▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=Xdxp3lbQKyQ
• Original title :: Ox Alpha is INSANE

The anonymous "Ox Alpha" model that appeared on OpenRouter and OpenCode turned out to be GLM 5.3 Flash from Z.AI. Theo argues it is the first small, cheap model that is genuinely pleasant to work with: not very smart, but extremely well behaved in agentic loops, at roughly a tenth to a hundredth the price of frontier models. It audited hundreds of pull requests in his repo for about twelve cents.

The anonymous drop

The model was announced with capacity for 100 trillion tokens per day, a number hard to fathom (all Gemini traffic is estimated around 150 trillion/day). That hinted at either hidden compute reserves or a very small, cheap model. Early private benchmarks on a public subset of SWE-style problems put it around 80% where other well-known models scored 52-65%, which Theo notes is misleading but not evidence the model is bad.

Releasing it anonymously and for free was, in his view, a smart launch: he would have skipped the blog post, but the free anonymous drop made him try it and be surprised. It became the top model on OpenRouter and in OpenCode during the testing window.

What it actually is

GLM 5.3 Flash: 320B total parameters, ~18B active per expert, a million-token context, and full multimodality (image, audio, video) - the non-Flash 5.3 could not even take a screenshot. A hybrid sparse/linear attention architecture keeps long-context serving cheap, so the price does not step up as the context grows. Open weights, so it can be self-hosted. Pricing at the time of recording was about 7.5 cents per million input tokens and 25 cents per million output - Gemini Flash 2 territory.

The reason Z.AI could give away so many tokens: the model runs entirely on Huawei Ascend 910 chips rather than Nvidia, heavily optimized (probably at some cost to throughput) around memory limits.

Intelligence versus behaviour

Theo's central framing is that model capability splits into two axes, not one:

• Intelligence :: how much knowledge is baked in, how well hard problems are reasoned through. GPT-4.5 and Gemini 3.1 Pro are the extreme cases - enormous knowledge, poor at staying on task. Gemini 3.1 Pro still tops his private "skate bench" (95-97%, the only model in the nineties) yet falls apart when asked to actually fix code.
• Behaviour :: how faithfully the model applies what it is told, stays on task, absorbs mid-task steering, and unblocks itself. This is what agentic work really measures.

GLM 5.3 Flash is the first model he has used that is genuinely stupid on the intelligence axis and genuinely excellent on the behaviour axis. He credits Z.AI's heavy RL post-training in provisioned environments where the model does real work.

The PR audit run

He wired the model into his own Codex fork via an OpenRouter binding and asked it to review the open PRs in the T3 Code repo, then mid-task redirected it ("ignore PRs made by Codex, only the ones from Fable and Claude Code") and added another instruction ("put the findings in HTML so I can read them on my phone"). Most open-weight agentic models lose the thread or stop early on that kind of steering; this one folded both in and kept going.

On an earlier run it spun up six subagents, three of which failed with provider auth errors. Rather than stalling, it noticed the failures were consistent, abandoned delegation, and completed the audit directly from GitHub metadata, diffs, reviews, CI comments, and local source. Theo calls the self-debugging the most impressive part.

It ranked hundreds of PRs by ease of merge and confidence, included real links rather than bare PR numbers (a small thing frontier models routinely get wrong), noticed which issues the merged PRs closed, and surfaced a real bug fix he would otherwise have missed. Total cost for auditing over a thousand PRs: 12 cents. The equivalent run on Fable, over a narrower set, had cost him more than $100.

Benchmarks and weak spots

Across Z.AI's published comparisons it sits neck-and-neck with Opus 4.8 and 5.6-tier models at under a tenth the price, and crushes the cost/intelligence Pareto frontier. On one coding bench it does the same work in ~70k output tokens that Opus 4.8 needs ~120k for.

The real complaint is token efficiency in the broader sense: ~47k tokens per task on artificial analysis, against ~20k for Luna on max and ~17k for Soul. That matters not because of price but because bloated context means slower runs and a higher chance of hitting the context limit before the task is done. Theo expects efficiency to improve now that the training run exists.

Creative and visual work: decent front-end output when given a design skill, surprisingly non-slop marketing copy, working animations, and a Blender scene the Z.AI team let it build over 12 hours that looks genuinely good. His "fish slop" game port worked and felt solid to control but carried logic bugs (fish stop moving after eating, eating when not hungry) - small-model errors on top of surprisingly good detail work.

Where it lands

On his tier list, GLM 5.3 Flash moves up next to Soul, and DeepSeek V4 Flash moves down - the latter has no vision in any downloadable form (the vision endpoint is an API feature, not a model capability) and drifts on long-running tasks. He argues that between Fable, Soul, and 5.3 Flash there is practically no reason to use anything else on the list. Against Muse Spark, 5.3 Flash matches the price without requiring you to hand over your code on the contributor tier.

The takeaway: at a fixed level of intelligence, cost falls exponentially; what is getting more expensive is more intelligence. For cheap, thorough, well-behaved agentic work that does exactly what you asked, this model is worth trying.