This week’s model releases are notable for expanding the hosted long-context LLM market rather than for a single headline benchmark or new modality breakthrough. Sakana AI and Inference Net both added new text-generation models to OpenRouter, giving developers more options for long-document analysis, summarization, and general assistant workloads—but also leaving important questions unanswered around pricing, output limits, benchmark performance, and deployment details.
| Model | Provider | Context | Pricing | Key Capabilities |
|---|---|---|---|---|
| Fugu Ultra v2 | Sakana AI | 1,000,000 tokens | N/A | Text generation, long-context analysis, summarization, general assistance |
| Fugu Max | Sakana AI | 1,000,000 tokens | N/A | Text generation, long-document analysis, summarization, general assistance |
| Schematron v2 Turbo | Inference Net | 128,000 tokens | N/A | Text generation, long-context analysis, summarization, general assistance |
| Schematron v2 Small | Inference Net | 128,000 tokens | N/A | Text generation, long-context analysis, summarization, general assistance |
Fugu Ultra v2: Sakana’s v2 update for very large document workloads
Fugu Ultra v2 is the most visibly ambitious release of the week: a hosted Sakana AI foundation model positioned for very large text workloads. The “v2” naming suggests an update to the Fugu Ultra line, though available release data does not specify architectural changes, training improvements, benchmark deltas, or pricing changes.
Its most practical capability is straightforward: it is a text-generation model intended for long-context use cases such as analyzing extensive document collections, producing summaries from large inputs, and answering questions across broad source material. That makes it relevant for workflows involving technical documentation, legal records, research corpora, enterprise knowledge bases, or other settings where chunking and retrieval pipelines can introduce complexity.
Technical specifications are currently limited. Fugu Ultra v2 is listed with a 1,000,000-token context window and was added to OpenRouter on September 11, 2026. It is a hosted model rather than an open-weight release. Verified capabilities include text generation and long-context processing; no multimodal input support has been confirmed. Max output length, pricing, throughput, latency, and benchmark results are not available in the provided listing.
The main benefit is operational simplicity for large-input tasks. A model that can accept very large prompts can reduce the need for aggressive pre-filtering, document slicing, or hierarchical summarization. It may also help preserve cross-document relationships that are easy to lose when information is split across many smaller requests.
The caveat is that context capacity does not automatically equal reliable long-context reasoning. Users should test whether the model can actually retrieve, compare, and synthesize information from distant parts of a large prompt. Without benchmark data or pricing, it is also difficult to judge cost-effectiveness. Compared with more fully documented hosted alternatives, Fugu Ultra v2’s current appeal is its apparent scale and availability—but evaluation will need to happen in user-specific workloads.
Fugu Max: A second Sakana option for million-token text analysis
Fugu Max arrives alongside Fugu Ultra v2 and appears to be another new hosted model in Sakana AI’s Fugu family. Based on the available metadata, it targets similar workloads: general text generation, long-document analysis, and summarization.
What makes Fugu Max notable is not a published benchmark lead or new modality, but the addition of another Sakana-hosted option for very large context tasks. In practice, that may matter if Fugu Max and Fugu Ultra v2 differ in latency, quality, cost, or routing behavior—though those differences are not yet documented in the data available this week.
Its confirmed technical profile is concise: Fugu Max supports a 1,000,000-token context window, is available as a hosted model via OpenRouter as of September 11, 2026, and is not open weight. The listing verifies text generation and long-context capabilities. Pricing, maximum output length, benchmark performance, supported languages, safety tuning details, and architecture are not specified.
The strength of Fugu Max is its potential role as a general-purpose long-document model. For teams experimenting with full-repository code review, large policy-document analysis, contract comparison, or multi-report synthesis, it may offer a simpler prompt-in/prompt-out path than building an elaborate retrieval stack first.
The limitations are similar to Fugu Ultra v2 but worth stating separately: there is not yet enough public information to know whether Fugu Max is the faster, cheaper, smaller, stronger, or more specialized member of the pair. The “Max” label sounds performance-oriented, but labels are not measurements. Users should run side-by-side tests on tasks that require precise citation, cross-reference resolution, and resistance to distraction across long inputs.
Compared with Fugu Ultra v2, Fugu Max currently looks like a sibling rather than a clearly differentiated tier. Until Sakana AI or OpenRouter exposes pricing, output caps, or evaluation results, the practical distinction between the two will have to be discovered empirically.
