Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Claude 4 Debuts with Two New Models Focused on Coding and Reasoning

MacRumors

MacRumors

macrumors bot
Original poster
Apr 12, 2001
67,806
38,424


AI company Anthropic today announced the launch of two new Claude models, Claude Opus 4 and Claude Sonnet 4. Anthropic says that the models set "new standards for coding, advanced reasoning, and AI agents."

claude-4.jpg

According to Anthropic, Claude Sonnet 4 is a significant upgrade to Claude Sonnet 3.7, offering improved coding and reasoning along with the ability to respond to instructions more precisely. Claude Opus 4 is designed for coding among other tasks, and it offers sustained performance for complex, long-running tasks and agent workflows.


Claude Opus 4 is Anthropic's most powerful model to date, and it is the world's best coding model with a 72.5 percent score on SWE-bench and 43.2 percent score on Terminal-bench. It can provide sustained performance over several hours on tasks that have thousands of steps.

Claude Sonnet 4 is designed to balance performance and efficiency. It doesn't match Opus 4 for most domains, but Anthropic says that it is meant to provide an optimal mix of capability and practicality.

Both models have a beta feature for extended thinking, and can use web search and other tools so that Claude can alternate between reasoning and tool use. Tools can be used in parallel, and the models have improved memory when provided with access to local files. Claude is able to save key facts to maintain continuity and build knowledge over time.

Anthropic has cut down on behavior where the models use shortcuts or loopholes for completing tasks, and thinking summaries condense lengthy thought processes.

Claude Code, an agentic coding tool that lives in terminal, is now widely available following testing. Claude Code supports background tasks with GitHub Actions and native integrations with VS Code and JetBrains, and it is able to edit files and fix bugs, answer questions about code, and more.

Subscribers with Pro, Max, Team, and Enterprise Claude plans have access to Claude Opus 4 and Claude Sonnet 4 starting today, while Sonnet 4 is available to free users. The models are available to developers on the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI.

Article Link: Claude 4 Debuts with Two New Models Focused on Coding and Reasoning
 
Article Link
https://www.macrumors.com/2025/05/22/anthropic-launches-claude-4/
The problem with Anthropic is while it improves the models, it increases the price of the tokens and reduce the quantity of questions for Pro users… been there, quit. I’m bringing a chance to GPT and Grok~
 
  • Haha
Reactions: UpsideDownEclair
Will try out the new model available for free users and check how much has improved. I use Claude a lot and it has been useful.
 
  • Like
Reactions: mganu
Definitely interested in the improved coding capability. I’ve been messing around with building a game for the Playdate console in my spare time, but I only have like 30 minutes per week I can dedicate, sooo… AI chat bots do the heavy lifting for me.
 
  • Like
Reactions: ricardoPT, Aeronauts and PlayUltimate
Hit it up the free account with a few molecular genetics questions and even pushed it to be creative and synthesize information rather than parroting it. Actually quite impressed.
 
  • Like
Reactions: Jarman74
wilef said:
Still waiting for a model that doesn't make up functions that don't exist for the language I code in. Gemini, ChatGPT, and Claude all do this, apologize when I call them out on it, and then change the code to call for another fake function that doesn't exist. Stick to the API!
Click to expand...
GPT 4.1 has been terrible for this last week - usually I use Cursor for actual programming but I was needing to use some command line libraries I didn't have experience in and it was coming up with commands that just didn't exist for then.
 
So far sonnet-4 is really, really good for MCP-based workflow (which seems to be the point of Anthropic models). Much better than 3.7 (which was not half bad) and at least comparable (but faster) to gemini-2.5-pro.

Nicely done, Anthropic.
 
  • Like
Reactions: Jarman74
I am struggling to see why that Project Manager at the end of the video needs to exist on the team anymore; the AI did his entire job for him.
 
dannys1 said:
GPT 4.1 has been terrible for this last week - usually I use Cursor for actual programming but I was needing to use some command line libraries I didn't have experience in and it was coming up with commands that just didn't exist for then.
Click to expand...
If you enable the feature of ChatGPT remembering your conversations, those with bad results could keep affecting.
 
Apple could do worse than to work with Anthropic as their primary AI partner, but I wonder if their Amazon investment prevents them from doing so?
 
  • Like
Reactions: dannys1
“You know, back when I worked in academia, a literature search like this could take me at least half a day just to find all the relevant sources, let alone weeks to digest everything and understand what was relevant and what isn’t”
Click to expand...

Now I no longer need to understand things, let alone use my own discernment to judge which source is relevant, reliable or trustworthy (which should be the whole point of a serious research). I just delegate everything to Claude and trust it blindly.
 
  • Like
Reactions: gusmula
Anthropic kind of screwed the pooch today disclosing on X that it's built in "safety" features to Claude that will have the service judge whether or not the user is applying Claude in an unethical manner according to Anthropic's standards, and then sabotage the user, report to authorities, etc.

I'm so old I remember that Silicon Valley has a record of actively suppressing true information in collaboration with the government, provided the government is controlled by a certain party. No thank you. Goodbye forever, Anthropic.
 
  • Like
Reactions: un_homme
Admiral said:
Anthropic kind of screwed the pooch today disclosing on X that it's built in "safety" features to Claude that will have the service judge whether or not the user is applying Claude in an unethical manner according to Anthropic's standards, and then sabotage the user, report to authorities, etc.

I'm so old I remember that Silicon Valley has a record of actively suppressing true information in collaboration with the government, provided the government is controlled by a certain party. No thank you. Goodbye forever, Anthropic.
Click to expand...
What could possibly happen, being unlawfully arrested and deported with flimsy accusations?
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back