Just yesterday Anthropic with their latest model, Opus 4.6 and OpenAI released Codex 5.3. OpenAI’s newest Codex model reflects the company’s increasing focus on winning over developers, delivering a massive increase in coding benchmarks. On Terminal Bench 2.0 Codex scored a 77.3, a huge improvement from 5.2’s 64. It is also much higher than Opus 4.6 which has a 65.4. I have been a Claude Code user for the past 6 months but seeing these benchmarks is definitely making me interested in trying Codex.

If you are using Claude Code today, you are already using Opus 4.6. Personally, I have’t felt a huge difference. Benchmarks don’t mean everything but 4.5’s Terminal Bench score was already 59.8 so the improvement here is nowhere close we are seeing with Codex.

Frontier appears as an Enterprise focused product, allowing for the creation of specialized agent “coworkers” that perform specialized tasks and functions. It looks like a cleaned up, closed-source version of OpenClaw.

In their announcement page they say that companies purchasing Frontier will get Forward Deployed Engineers to help them integrate it. They list a few examples of problems they’ve solved with this system but details are vague on how much the Frontier platform was actually used and who the work was done for. For example one of the solutions they say they made is a “user personalization” tool that basically just reads as a vector search pipeline deployed on a website’s keyword search bar.

I don’t think it’s time for specialized AI agent solutions to worry about this platform yet but time will tell.

OpenClaw costs a lot to run and you can plug in any model capable of tool calls to use with it. I introduced a new leaderboard on Price Per Token where anyone can vote on their favorite models use. Check it out if you need some guidance on models to try!

Alex