Best LLM in 2026: the models Standards recommends

An LLM, or large language model, is the AI model that powers ChatGPT, Claude and Gemini: it reads, writes and reasons in plain language, and you pay for it by the million tokens processed. LLM rankings don't all measure the same thing, which is why they disagree. Artificial Analysis aggregates intelligence benchmarks. Multiple.chat counts the tokens people actually spend, which measures popularity. The Standards model catalog keeps a third kind: for each model, the tasks it can do (conversation, writing, code, analysis, extraction), the level of demand it holds on each (light, standard or heavy), and its price per million tokens. That ranking serves a team paying per use that wants the model that does the job at the lowest fair price. The table below sums up the three on October 30, 2026.

Standards recommends three models on October 30, 2026, in this order: OpenAI's Luna 6, Google's Gemini 3.8 Flash and DeepSeek V4.1 Flash. All three are balanced models, qualified at standard level or above on every task, and priced under a dollar per million input tokens. The table below is read live from the catalog, so if the recommendation changes tomorrow, the table changes with it.

Luna 6 holds standard level everywhere for $0.10 per million input tokens, which makes it a sound pick for everyday work. Gemini 3.8 Flash and DeepSeek V4.1 Flash both reach heavy level for code, and Gemini 3.8 Flash costs more than DeepSeek V4.1 Flash. None of the three tops the performance leaderboards, and that isn't the criterion: a recommended model here is one you can leave running on a team's daily work all day without watching the bill. For heavy tasks, the catalog qualifies other models, and Standards' Auto mode picks them when the work in progress needs it. The catalog's default model, Claude Haiku 5.5, isn't on this list; it also serves as the utility model.

Ranking LLMs by task gives you a more useful table than a single score, because a model that excels at code can be average at writing. Code is where the gap shows most. On October 30, 2026, nine models in the Standards catalog are qualified heavy for code, at prices that run from $0.15 to several dollars per million input tokens. These are the cheapest of them.

On that date, no model in the catalog is qualified heavy for extraction, so on long, ambiguous documents that kind of extraction deserves a human read. Mistral's three models, Ministral 14B, Medium 3.5 and Large 3, are qualified light on every task. They suit short, high-volume requests; a multi-step analysis will go to another provider.

The Standards catalog updates without a new product version, in four steps, one of which goes through a person. The docs walk through the app that runs it, which is built on Standards too.

1. Every six hours, a scheduled job reads the models the Vercel AI Gateway offers and writes one record per model into a Standards workspace.

2. A person reviews the records: publish, reject, or confirm a price drop. A price drop waits for that check because it can be a promotion, and the published price works as a spending ceiling.

3. Every thirty minutes, a second job publishes the approved records as a JSON file at models.standards.new.

4. By default, every Standards deployment rereads that file at startup and every 30 minutes, so a model published in the morning can be selected that afternoon.

On October 25, 2026, a Standards update brought its fallback list back in line with the published catalog. That list is the set of models built into the product, used when the catalog can't be reached, and it had been frozen since late September. The catch-up counts eight additions and twelve removals.

• In: Luna 6, Sol 6.1, Claude Opus 5.5, Sonnet 5.5, Grok 4.7, Muse Spark 1.3, Muse Glimmer and Kimi K3.1.

• Out: twelve models, including GPT-6 Astra, Claude Opus 5, Sonnet 5, Haiku 4.5, Fable 5.1 and Kimi K3.

A conversation set to a retired model doesn't break: Standards moves it to the closest live model. The same day, the fallback list took Claude Haiku 5.5 as its default and utility model, as the published catalog does: it's the one that writes conversation titles and memory.

For what an agent may do on its own once its model is set, read our article on AI agent security for firms that hold client data. To connect Claude or ChatGPT to your own client data, see which CRMs run an MCP server.