This week we got not one but two frontier models dropped within 20 minutes of each other. It is clear that Anthropic and OpenAI are at war, and the first battleground and most important battleground is around coding.
Oh my goodness. Look, it is a good day around here when we get news about upcoming models. It's a great day when we actually get a new model that we get to play with. And it is just something else entirely when within 20 minutes of each other, two of the leading frontier labs drop dueling frontier models, in this case with an incredibly clear focus. The models are of course Claude Opus 4.6 from Anthropic and GPT-5.3 Codex from OpenAI. And for sure the first thing that people noticed is the sequence. Now, none of the frontier labs have ever been above trying to get out in front of one of their peers when it comes to a big announcement, but we've never seen anything like this. Purcell's Ali writes, 'Versus OpenAI is like Kendrick versus Drake, but for nerds.' This indeed was a common refrain. Aush wrote, 'Anthropic drops a model and OpenAI responds 15 minutes later. This is basically a rap beef now.' Peter Yang writes, 'How did Opus 4.6 and GPT-5.3 drop back to back in like 20 minutes? Is this a new cold war or something where there are spies in each company?' The answer, Peter, is almost certainly yes.
But obviously way more important than the popcorn ball watching of the competition is what the models actually have for us. On the Opus side, Anthropic says there are key coding improvements, including better code review and debugging skills to quote catch its own mistakes. Interestingly, it does also feel like Anthropic, because Claude is so associated with coding, used the chance to build on the themes that they've been exploring with Claude Cowork to talk about how Opus 4.6 was better for everyday tasks as well. Still with a focus, of course, on white collar work, but things like running financial analyses, doing research, and using and creating documents, spreadsheets, and presentations. In fact, in their presentation of the benchmarks, they actually put knowledge work right up front. On the benchmark side, they claimed the leading score on the agent coding benchmark Terminal Bench 2.0 and also claimed the top spot on the leaderboard for Humanity's Last Exam. Now, Humanity's Last Exam began as a general knowledge test, but is increasingly a measure of reasoning and tool use capability. In terms of new features, Opus 4.6 now supports million token context windows. And for folks who feel like they constantly run against those limits, that has got to be welcome news.
They've also introduced something that I'm really excited about called agent teams. Now, Anthropic seemingly got the memo that swarms was sort of an intimidating name because they are explicitly trying to move away from the naming of agent swarms and calling it teams instead. The feature basically allows users to set a whole team of Claws to work on a particular problem, including a coordination layer to ensure they're working on separate tasks that all contribute to the whole. They write that agent teams are most effective for tasks where parallel exploration adds real value. Some of the examples they give are cross-layer coordination where you've got coding changes that span front end, back end, and tests. But they also point to research and review where, for example, multiple teammates can investigate different aspects of a problem simultaneously, then share and challenge each other's findings. The difference between sub-agents and agent teams basically comes down to the extent to which you need your agents to communicate with one another. Anthropic sums up: use sub-agents when you need quick, focused workers that report back. Use agent teams when teammates need to share findings, challenge each other, and coordinate on their own. Anthropic has also added a feature called adaptive thinking. It allows the model to pick up on context clues to determine how much reasoning effort to expend on a particular task. Users also have manual controls available to dial up or down the amount of reasoning effort being deployed.
Now, to demonstrate the power of their new agent swarm/agent teams, they tested a fully autonomous coding task. Anthropic writes, 'We tasked Opus 4.6 using agent teams to build a C compiler and then mostly walked away.' The task made use of the Ralph loop for continuous work, which we talked about in a show recently, but basically means a continuous loop where it's always checking in so it doesn't get stuck and continue to take on the same challenge even if it doesn't solve it in one session. Overall, the process consumed around 2 billion tokens, generating over 140 million output tokens and costing around $20,000 using standard API pricing. The test was performed without internet access and only used the standard Rust library. Anthropic also noted that this version of Claude was also created by Claude, with the AI model now the key driver of all coding within Anthropic. Now there are a bunch of additional features that they talked about. Opus 4.6 shows significant improvement on long context retrieval and long context reasoning, which should lead to a much better experience in long horizon tasks as well as coding work that requires the model to access large code bases. Overall, they said with Opus 4.6, 'We found that the model brings more focus to the most challenging parts of a task without being told to, moves quickly through the more straightforward parts, handles ambiguous problems with better judgment, and stays productive over longer sessions.' So, that was 4.6, and then literally 15 minutes later, OpenAI dropped GPT-5.3 Codex.
