Nex-N2.6 Is Live: Smaller Base, Higher Price, No Benchmarks

Nex-AGI shipped a new agent model today, and the launch reads like a dare. Nex-N2.6 went live on November 10, 2026, priced at $0.12 per million input tokens and $0.40 per million output tokens. That is 60% more than Nex-N2.5 Pro, the model it succeeds, on both sides of the meter. It also arrives with no launch post, no readable model card, and no benchmark table of any kind.

The price rise is the story, but the base swap underneath it is the bigger change. Nex-N2.5 Pro was post-trained on a Qwen3.5 mixture-of-experts checkpoint. Nex-N2.6 is post-trained on GLM-5.3-Flash, Z.ai's 320-billion-parameter multimodal model, and carries roughly 321 billion parameters of its own. A newer model from the same lab, on a smaller base, at 1.6 times the rate, with less evidence attached than its predecessor had.

What the listing actually shows

Nex-AGI has published nothing about Nex-N2.6. The vendor's own website still advertises the Nex-N2.5 family and does not mention N2.6 at all, as of this writing. There is no entry for it in the Nex-AGI GitHub organisation, no record on ModelScope, and the Hugging Face repository named in the listing returns an authentication prompt instead of a public model card. So everything below comes from the commercial listing, and it is worth saying that plainly before any of it is quoted as fact.

What the listing does establish is specific. Nex-N2.6 serves text and image input and returns text only, with no audio or video channel and no image generation. Its context window is 262,144 tokens, the same serving window as every model in the N2.5 family, with a maximum output of 235,929 tokens in a single call. It supports function calling, forced tool choice, structured outputs, and logprobs, which is the right tooling set for an agent model and the one part of this release that is fully specified. It is served at fp8 precision, and the listing 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. Those health figures are listing-reported, not independently measured.

One behaviour change matters more than the rest. Reasoning on Nex-N2.6 is mandatory, not optional. The endpoint accepts only two effort settings, high and medium, and defaults to high. You cannot turn thinking off. Nex-N2.5 Pro accepted none, medium, and high, and let you switch thinking off entirely for short, mechanical turns. On an agent loop that fires dozens of small tool calls to find a file, that difference in default behaviour will move your bill more than the 60% headline rate does, because every call now spends output tokens thinking first.

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, shipped in early September 2026, is a 397-billion-parameter mixture-of-experts model with about 17 billion parameters active per token, built on Qwen3.5, with its 122 weight shards sitting on Hugging Face under Apache-2.0 for anyone to download. 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 parameters with 18 billion active, built on a hybrid sparse-and-linear attention design, and licensed MIT.

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 Qwen's token boundaries to control cost, are not transferable by assumption. Token counts will move, so any cost model or context-budget logic that counts tokens is approximate until it is re-measured. And the latency profile changes: a turn Pro could answer instantly with reasoning off now thinks first, at high effort, by default.

The one behavioural claim Nex-AGI makes is precisely about this interface. According to the listing, the model reduces repetitive reasoning and follows the tools available in the current environment more closely than Pro. That is a plausible thing to improve when you re-base onto a model that was itself tuned for agentic work. It is also exactly the kind of claim that only shows up in your own traces, not in a benchmark table, and Nex-AGI has published no benchmark table. Read charitably, it says N2.6 spends fewer tokens re-deriving what it already knows. Read sceptically, it says the model was tuned for a failure mode the vendor observed, without showing the observation. Either way, it is one sentence, and it is the entire comparative claim.

There is also a context worth stating. Nex-AGI is the open-source agent-model ecosystem built by the Shanghai Innovation Institute with partner research groups and startups, and the N2.5 family was its first full post-training effort at trillion-parameter scale. The family earned attention honestly: the 1.6-trillion-parameter Nex-N2.5 Max topped BrowseComp at 92.6 in the vendor's own published table, ahead of Claude Opus 5 and GPT-5.6 Sol on that benchmark. A lab with that track record shipping a successor with no table at all is a choice, and it is fair to ask what the table would have shown.

What Nex-N2.5 Pro could prove, and N2.6 cannot yet

The older model's numbers are vendor-reported, from Nex-AGI's own harness and prompt settings, with 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. From the table Nex-AGI published for the N2.5 family:

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.

There is one independent reference point, and it belongs to the older generation too. Artificial Analysis lists Nex-N2-Pro, the June 2026 model on the same Qwen3.5 base, at a Coding Index of 59.1 and an Agentic Index of 23.7. No Artificial Analysis score for Nex-N2.6 has appeared yet, and no Nex-N2.6 page exists on the site as of publication. That is not the same as a bad score. It is an empty column, and an empty column is the honest description of this launch's evidence base.

