Open Source Is Closing the Gap Faster Than Anyone Expected.

Open source models are nowhere close to flagship models like Opus and Gemini — yeah, in 2024.

Now, in mid-2026, the scenario is a lot different.

I’ll be upfront about why I’m writing this. I’d been using Claude Code for a couple of months as my daily driver — it’s genuinely great, no complaints. But on a whim I decided to try out the GLM Coding Plan just to see what the hype was about. And honestly? The results were too good for me to not write about it.

This blog is for the rest of public — students, freelancers or anyone who’s just curious about how far open source has actually come. If you’re on a tight budget and want frontier-adjacent coding capability without the frontier price tag, keep reading.

Back in 2024 and even most of 2025, comparing open source models to closed frontier models felt almost unfair. Claude, GPT, and Gemini had years of head start, insane compute budgets, and RLHF pipelines that open labs simply couldn’t replicate on a startup budget. Every open release would get compared to whatever Anthropic or OpenAI shipped, and it would lose. Understandably so — you can’t out-resource a hyperscaler with community compute and a handful of GPUs.

Then DeepSeek R1 happened, and it kind of broke the assumption everyone was operating on.

It wasn’t just that R1 was good — it was that it proved a certain level of reasoning capability could be trained without the eye-watering compute budgets everyone assumed were mandatory. That single release created an entire field of momentum that nobody, including the closed labs, really saw coming. Suddenly every Chinese lab — Zhipu (Z.ai), Alibaba (Qwen), Moonshot, and others — had both the incentive and a rough blueprint to chase frontier-level reasoning and coding performance without frontier-level budgets. And they’ve been sprinting ever since.

Benchmarks are one thing. What actually matters is cost per unit of real coding output. Here’s where things stood in late June 2026:

Pro and Verified are different, harder-vs-easier benchmark variants, so don’t read the percentages across rows as directly comparable — the point is the price gap, not a leaderboard.

Can GLM-5.2, MiniMax M3, or Kimi K2.7 directly beat Opus 4.8 on the hardest, most ambiguous engineering problems? No, not consistently. That’s usually where the genuinely gnarly stuff lives — deep architectural decisions, edge cases the model has to infer intent for.

But that’s not what most coding work actually is. The 80% of tasks that make up a normal engineering day — CRUD endpoints, failing tests, refactors, scaffolding off an existing pattern, debugging a stack trace — these three are landing reliably now, especially with well-structured prompts and decent context. Put cost into the equation and the picture flips: paying 10-25x more per token stops making sense. Per dollar, these open models aren’t just “good enough” — for that 80%, they’re often the better deal outright. The remaining 20% is where you keep a closed frontier model in the toolkit, as a targeted tool, not a full replacement.

Now, the actual money math. Claude’s $20/month Pro plan isn’t the real comparison here — it doesn’t carry enough usage headroom for daily agentic coding, so anyone doing real work is really looking at the $100/month Max plan. Put that $100 next to the GLM Coding Plan, which starts from $18/month, or against even cheaper plans for MiniMax M3 or Kimi K2.7. For an equivalent volume of daily coding tokens, you’re looking at a fraction of that $100 — often low enough that the same monthly budget covers several team members instead of one.

The other thing worth calling out: these labs aren’t just cloning closed-source architecture at a discount. They’re trying things the big closed labs likely never will, because they don’t need to take the risk.

MiniMax built its own sparse-attention mechanism (MSA) specifically to make long-context inference faster instead of just throwing more compute at it. Kimi K2.7 leans hard into native sub-agent orchestration — its “Agent Swarm” setup can coordinate up to 300 sub-agents across thousands of steps in a single session, which is a genuinely different bet on how agentic coding should work compared to a single model looping on its own. And that's not all — Kimi K2.7 ships as a full 1-trillion-parameter model, using a mixture-of-experts setup that keeps only 32 billion parameters active per token so it stays practical to run.

None of this is guaranteed to win. But it’s real architectural risk-taking, out in the open, from labs that don’t have a comfortable moat to protect. That’s arguably a bigger deal long-term than any single benchmark score.

Anthropic’s Fable and Mythos models pushed the closed-source lead back out. It is way more capable then any other model in the market. However, then US government happened & effectively became the savior of the open-source narrative here — the model that would’ve stretched the gap back out got pulled off the table before it could.

And the second point matters just as much: until Fable’s full, unrestricted capability is actually available to the public instead of gated behind a small partner program, calling it proof the gap widened again is premature. Right now it reads more like a marketing move than a genuine, provable model — and until that changes, the real story is still the one playing out in the open-weight world.

Fable or no Fable, the underlying trend hasn’t changed. A year and a half ago, “open source vs closed frontier” wasn’t a real conversation for coding — it was a foregone conclusion. Today it’s a genuine cost-performance decision.

If someone else is paying the bill, take Claude Code and don’t look back. If you’re paying out of pocket, it’s worth trying something like the GLM or Minimax Coding Plan yourself. A year ago it was a different story. Today it’s just the obvious thing to say.