Nex-N2.6 Swaps Qwen for GLM: Nex-AGI's New Agent Model Costs More and Says Less

Nex-N2.6 went live on Nex-AGI's own commercial API on 10 November 2026 with no launch post, no readable model card and no benchmark table. What it did arrive with is a base-model change: where Nex-N2.5 Pro was post-trained on a Q​wen3.5 mixture-of-experts checkpoint, Nex-N2.6 is post-trained on GLM-5.3-Flash, Z.ai's 320-billion-parameter multimodal model, and the result carries roughly 321 billion parameters of its own. It also arrived at a higher price than the model it supersedes — $0.12 per million input tokens and $0.40 per million output, against Nex-N2.5 Pro's $0.075 and $0.25, a 60% increase on both sides of the meter. For a family that has spent two releases arguing it delivers frontier agent behaviour at open-model prices, a successor that costs more and documents less is the interesting part of this launch, not the launch itself. Nex-N2.6 is a single-endpoint model, and that is the operational risk described above. OrcaRouter carries 200+ models behind one OpenAI-compatible endpoint with automatic failover, so the fallback path you keep for an unmeasured model can be a routing rule rather than a second integration — and when we do host a model, the provider's list price is passed through at zero markup.

What is actually there to see

The model is not a rumour. It is callable right now, served by Nex-AGI itself at fp8 precision, and the endpoint reports clean health — 100% uptime over the windows the catalogue measures, around 35 output tokens per second, and a p50 latency of 2.81 seconds. Everything below comes from that listing rather than from the vendor, because the vendor has published nothing else:

• Price — $0.12 input / $0.40 output per million tokens, cache reads at $0.024. The listing shows that as 20% off a $0.15 / $0.50 struck-through list, so treat the current number as an introductory rate with an unstated expiry.

• Context — 262,144 tokens, the same serving window as every model in the N2.5 family, with a maximum output of 235,929 tokens.

• Modality — text and image in, text out. There is no audio or video channel, and no image generation.

• Reasoning — mandatory, not optional. The endpoint accepts only two effort settings, high and medium, and defaults to high. You cannot turn thinking off on this model.

• Tooling — function calling, forced tool choice, structured outputs and logprobs are all supported, which is the right set for an agent model and the one thing about this release that is fully specified.

• Weights — the listing links a Hugging Face repository for Nex-N2.6. That repository returns an authentication prompt rather than a public model card, so the weights, the config and any vendor evaluation tables are not readable as of publication.

The base swap is the real change

Model families usually iterate by scaling the same post-training recipe. Nex-AGI did something more disruptive: it changed the foundation. Nex-N2.5 Pro, released 8 September, was built on a 397-billion-parameter Q​wen3.5 mixture-of-experts base with about 17 billion parameters active per token — the architecture and tokenizer are visibly Q​wen-derived, and its card says so. Nex-N2.6 is built on GLM-5.3-Flash, a different lab's model, a different tokenizer, and a different parameter budget: Z.ai's card describes 320 billion total with 18 billion active, built on a hybrid sparse-and-linear attention design rather than the Q​wen stack.

That matters beyond trivia. A changed base means the tool-call parser, the chat template, the tokenizer and the failure modes all changed with it. Prompts tuned against Nex-N2.5 Pro — especially long system prompts that teach a specific tool-call dialect, and anything that leans on Q​wen's token boundaries to control cost — are not transferable by assumption. The one behavioural claim Nex-AGI makes is precisely about this interface: the model, it says, reduces repetitive reasoning and follows the tools available in the current environment more closely. That is a plausible thing to improve when you re-base onto a model that was itself tuned for agentic work, and it is also exactly the kind of claim that only shows up in your own traces, not in a benchmark table. Vendor-reported, unreproduced, and stated in one sentence.

The second-order effect is cost predictability. Nex-N2.5 Pro could be run with reasoning disabled entirely, which made it cheap for short, mechanical turns. Nex-N2.6 cannot. Every call thinks, at high effort by default. On an agent loop that fires forty small tool calls to find a file, that difference in default behaviour will move your bill more than the 60% headline rate does.

A successor that costs 60% more

The pricing direction is worth pausing on, because it inverts the usual generational story. Per million tokens: Nex-N2.5 Pro at $0.075 in / $0.25 out, Nex-N2.6 at $0.12 in / $0.40 out. Same context window. Same modality. A newer model from the same lab, on a smaller base, at 1.6 times the rate.

