Nex-N2-Pro scores like a frontier coder and bills like a budget one. The free weights are the real story.

What Nex AGI actually shipped

Nex AGI is not a name most people had a month ago. The lab put Nex-N2-Pro on Hugging Face and GitHub in early June, and it surfaced on Artificial Analysis dated June 2. The model is a mixture-of-experts design: 397 billion total parameters, 17 billion active per token. That active count is the number that matters for inference cost, and 17B is small, which is why the hosted price can sit where it does.

The base is borrowed. Nex AGI says the model is post-trained on a Qwen3.5-397B-A17B foundation, so the heavy pre-training bill was someone else's. There is a smaller sibling too, Nex-N2-mini at 35B total and 3B active, for people who want to run something local without a GPU cluster. Both ship under Apache 2.0, which is the permissive end of the license spectrum: you can use the weights commercially, modify them, and serve them without asking anyone.

On the practical specs, Nex-N2-Pro carries a 262K context window, accepts text and image input, and outputs text. It is built for coding and agentic work rather than chat, which shows up in both the benchmark sheet and the way the lab talks about it.

What a month of it costs

Sticker prices are abstract until you attach them to a workload. Here are three coding profiles run through Nex-N2-Pro, the cheaper DeepSeek V4-Pro, the open-weight peer GLM-5.2, and GPT-5.5 as the frontier anchor. All flat-rate, no cache discount applied.

Profiles assume a coding mix where input outweighs output roughly 3.5:1 to 4:1. Nex-N2-Pro column accented for reference, not because it wins.

The against-GPT-5.5 story holds at every size. On the volume row, $175 against $1,950 is an 11x gap, and the multiple barely moves whether you are a hobbyist or a team because the ratio between the two rate cards is fixed: a tenth of the input, a twelfth of the output.

The against-DeepSeek story is the one Nex AGI would rather you not run. DeepSeek V4-Pro is cheaper in every row, by a wide margin on the team profile, $100 against $175. So if your only goal is the lowest possible bill on text-only coding, Nex-N2-Pro is not the answer. Its case rests on what it scores, and on the license.

The benchmark reality

This is where the "rivals GPT-5.5" framing needs trimming. Nex AGI's own card reports an 80.8 on SWE-Bench Verified, a 75.3 on Terminal-Bench 2.1, and a 90.7 on GPQA Diamond. Those are genuinely strong coding and reasoning numbers, and on raw SWE-Bench Verified the 80.8 lands a hair above DeepSeek V4-Pro's self-reported 80.6.

Then look at the one number that is normalized across models. Artificial Analysis puts Nex-N2-Pro's Intelligence Index at 41, which had it sitting around seventh on the board in late June (that ranking moves as the index re-runs). That is excellent for the price. It is also clearly a step below GPT-5.5, which held the top of that same index back in April. A high SWE-Bench Verified score on the vendor's own harness and a top-of-the-board normalized index are not the same claim, and Nex-N2-Pro has the first, not the second.

The honest summary: this is a model that will clear a lot of real coding evals at a tenth of frontier cost, not a model that quietly matches the best closed system for free. At this price, the first thing is usually enough.

The catch: renting it is the shaky part

When Nex-N2-Pro first showed up it had a free slug on OpenRouter and a cut-rate hosted tier around $0.25 input and $1.00 output, cheaper than DeepSeek. That is the version that went a little viral. It is also the version that is already gone: OpenRouter shows the model deprecated around June 22, roughly two weeks after it landed, and the free window closed with it.

This is the recurring trap with models from small labs. The launch-week hosting is a promotion, sometimes a single provider eating cost for attention, and it does not survive contact with a real inference bill. Build a pipeline against a $0.25 endpoint that vanishes in a fortnight and you are re-platforming, not saving money.

Which is exactly why the Apache 2.0 license is the headline and the rate card is the footnote. The weights are not going anywhere. If Nex-N2-Pro fits your evals, the plan that holds is to run it yourself, on rented GPUs or your own, where the only meter is hardware. The hosted reference rate of $0.50/$2.50 is a useful ceiling for budgeting, but treat any cheaper hosted offer as temporary until proven otherwise.

Who should care

The strongest reason to look here is self-hosting. An Apache 2.0 model with an 80.8 SWE-Bench Verified and a 17B active footprint is cheap to serve and unencumbered, so if data control or getting off per-token billing is the point, Nex-N2-Pro earns a spot on the shortlist next to the other open coders.

If you just want the cheapest hosted coder and nothing more, start with DeepSeek V4-Pro instead. It comes in lower, its context runs to 1M, and its API has not vanished on anyone. Nex-N2-Pro only edges past it when the open license or image input is something you actually need.

And if you are still paying GPT-5.5 rates for routine coding, the 11x gap is worth a weekend of testing whichever budget model you land on. The savings come from leaving the premium tier. They do not come from picking flawlessly among the cheap models once you have left it.

Sources

• - Nex-N2-Pro model card and benchmarks: huggingface.co/nex-agi/Nex-N2-Pro
• - Nex-N2 repo (architecture, base model): github.com/nex-agi/Nex-N2
• - Pricing, AA Intelligence Index, throughput: artificialanalysis.ai/models/nex-n2-pro
• - Hosted rate and deprecation status: openrouter.ai/nex-agi/nex-n2-pro
• - Competitor list prices (DeepSeek, GLM, GPT-5.5): tokencost.app/pricing