One consequence is 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. Do not assume N2.6 carries the same Apache-2.0 licence as its predecessor. No licence has been published for it at all.

The price math, in plain numbers

The pricing direction inverts the usual generational story. Per million tokens, Nex-N2.5 Pro costs $0.075 in and $0.25 out. Nex-N2.6 costs $0.12 in and $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. The listing shows the current numbers as 20% off a struck-through $0.15 / $0.50 list rate, so treat the current price as an introductory rate with an unstated expiry. If that discount lapses, N2.6 at list would cost exactly double its predecessor on both sides of the meter.

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. On Nex-N2.6 that is $1.20 plus $0.80, or $2.00 a month. The same workload on Nex-N2.5 Pro is $0.75 plus $0.50, or $1.25. At list price, N2.6 would be $2.50 for the same work. Cache reads cost $0.024 per million tokens, which is 80% below the input rate, so a workload that reuses long system prompts will feel the rise less than one that does not. At November 10, 2026 market rates, verified at publication, $2.00 is about £1.49 or C$2.83, and $1.25 is about £0.93 or C$1.77.

Both numbers are rounding errors next to the frontier models, and that is the honest read: Nex-N2.6 is more expensive than its predecessor and still roughly a sixteenth of the cost of a flagship, so the price rise is a signal about positioning rather than a budget problem. 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.

The table holds one quiet surprise. Nex-N2.6's input price is 20% above Claude Haiku 5.5 and GPT-6 Luna, while its output price is 20% below theirs. For a workload with five input tokens for every output token, the two cancel out exactly: 1 million input plus 200,000 output tokens costs $0.20 on Nex-N2.6 and $0.20 on Haiku 5.5. Output-heavy work favours N2.6; input-heavy work favours Haiku. The comparison is not a clean win for either side, which is why the workload mix matters more than the headline rate. One caveat: Haiku's $0.10 / $0.50 rate applies up to 100,000 tokens per prompt, and rises above that threshold, while N2.6 charges one rate across its whole 262,144-token window.

Against its own base model, the picture is stranger still. The raw GLM-5.3-Flash lists at $0.15 in and $0.50 out through Fireworks, which is above Nex-N2.6's $0.12 / $0.40. A tuned derivative listing below its own base on a different host is a reminder that gateway rates are not vendor rates, and that comparing prices across hosts is comparing two different serving stacks. What is comparable is the direction: Nex-AGI now charges more for a model built on a base Z.ai sells cheaper, and the difference is the agentic tuning. Whether that tuning is worth the premium is precisely what nobody has measured.

The evidence gap, item by item

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.

• No launch post or announcement of any kind from Nex-AGI.
• No model card, no benchmark table, and no published active-parameter count.
• No weights: the linked Hugging Face repository is not publicly readable, and no Nex-N2.6 repository appears in the Nex-AGI organisation's public listing.
• No independent score: no Artificial Analysis entry, and no score on any public leaderboard we checked.
• No second host: the model is callable from Nex-AGI's own API and through OpenRouter, and no other provider lists it, so there is no failover path.
• No licence: nothing published, so do not assume the Apache-2.0 terms of the N2.5 family carry over.

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. 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.

One more honest note about the sourcing of this article. The detailed analysis of the release comes from OrcaRouter, a commercial model gateway that sells routing across providers. Its factual claims check out against independent records: the price, context, base, and parameter count all match OpenRouter's own listing, the missing Hugging Face repository is visible in Nex-AGI's public organisation page, and the vendor site really does show only the N2.5 family. But OrcaRouter has a commercial interest in the routing story, so treat its framing as one interested party's reading and the numbers as the part that survives independent checking. Every number in this article was checked against OpenRouter, Hugging Face, Z.ai's model card, or Nex-AGI's own site, not taken from the gateway's word alone.

Who should try it, 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. At $0.12 / $0.40 the experiment costs almost nothing to run.

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. The practical way to test it 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.

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.

Common mistakes

Treating the version number as an upgrade. N2.6 is a re-base onto another lab's architecture, sold at a premium, with the evidence that would justify it still unpublished. A higher number does not mean a better model here.

Migrating prompts blindly. The tokenizer, chat template, and tool-call parser all changed with the base. Re-test system prompts that teach Pro its tool dialect, and re-measure token counts before trusting any cost model built on Pro.

