Best LLM for Hermes Agent

Top of the list on arithmetic. It is the cheapest model here per token, its cache-hit rate is a tiny fraction of its input rate, and a long-running Hermes session re-sends the same system prompt, memory and tool schemas on every turn, which is exactly what a cache discount pays for. The catch is the evidence: the 90.6 Terminal-Bench 2.1 score that puts it first is DeepSeek's own, and no independent lab has measured it yet. Peak hours on weekdays bill double the off-peak figure shown here, which matters for cron jobs scheduled into those windows.

The best-measured cheap option for a shell-heavy agent: 77.5 on Terminal-Bench 2.1 with a 1M context. Two dates to know before you build on it. Google's current rate is a cut that its own page says reverts to the launch price on January 1, 2027, and Vertex AI retires the model on November 19, 2026, while the Gemini API keeps serving it and now routes deprecated 3.5 Flash traffic to it. Thinking tokens bill as output, so agent runs cost more than the sticker suggests.

The highest SWE-bench Verified score on the page at 96.0 (Anthropic-reported), and 86.1 on Terminal-Bench 2.1 as measured by Artificial Analysis, the best independent shell score here. On a list this price-sensitive it still lands near the top, which says how wide the quality gap is. Use it as the main model for the hard, rare job, or for a subagent you send the difficult part to, rather than as the thing that answers every Telegram message.

MIT-licensed open weights with a 1M context. Its 77.9 on Terminal-Bench 2.1 is Artificial Analysis's run of the max reasoning setting, not the default, so read it as an upper bound for a plain call (Z.ai self-reports 81.0). Hermes has a built-in z.ai provider, so it is one key away.

The safe default if you want a model whose scores are well documented: 85.2 on SWE-bench Verified and 80.5 on Terminal-Bench 2.1 (the latter at max effort, so an upper bound), full 1M context, no long-context surcharge. Its tokenizer emits roughly 30% more tokens than Sonnet 4.6 for the same text, so compare real bills, not sticker rates, if you are moving from an older Claude.

MiniMax now labels its half price permanent, which makes M3 one of the cheapest 1M-context models with an independently measured Terminal-Bench 2.1 score (65.2, Artificial Analysis). Native image and video input means it can fill the vision slot too. Hermes supports MiniMax by API key or by browser OAuth. Rates double above 512K input, so very long sessions should be compressed before they get there.

Open weights from Moonshot, 65.9 on Terminal-Bench 2.1, and fast for its size. Its 262K context is well clear of Hermes's 64K minimum but a quarter of the 1M models above it, so leave room for memory and tool output on long sessions. Hermes has a built-in Kimi provider.

Not a main model for most people. It is here as the obvious auxiliary pick: Hermes sends context compression, web-page summarising, session titles and approval checks to side slots, and every slot defaults to your main model. Pointing those at a cheap fast model is the easiest saving in Hermes, and this one posts 53.6 on Terminal-Bench 2.1 at a fraction of a frontier rate.