Now, the first thing you might notice is that we don't have GPT-5.3 yet. And I think it's quite clear that the choice to release the coding-tuned version of GPT-5.3 as a standalone before the release of the regular GPT-5.3 tells you everything you need to know about how these labs see the importance of various use cases. OpenAI says that this model advances both the coding performance of GPT-5.2 Codex and the reasoning and knowledge abilities of regular GPT-5.2 in a single package. Similar to Anthropic, OpenAI's models are now the core of the development team. With OpenAI writing, 'GPT-5.3 Codex is our first model that was instrumental in creating itself. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations. Our team was blown away by how much Codex was able to accelerate its own development.' Max Stober from the ChatGPT team shared that a recently announced feature where ChatGPT had full support for MCP apps was built entirely with GPT-5.3 Codex. Max writes, 'Zero lines of code written by hand. Most times the Codex CLI worked autonomously for hours and implemented parts of this first try.' Now on the benchmark, self-reported of course, Codex 5.3 is a significant jump on SWE-bench Pro compared to the previous version. And maybe even more notably than performance, the performance that was achieved was achieved with far fewer tokens, demonstrating OpenAI's work on token efficiency. The model claims a new absolutely state-of-the-art score of 77.3% on Terminal Bench 2.0, which beats Codex 5.2 at 64% and is, if true, much higher than Opus 4.6 at 65.4%.
For their autonomous coding demo, OpenAI showed off a racing game. Codex 5.3 used the web game skill and was fed generic prompts throughout the process like 'fix the bug' and 'improve the game.' The result demonstrated that Codex 5.3 can churn through a task using millions of tokens without human intervention. Like Anthropic, OpenAI also highlighted various non-coding work that Codex 5.3 excels at. While for developers the model is trained to be able to debug, deploy, write PRDs and tests and a lot more, basically supporting the entire development life cycle. For non-programmers, they showed off tasks like a set of financial advice slides, a retail training document, and a fashion presentation deck. Now, interestingly, on GDP Val, Codex 5.3 was very similar to previous models. But on the OSWorld benchmark, which measures compute use in real world tasks, Codex 5.3 scored 64.7%, which almost doubles the performance of GPT-5.2. Summing up, OpenAI wrote, 'With GPT-5.3, Codex is moving beyond writing code to using it as a tool to operate a computer and complete work end to end. By pushing the frontier of what a coding agent can do, we're also unlocking a broader class of knowledge work from building and deploying software to researching, analyzing, and executing complex tasks. What started as a focus on being the best coding agent has become the foundation for a more general collaborator on the computer, expanding both who can build and what's possible with Codex.'
Basically summing up, you got two labs which are both reaffirming everything we talked about on the 'Code AGI is Functional AGI' episode, that by expanding the capability set around coding, it unlocks use cases that are far beyond coding and are core to economically valuable knowledge work. So what were the first impressions? The companies that had early access to these models, big surprise, liked them. AJ Orbach from Triple Whale said, 'We've had early access to Claude Opus 4.6 and have been testing it over the past week. It's the best model in the world for front-end report design, ad anatomy, and copywriting, and orchestrating other tools to actually build creatives. What's kind of crazy is how good these models are getting at long-running agentic work.' Boxes and Levy said that overall, Opus 4.6 represented a 10% jump over Opus 4.5 on their hardest knowledge work tasks. In terms of the features that people were most excited about, on the Opus side, many, many focused on this 1 million token context window. Now, as Mano Ventures DD Doss pointed out, part of why people were excited about this is that it came with, as he put it, insane state-of-the-art performance on all the long context benchmarks. In other words, this isn't one of those claims of a long token context window where functionally it actually doesn't work. Others honed in quickly on the Swarm/Team mode. McKay Wrigley shared a test between Opus 4.6 with the teams mode and 4.6 without it and found that the team's mode was 2.5 times faster and done better. He also pointed out, 'Reminder that swarms is available in the Claude Agent SDK as well. You can build swarms into any product literally right now.' Kieran Clawson from Every thinks the implications are even bigger. He wrote, 'Been running agent swarms for a few weeks now. I think this is the future, but I'm relearning what feature development even means.' On the 5.3 Codex side, the things that people were taking note of were things like the new token efficiency. Andy Henny writes, 'The biggest 5.3 Codex news is that it's roughly three times more token efficient. 5.3 high is smarter than 5.2 high, but uses one-third the tokens, making it faster and letting your weekly limit last about three times longer. Good job.' But of course, the real question is which of these models is better? Perennial early tester Simon Willis kind of shrugged his shoulders. In his blog post about them, he said, 'I don't have much interesting to say about these models yet, to be honest. They're both incremental improvements on their predecessors and very capable.' Maybe my favorite tweet, which maybe you'll get right away, but did take me a second, came from the prime agent who wrote, 'I cannot believe how much better 5.3 is than 4.6. After some internal testing, results show it's 15.2% better.' Now, rather than explain, I'll let you try to work that one out for yourselves.
