Nex-N2.6 vs Nex-N2.5 Pro: A Smaller Base, a Higher Price, and a Migration Nobody Has Benchmarked

Two months separate Nex-N2.5 Pro and Nex-N2.6, and the newer model is the smaller one. Nex-N2.5 Pro, which Nex-AGI shipped on 8 September 2026, is a 397-billion-parameter mixture-of-experts model post-trained from a Q​wen3.5 base, licensed Apache-2.0, with its 122 weight shards sitting on Hugging Face for anyone to download. Nex-N2.6, live since 10 November 2026, is post-trained from GLM-5.3-Flash, carries roughly 321 billion parameters, lists at $0.12 / $0.40 per million tokens against its predecessor's $0.075 / $0.25, and points at a Hugging Face repository that will not open without credentials. Same 262,144-token window. Same text-and-image input. A different foundation underneath, a bill 60% higher, and — so far — not one published number comparing the two.

If you are already running Nex-N2.5 Pro, that combination should make you cautious rather than curious. This is not a drop-in version bump. It is a re-basing onto another lab's architecture, sold at a premium, with the evidence that would justify it still unpublished.

What the version number hides

Version numbers imply continuity. In this family the number is doing something else: N2.5 Pro and N2.6 share a serving window and an input modality, and almost nothing below that.

• Foundation — Nex-N2.5 Pro is post-trained from a Q​wen3.5 mixture-of-experts checkpoint, Q​wen tokenizer and all; Nex-N2.6 from GLM-5.3-Flash, Z.ai's 320-billion-parameter natively multimodal model with a hybrid sparse-and-linear attention design. Different lab, different tokenizer, different chat template, different tool-call parser.

• Size — 397 billion total parameters with about 17 billion active per token, versus roughly 321 billion total with an active count Nex-AGI has not published for N2.6.

• Price per million tokens — $0.075 in / $0.25 out for Pro; $0.12 in / $0.40 out for N2.6, which the listing presents as 20% off a $0.15 / $0.50 list rate. So the eventual price may be double, not merely 60% higher.

• Reasoning — Pro accepts none, medium and high effort and defaults to high, so you can switch thinking off for cheap mechanical turns. N2.6 accepts high and medium only, defaults to high, and cannot be turned off at all.

• Weights — Pro is Apache-2.0 on Hugging Face, fully downloadable. N2.6's linked repository is not publicly readable.

• Vendor benchmarks — Pro has a full published table across coding, agentic and multimodal suites. N2.6 has one sentence: it reduces repetitive reasoning and follows the tools available in the current environment more closely than Pro.

The one number the vendor did give

That sentence is the entire comparative claim, and it is a specific kind of claim: it describes tool-call discipline, not capability. Read it charitably and it says N2.6 spends fewer tokens re-deriving what it already knows and pays more attention to the tools you actually exposed. Read it sceptically and it says the model was tuned for a failure mode the vendor observed, without showing the observation.

For anyone running an agent, that distinction matters more than a benchmark point. The expensive failure in a long-horizon agent loop is rarely a wrong answer; it is a correct answer reached after nineteen redundant reasoning steps and four wasted tool calls. A model that genuinely reduces repetition can be worth more than its 60% premium even if it scores identically on a static eval, because the bill you pay is tokens per completed task, not tokens per call.

But the corollary is uncomfortable. Nex-N2.6's mandatory reasoning interacts directly with that claim: a model that always thinks at high effort spends output tokens on every turn, including the trivial ones. If Pro with reasoning disabled handled your short calls for a fraction of a cent and N2.6 must think through each one, the repetition savings have to be substantial just to break even. Whether they are is exactly the thing nobody has measured.

What Nex-N2.5 Pro's published table still tells you

The older model's numbers are vendor-reported — Nex-AGI's own harness, its own prompt settings, no independent replication — and they should be read as directional. But they exist, they are specific, and they establish a floor that N2.6 has not cleared in public:

• Coding — Terminal-Bench 2.1 at 82.7, SWE-Bench Pro at 61.2, DeepSWE v1.1 at 55.8.

• Agentic — AutomationBench v1.0.6 at 44.2, Toolathlon Verified at 68.5, BrowseComp at 89.7, Job Bench at 41.4.

