Look, I want to love MiniMax. I really do. And I'll keep going back to it, I'll keep trying. I'll keep giving it a chance. MiniMax M2.7 was released this week with the usual fanfare and hype that comes with an incremental release. It's 100000000x better than before, it's top in all the benchmarks, it's "almost" Opus-level. Blah blah blah. I'm a kid at heart, so when M2.7 was announced I got excited and had to give it a spin.
The most irritating thing for me is its inability to use tools. I use a self-hosted Gitea instance for Git and okay, it's not super-mainstream, it's clearly not in MiniMax's training data and it can't natively use it. But even with AGENTS.md direction, with the official MCP, with the tea CLI (which are both admittedly a little bit ... under developed) and a custom API wrapper/helper MiniMax consistently fails to fetch Gitea Issues, read comments/reviews on PRs and generally use Gitea. Which, for me, is just a massive spanner in the works.
It'll usually come back with "I couldn't read the comment/review/issue, what's the problem? Let me know and I'll help" or something. If I tell it specifically how to use the tools, it'll get there. This is something that even Haiku and Codex's mini models has handled for several versions now.
Then there's Git hygiene. One of my pet-peeves is force pushing and rebasing without good need. I like a clean history, I like auditable trails. I feel like it's absolutely essential when you're working with AI to actually have a clean history of what's happened, what changed. Claude Opus, in fairness, is also quite partial to a bit of force pushing, and I usually have to have rules against it in CLAUDE.md, but MiniMax IS SOMETHING ELSE. It just force pushes and rebases for ABSOLUTELY NO REASON.
It opens a PR, I get it reviewed, I ask it to address the feedback and every commit thereafter is a force push and a rebase. WHY? I ask it why, and apparently because "it panicked"... okay. Sure. Fine.
I haven't properly done any kind of objective test of M2.7's coding abilities yet, I just got it to fix a minor frontend bug which it did reasonably well with. After a couple of rounds of revision (which Codex 5.4 and Opus 4.6 would probably need too) we got there. But for me, its inability to figure out tools and its poor basic Git hygiene just irritate me to the point where I just don't want to use it.
For the price, okay, it's great. For the actual cost I have no complaints about it. It's clearly distilled from Claude and it's clearly trying to offer a Claude-like experience. But it isn't Claude. It isn't near-Claude. It's a Temu Claude, and that's fine. But can we stop pretending that's "Near-Opus level for a fraction of the cost" because only one part of that is true.
At the end of the day, you get what you pay for. There's no getting around that. Although in this regard, in terms of value for money, Codex wins. In terms of model capability vs cost, Codex wins every day. I just wish OpenAI would offer a middle £90-100 subscription rather than the £20-£200 jump, but that's another issue. And it absolutely will have its uses.
In the next few weeks I'm going to do an objective head-to-head between all the latest releases with a fixed task. I'll also try a project using the the MiniMax API baked into an application in some way. I am determined to find a place for it in my workflow, in the things I do. Right now, I don't know what that is. But I really, really do want to love it. I want to make use of that cheap API cost. And I'll continue to keep trying to love it, but for now, we just really, really need to stop pretending that it's "near-Opus level" because it just isn't.