Some people did try to actually make a comparison, though, and many did it with their own benchmarks. Neil Chudley writes, 'TLDR: Opus 4.6, 1 million context, enterprise and knowledge work, agent teams, and Claude Code not benching as high as Codex 5.3,' although he does point out that he doesn't really care about self-reported benchmarks. On 5.3 Codex he writes, 'Wins code benchmarks, faster mid-task steering, but less than half the context window of Opus.' His conclusion: 'Going to have to try Codex, I guess.' With things still pretty fresh on a performance side, Latent Space focused on the blood sport that model releases have become. They write, 'If you think the simultaneous release of Claude Opus 4.6 and GPT-5.3 Codex is sheer coincidence, you're not sufficiently appreciating the intensity of the competition between the leading two coding model labs in the world right now.' Now, ultimately, they gave this round to Anthropic from a developer attention perspective, getting out ahead of OpenAI and offering a huge range of new features to try out and a $50 credit for tinkering away this weekend. At the same time, they noted that OpenAI won across most benchmarks and delivered a model that was 25% faster than their previous generation and has higher token efficiency. Their head-to-head comparison, nent though it is, showed OpenAI was stronger in coding and speed while Opus had the edge on agent orchestration and long context tasks. However, and I think this is extremely important if you're trying to have a definitive answer right now, they noted that the pair of models are so close that it's likely that, quote, 'all first-party reactions are either biased or superficial.'
This idea that of Opus for orchestration, Codex for coding is something that I've been seeing even for the last couple of weeks though, and it'll be interesting to see if with the release of these new models that framework or narrative reinforces itself among developers. When Vibe Code apps Riley Brown asked what the vibe check is so far and how people were feeling about 5.3 and 4.6, a lot of the comments were basically about how OpenAI's Codex models had gotten really good and that people who hadn't tried them because they were so hooked on Claude Code really owed it to themselves to give them a try. Still, when developer Ryan Carson asked, 'Which model are you going to code with this week?' out of more than 700 voters, 53.3% said Opus 4.6 compared to 24.9% for Codex 5.3, suggesting that Anthropic still has the developer devotion. The team at Every had early access to both models with Dan Shipper's big thesis being that the models are converging. He writes, 'Opus 4.6 has all the things we love about 4.5, but with the thorough, precise style that made Codex the go-to for hard coding tasks, and Codex 5.3 is still a powerful workhorse, but it finally picked up some of Opus' warmth, speed, and willingness to just do things without asking permission. From this, we can only conclude that both labs are moving steadily towards a sort of coding model. One that's wicked smart, highly technical, and fast, creative, and pleasant to work with.' Why the convergence? Because Dan writes, 'A great coding agent turns out to be the basis for a great general-purpose work agent. The behaviors that make AI useful for software development—parallel execution, tool use, planning before asking, knowing when to dig deep versus when to ship—are the same behaviors that make AI useful for any knowledge work. And that is the holy grail of AI.'
Now, one of the things that I will be watching closely is not the day one or even the week one reactions. Remember, after we got Opus 4.5 and GPT-5.2 Codex, it took more than a month and everyone going home for the holidays for people to really fully appreciate just how much the capability set has shifted. Some are reporting an early continuation of that, having to relearn how to do things. Anthropic's Alex Albert writes, 'The jump in autonomy is real. The biggest shift for me personally has been learning to let it run, give it the context, step away, and come back to something pretty amazing. The way we work alongside models is starting to completely change.' OpenAI is in fact throwing down a gauntlet, trying to encapsulate this shift. OpenAI President Greg Brockman writes, 'Software development is undergoing a renaissance in front of our eyes. If you haven't used the tools recently, you're likely underestimating what you're missing. Since December, there's been a step function improvement in what tools like Codex can do. Some great engineers at OpenAI yesterday told me that their job has fundamentally changed since December. Prior to then, they could use Codex for unit tests. Now it writes essentially all the code and does a great deal of their operations in debugging. Not everyone yet has made that leap, but it's usually because of factors besides the capability of the model. Every company,' Greg continues, 'faces the same opportunity now and navigating it well requires careful thought.' And so the rest of the post then he says is about how OpenAI is retooling their teams towards agentic software development. As a first step, he writes, 'By March 31st, we're aiming that one, for any technical task, the tool of first resort for humans is interacting with an agent rather than using an editor or terminal. Two, the default way humans utilize agents is explicitly evaluated as safe, but also productive enough that most workflows do not need additional permissions.' Greg Brockman then is saying that in less than two months, agent-first is the way that technical teams will build at OpenAI. If that doesn't put a starting gun on how we all should be thinking about changing our own workflows, I don't know what will.
Ultimately, it is a very exciting moment. One which, as we'll see on our next episode, even has people rethinking the whole idea of an AI bubble. For now, though, that is going to do it for today's AI Daily Brief. Appreciate you listening or watching as always and until next time, peace.