• Computer use — OSWorld-Verified at 82.2, OSWorld-2 at 56.4, OSWorld-G at 87.4, WebArena-Verified at 67.6.

• Knowledge work — GDPval-AA v2 at 1,628 Elo.

Set against the comparison columns Nex-AGI printed beside them, the shape is informative: Pro trails Claude Opus 5 on most rows but sits within a few points of GPT-5.6 Sol and Kimi-K3 on tool-use suites, and it beats Qwen3.8-Max on OSWorld-2. That is a credible open-weight agent model, not a leader. N2.6 has to beat it to be worth a migration, and there is currently no page on which to check.

One consequence worth stating plainly: because Pro's weights are downloadable under Apache-2.0, you can run its evaluation yourself, on your tasks, without asking anyone's permission. That is not true of N2.6 in any sense — not the weights, not the card, not a public score. For a team making a six-month infrastructure bet, verifiability is a feature, and this release moved backwards on it.

Migration mechanics, honestly

Switching is a base URL and a model string away, which is why it is tempting to treat as free. It is not. A changed tokenizer means token counts move, so both your cost model and any context-budget logic that counts tokens are now approximate until re-measured. A changed chat template means the system prompt that reliably taught Pro its tool dialect may teach N2.6 something else. And a mandatory reasoning mode means the latency profile changes: an agent loop that could answer a trivial "which file contains this symbol" turn instantly on Pro now thinks first, at high effort, by default.

The practical way to do this is a shadow run rather than a switch. Send the same real traffic to both models for a week, grade the outcomes with something that is not the model itself, and measure tokens per completed task rather than latency or price per token. Whatever the result, keep Pro reachable — it is the only one of the two you can also self-host, and a downloadable Apache-2.0 checkpoint is a fallback that no hosted-only model can offer.

This is also the point where a gateway earns its place rather than being a convenience. OrcaRouter takes the model behind one OpenAI-compatible endpoint and lets you define a fallback chain in the routing DSL — cheap model first, escalate by rule rather than by redeploy — so testing N2.6 against Pro costs a config change instead of a code change, and reverting costs another. Both are, to be clear, models we do not host: neither Nex-N2.5 Pro nor Nex-N2.6 is in the OrcaRouter catalogue today, so in practice this means keeping your own keys and letting the routing layer hold both paths. The arithmetic of the fallback is the same either way.

Where the cheaper model is now the wrong choice

For most of this family's users the migration case is weak, and it is worth saying so directly. If Pro is working, the only reasons to move are the two behaviours Nex-AGI names — less repetition and tighter tool adherence — and you cannot verify either before you try it. Everything else about the move is neutral or negative: a smaller base with an unpublished active-parameter count, a higher price that may double when the discount lapses, no way to disable reasoning, and a model card you cannot read.

There is a narrow but real case on the other side. Teams whose agents are bottlenecked on wasted turns, whose prompts are short enough that the 262K window is not a constraint, and who have a grading harness good enough to tell a real improvement from noise can profit from moving early — the absolute prices are low enough that the experiment costs less than the meeting about it. The same applies to anyone who wants a GLM-family tokenizer and tool-call behaviour behind an agent loop; N2.6 is a post-trained GLM-5.3-Flash with agentic tuning, which is a different proposition from the raw base model.

What would settle it

Three publications would turn this from a guess into a comparison, and none requires anything exotic. First, Nex-N2.6's own card with the same table Nex-AGI printed for the N2.5 family — the vendor already knows how to publish one. Second, the weights under a licence you can use, which would let the community run Pro and N2.6 side by side on equal footing instead of trusting a sentence. Third, a first independent score on a public leaderboard, even a mediocre one, because a mediocre measured model is more useful than an unmeasured promising one.

Until then

The honest summary is that Nex-AGI shipped a re-based, more expensive, less documented successor to a model that is still available, still cheaper, still downloadable, and still the only one of the two with numbers attached. That is not a reason to ignore N2.6 — early access to a model tuned for tool discipline is worth something to the right team — but it is a reason not to migrate a working production path on the strength of a version number and a discount.

If you want a rule: switch when your own traces show the problem Nex-N2.6 claims to fix, keep Nex-N2.5 Pro deployed as the fallback either way, and re-read the price the day the 20% lapses. The successor's case rests entirely on behaviour the vendor described and nobody has measured.