Schematron v2 Turbo: Inference Net’s speed-coded member of the v2 family
Schematron v2 Turbo is a hosted Inference Net language model added to OpenRouter on September 12, 2026. It sits in the Schematron v2 family and is positioned by name as the likely higher-throughput or more responsive option, though no latency, pricing, or performance figures are provided in the verified data.
The model’s core capabilities are text generation and long-context processing. It is best suited, based on the listing, for general assistance, long-context analysis, and summarization. A 128,000-token context window makes it suitable for substantial documents, multi-file inputs, transcripts, reports, and structured analysis tasks without reaching the extreme scale of the Fugu releases.
Technically, Schematron v2 Turbo is a hosted, closed-weight model with 128,000 tokens of context. Max output length is not listed. Pricing is not available. No multimodal support is confirmed, so it should be treated as a text model unless additional provider documentation says otherwise. Benchmark data is also absent, which means claims about reasoning quality, coding ability, factual accuracy, or instruction following cannot be verified from the listing alone.
The likely benefit of Schematron v2 Turbo is practicality. Many real workloads do not require million-token prompts; they require enough room for a long report, a documentation set, or a batch of related files, plus responsive generation. If the “Turbo” designation corresponds to lower latency or better throughput, it could be attractive for interactive assistant use—but that remains to be validated.
The downside is uncertainty. Without pricing or benchmark disclosure, users cannot yet assess whether Turbo is optimized for speed, quality, cost, or simply product positioning. Compared with the larger Fugu models, Schematron v2 Turbo offers a smaller but still substantial context window; compared with Schematron v2 Small, it likely represents the more capable or faster variant, but the exact trade-off is not documented.
Schematron v2 Small: A potentially lighter option for long-context assistance
Schematron v2 Small is the second Inference Net release this week and appears to be the lighter member of the Schematron v2 pair. Like Turbo, it is a hosted text-generation model with long-context support and availability through OpenRouter as of September 12, 2026.
The most interesting aspect of Schematron v2 Small is its likely role as a more economical or efficient model for routine long-context tasks. The listing does not confirm lower pricing, smaller parameter count, faster latency, or reduced capability, but the “Small” label typically implies a model intended for less expensive or lower-latency use cases.
Its verified specifications match Schematron v2 Turbo in the areas currently disclosed: 128,000-token context, hosted availability, closed weights, text-generation capability, and suitability for general assistance, summarization, and long-context analysis. Pricing, max output length, modalities, benchmarks, and license details beyond non-open-weight status are not provided.
The strength of Schematron v2 Small may be fit-for-purpose deployment. Not every task needs a top-tier model. Summarizing internal notes, extracting action items, scanning documentation, or answering questions over a medium-sized source packet may benefit more from predictable latency and cost than from maximum reasoning depth. If Small is priced accordingly, it could become a practical default for high-volume text workflows.
The limitation is that this value proposition is inferred rather than proven. Users should not assume cost savings, speed, or acceptable quality until the provider publishes pricing and evaluation details or until they run their own tests. Compared with Schematron v2 Turbo, Small is probably the efficiency-oriented sibling; compared with the Fugu releases, it is a more bounded long-context option rather than a million-token document engine.
Practical note: where these models may help engineering teams
Long-context text models can be useful in software maintenance when the input itself is large: dependency manifests, changelogs, migration guides, release notes, configuration files, or multi-package repositories. A model with enough room to inspect related files together can help summarize upgrade risks, compare version changes, or draft migration checklists. That said, these tasks still require verification against authoritative sources; long context reduces friction, but it does not remove the need for deterministic tooling and human review.
Bottom line
This week’s releases broaden the menu of hosted long-context text models, with Sakana AI pushing into very large-input territory through Fugu Ultra v2 and Fugu Max, and Inference Net adding two 128k-context Schematron v2 options that may serve more everyday long-document workloads. The biggest caveat is transparency: without pricing, output limits, benchmarks, or modality details, the models are promising but not yet easy to evaluate on paper.
The direction is clear, though. Hosted LLM providers are continuing to make longer-input analysis a standard feature rather than a niche specialty. The next competitive frontier will be proving not just how much text a model can accept, but how accurately, affordably, and reliably it can reason over that text.