Run it through a realistic agent workload — ten million input tokens and two million output tokens a month, which is one moderate coding agent on a busy team — and the numbers are $1.20 plus $0.80, so $2.00 a month on Nex-N2.6. The same workload on Nex-N2.5 Pro is $0.75 plus $0.50, or $1.25. Both are rounding errors next to the frontier models. That is the honest read: Nex-N2.6 is more expensive than its predecessor and still roughly one-sixteenth the cost of a flagship, so the price rise is a signal about positioning rather than a budget problem.

Where the rise does bite is at volume, and where the mandatory reasoning bites is in the tail. A model that always thinks at high effort spends output tokens you did not ask for, and $0.40 per million output tokens on a reasoning-heavy loop is a different number from $0.40 on short answers. Measure tokens per completed task, not price per token, before concluding anything.

One structural note on how these rates travel: a gateway that passes provider list price through at 0% markup has no separate rate card to reconcile, so a vendor cut or a discount lapsing shows up the same day rather than at the next contract renewal. OrcaRouter works that way. It is also the honest place to say what we do not have — no Nex-AGI checkpoint is in the OrcaRouter catalogue today, so Nex-N2.6 is a model you call directly from one provider at one price. That single-provider shape is the actual operational risk in this release, and it is worth planning around even if you never touch our product: a model with one endpoint, no published weights and no second host has nowhere to fail over to.

The evidence gap is the part to price in

Nex-AGI's previous release came with a full apparatus: a vendor site with benchmark charts, a Hugging Face card with text and multimodal tables comparing the family against Claude Opus 5, GPT-5.6 Sol, Kimi-K3 and Qwen3.8-Max, and a documented serving recipe on eight H100s. Nex-N2.6 came with a paragraph.

Check what is missing and it is a consistent list. The vendor's own website still advertises the Nex-N2.5 family and does not mention N2.6 in any form we could load. The GitHub organisation has repositories for Nex-N2.5, NexRT and the NexAU agent framework, and none for N2.6. ModelScope returns no record for the checkpoint. The Hugging Face repository named in the listing is not publicly readable. There is no entry on Artificial Analysis, so there is no independent score at all — not a weak one, none. As of publication, the total public evidence base for this model is a price, a context length, a modality, a parameter count and one sentence of positioning.

That is not the same as a bad model. It is a model whose claims you cannot check, from a lab that has previously shipped the checks. The right posture is the one you would take with any unmeasured dependency: try it on work you can grade yourself, keep the previous model wired up as a fallback, and do not move a production path onto it because the version number went up.

Who should move, and who should wait

There is a real case for switching now, and it is narrow. If your agent harness spends its budget fighting the model rather than the task — re-reading the same file, re-deriving a conclusion it already reached, or emitting tool calls in a dialect your schema rejects — then a model explicitly tuned to cut repetitive reasoning and honour the available tools is aimed directly at your problem, and at $0.12/$0.40 the experiment costs almost nothing to run. Image input plus function calling in one endpoint also makes it usable for screenshot-driven QA without a second model in the loop.

The case for waiting is broader. If your workload is already working on Nex-N2.5 Pro, the base swap alone is a migration, and you would be doing it for a model that costs more per token, cannot be run with reasoning off, cannot be downloaded, and has no published score. There is no evidence yet that N2.6 is better at anything except the two behaviours its author names.

What would change the picture, in rough order of usefulness: the model card opening with its own benchmark tables, a first independent score appearing on a public leaderboard, the weights actually publishing under a licence you can use, and a second provider hosting it so the single-endpoint risk goes away. Until at least two of those land, Nex-N2.6 is a promising unmeasured option rather than a recommended upgrade — which, given the price rise, is the whole story.

What the base change looks like on paper

The comparison that matters is not Nex-N2.6 against a frontier model. It is Nex-N2.6 against the checkpoint it replaces, because that is the migration anyone already on this family is being asked to make.

Read down that list and the pattern is a model that got smaller, more expensive, less configurable and less inspectable in a single step. That is not automatically a worse trade — a 321-billion-parameter base tuned for agentic tool use can beat a 397-billion-parameter one that was tuned for something else, and the per-token price is small in absolute terms. But it is a trade, and the vendor has not published the comparison that would let anyone judge it. Until that changes, the base swap is the story and the benchmark table is the missing chapter.