agents and then collaborate in parallel to complete development.
As long as you give a request, multiple sub-agents will work in parallel and verify each other at the same time.
The final output is not a piece of code, but a fully usable, testable, and iterable functional module.
It can be said that this is no longer Copilot, but an entire development team that does not ask for leave!
Breaking news: Claude Code’s “swarm” mode is coming
In fact, a few days ago, there was news about Claude Code’s swarm mode (Agent Swarm).
It can be guessed that this bee swarm model has an extremely close relationship with the Dev Team that is said to be online.
The key person who broke the news was Mike Kelly. He broke the news on X: There is a Swarm function hidden in Claude Code that has not yet been made public.
Moreover, he also released a Fork project-Claude Sneak Peek for everyone to try.
The so-called Swarms is a "general commander" AI, plus a group of professional sub-agents that are on call and ready to go, forming a complete collaborative system. This model is as follows -
Hierarchical: commander-in-chief → team leader → executor
Dependency: Task A must be completed before task B can be started.
Broadcast: a message that all sub-agents receive simultaneously
Messaging System: agents can communicate with each other
In this way, all AI operates like a company.
In fact, as early as July last year, Anthropic launchedCustom Sub Agents. However, they have many fatal flaws. For example, the memory is fragmented and requires manual patching by humans, and the creation of sub-Agents also relies entirely on humans.
But this time the "swarm" is different!
A developer's practical task began: let AI create a web front-end for the tool, allowing you to download videos from X in the terminal.
Let Claude read aplan.md, requiring it to split tasks and execute them using Swarm.
Here comes the amazing scene: at the beginning, there was no custom child agent.
After performing the mission, a miracle happened! Claude automatically created Team Leader, Front-end Builder, Backend Builder, Component Builder and QA Tester.
These agents are all working in parallel, and their status is visible in real time. You can even enter any Agent and check its system prompt to see what they are thinking.
The most exaggerated part is that Agent can also "give birth to children by himself." Halfway through the task, Claude found that it was not enough, so he generated PI Server Agent, front-end construction agent, CSS special agent, API integration agent, etc.
The number of Agents directly expanded to 8, and no one intervened during the whole process! It can be said that this is no longer a multi-agent, but AI self-organizing.
After the task, Claude also output two documents: project summary and execution report, which are already project management level outputs.
However, why has such a powerful bee swarm function not been opened yet? The reason is that this super scheduler, which can read all your context, is too dangerous and cannot be released easily.
It is conceivable that if Swarm is truly open, Claude Code will no longer be just an AI that writes code, but a system that can form its own R&D team.
This will be a complete paradigm upgrade.
Killing intent at the right time: confronting OpenAI head-on
In the past few days, OpenAI has been making fierce moves in the same period. Ultraman has already announced that the upcoming release month of Codex Family Bucket will be here.
Anthropic chose this time to release Sonnet 5, and its intention is obvious: Come on, fight head-on!
And the answer to the reason why we have this confidence is one word - money.
By the end of January 2026, Anthropic has completed more than US$10 billion in financing, with a valuation of US$350 billion, and the final financing scale may reach US$20 billion.
According to Menlo Ventures' report in 2025 Q4, Anthropic's enterprise market share has accounted for 40%, far ahead of OpenAI's 27% and Google's 21%.
And in just half a year, Claude Code’s annual revenue exceeded US$1 billion.
Recently, Anthropic has directly triggered a series of programming singularities with Cowork, and even hatched the birth of new species such as Clawdbot and Molbook.
If this leak is true, the Claude Sonnet 5 (Fennec) will be the first model to truly "work as a team."
February 3, 2026 will likely be remembered as one of the watershed dates in software engineering.
References:
https://x.com/RichOBray/status/2018351965323850149
https://x.com/pankajkumar_dev/status/1753448405523136512