Budgeting on the intro price. The listing shows $0.12 / $0.40 as 20% off a $0.15 / $0.50 list rate, with no expiry stated. Budget at list price, and re-read the rate the day the discount lapses.

Assuming the weights are open. The N2.5 family shipped under Apache-2.0. N2.6's repository is gated and no licence has been published. Do not plan a self-hosted deployment around a model you cannot download.

Judging it on price per token. Mandatory reasoning at high effort spends output tokens on every turn, including trivial ones. The number that matters is tokens per completed task, and nobody has published it yet.

Frequently asked questions

Is Nex-N2.6 open source?

Not as far as anyone can verify. Its predecessor shipped under Apache-2.0 with downloadable weights, but Nex-N2.6's linked Hugging Face repository is not publicly readable and no licence has been published. Until the repository opens, treat it as a hosted-only model.

How much does Nex-N2.6 cost?

$0.12 per million input tokens and $0.40 per million output tokens, with cache reads at $0.024. The listing presents that as 20% off a $0.15 / $0.50 list rate, with no expiry stated, so the eventual price may be higher. At November 10, 2026 market rates, $0.12 is about £0.09 or C$0.17, and $0.40 is about £0.30 or C$0.57.

Is Nex-N2.6 better than Nex-N2.5 Pro?

Nobody has published the comparison that would answer that. Pro has a full vendor benchmark table, including BrowseComp 89.7 and OSWorld-G 87.4. N2.6 has one sentence from its author: it reduces repetitive reasoning and follows available tools more closely. That is a claim about behaviour, not a measured result.

Where can I call Nex-N2.6?

Through Nex-AGI's own commercial API and through OpenRouter, where it is listed as nex-agi/nex-n2.6 with a 262,144-token context window. No other provider lists it, so there is no second host to fail over to.

Will my Nex-N2.5 Pro prompts work on N2.6?

Probably, but not by assumption. The base changed from Qwen3.5 to GLM-5.3-Flash, which means a different tokenizer, a different chat template, and a different tool-call parser. Re-test prompts that rely on a specific tool-call dialect, and re-measure token counts, before moving production traffic.

When will benchmarks appear?

There is no published date. The three things that would settle it are the model card with Nex-AGI's own benchmark table, a first independent score on a public leaderboard, and the weights publishing under a usable licence. Until at least two of those land, the model is an unmeasured option.

Sources

• OpenRouter — Nex AGI: Nex-N2.6 model page (price $0.12 / $0.40, cache $0.024, 262,144 context, GLM-5.3-Flash base, about 321B parameters, listed November 10, 2026)
• OpenRouter — Nex AGI provider page (six Nex AGI models; Nex-N2.6 most recently added, November 10, 2026)
• OrcaRouter — Nex-N2.6: GLM-5.3-Flash Base, Higher Price, No Benchmarks (listing details, latency and uptime figures, evidence-gap audit)
• OrcaRouter — Nex-N2.6 vs Nex-N2.5 Pro: Base Swap, Price Rise, No Numbers (N2.5 Pro benchmark table, migration mechanics)
• Hugging Face — zai-org/GLM-5.3-Flash model card (320B total, 18B active, hybrid sparse-and-linear attention, MIT licence, 6,392,890 downloads)
• Nex-AGI — official site (Nex-N2.5 family only; no mention of N2.6 as of November 10, 2026)
• Artificial Analysis — Nex-N2-Pro model page (Coding Index 59.1, Agentic Index 23.7; no Nex-N2.6 page exists yet)
• Frontier Now — China's 1.6T open-source model Nex-N2.5 tops (N2.5 family background, September 2026 release, BrowseComp 92.6 for Max)
• Z.ai — GLM-5.3-Flash blog (base model announcement, linked from the official model card)
• SpaceXAI — developer pricing (Grok 4.7 $2.00 / $6.00 per million tokens; rate card read November 7, 2026)
• Exchange rates: November 10, 2026 market rates, verified at publication — 1 USD = £0.745 / C$1.413 (Finnhub and ForexInvestors). All GBP and CAD figures in this article are approximations at those rates.

Related reading on AI Core Updates: DeepSeek V4.1 Flash 1017: The $0.15 Coding Model and Its New Search API, Tencent Hy4: 770B Open-Source Model at $0.90/M — Vision, Computer Use, Weights Pending, Kimi K3.1 Pro Is Live: $2 API, 4th on AI Index, Open Weights Nov 4, Fireworks Forge: 1,500 Tokens/Second Inference and Half-Price Agents.