# Cheapest Capable LLM

> Updated Oct 20, 2026 · evidence through Oct 20, 2026. Synthesized from cited community evidence and open benchmarks, not vendor marketing. Canonical page: https://modelmatchbook.com/best/budget-cheapest

Qwen3.8 27B ranks first for Cheapest Capable LLM as of October 20, 2026. It scores 72.66% on LiveBench Instruction Following (#15 of 59), including paraphrasing, simplifying, story generation, and summarization. DeepSeek V4.1 Flash is the next-ranked alternative. It scores 70.03% on LiveBench Instruction Following (#25 of 59) and ranks #26 on LMArena (Elo 1475). [1,2,3,4,5,6,7]

## 1. Qwen3.8 27B
https://modelmatchbook.com/models/qwen/qwen3.8-27b-20260814

Qwen3.8 27B scores 72.66% on LiveBench Instruction Following (#15 of 59), including paraphrasing, simplifying, story generation, and summarization.

**Best when:** Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly. [1]

## 2. DeepSeek V4.1 Flash
https://modelmatchbook.com/models/deepseek/deepseek-v4.1-flash-20260910

DeepSeek V4.1 Flash scores 70.03% on LiveBench Instruction Following (#25 of 59) and ranks #26 on LMArena (Elo 1475).

**Best when:** Use for human-preference-aligned outputs where LMArena ranking matters, placing #26 of 152 with Elo 1475. [3]

**Watch out for:**
- Advertised as FREE in model selector but billed credits anyway, causing negative balance (-0.26 USD) and breaking web_search with HTTP 402 insufficient_credits. [7]

## 3. GLM 5.3
https://modelmatchbook.com/models/z-ai/glm-5.3-20260816

GLM 5.3 scores 69.3% on LiveBench Instruction Following (#27 of 59) and ranks #21 on LMArena (Elo 1479), with DeepInfra offering the cheapest confirmed pricing at $0.56/M input and $2.50/M output, though the model was deliberately left unpriced on some public rate cards.

**Best when:** Select DeepInfra as your provider for the cheapest confirmed pricing at $0.56 per million input tokens and $2.50 per million output tokens, with monthly cost of $106 for 100M input + 20M output tokens versus $342 at the most expensive provider. [8]

## 4. Gemini 3.8 Flash
https://modelmatchbook.com/models/google/gemini-3.8-flash-20260902

Gemini 3.8 Flash scores 81.41% on LiveBench Instruction Following (#1 of 59), including paraphrasing, simplifying, story generation, and summarization.

**Best when:** Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly. [9]

## 5. Kimi K3
https://modelmatchbook.com/models/moonshotai/kimi-k3-20260715

Kimi K3 scores 71.36% on LiveBench Instruction Following (#20 of 59) and ranks #14 on LMArena (Elo 1487), placing it mid-pack on both benchmarks with no cost-specific evidence.

**Best when:** Use for balanced performance with #14 LMArena ranking (Elo 1487) and 71.36% LiveBench Instruction Following. [10,11]

## 6. Gemini 3.7 Flash
https://modelmatchbook.com/models/google/gemini-3.7-flash-20260813

Gemini 3.7 Flash scores 79.93% on LiveBench Instruction Following (#2 of 59) and ranks #13 on LMArena (Elo 1488).

**Best when:** Use for near-top instruction-following performance at 79.93%, ranking #2 of 59 on LiveBench. [12]

**Watch out for:**
- Listed with $0 cost for input and output in some model catalogs, causing session usage under-reporting and budgeting errors. [13]

## 7. DeepSeek V4 Flash 0423
https://modelmatchbook.com/models/deepseek/deepseek-v4-flash-20260423

DeepSeek V4 Flash 0423 scores 63.14% on LiveBench Instruction Following (#44 of 59), including paraphrasing, simplifying, story generation, and summarization.

**Best when:** Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly. [14]

## 8. GPT-5.4 Nano
https://modelmatchbook.com/models/openai/gpt-5.4-nano-20260317

GPT-5.4 Nano scores 67.2% on LiveBench Instruction Following (#31 of 59), including paraphrasing, simplifying, story generation, and summarization.

**Best when:** Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly. [15]

## FAQ

**What is the best model for Cheapest Capable LLM?**

Qwen3.8 27B ranks first as of October 20, 2026. Qwen3.8 27B scores 72.66% on LiveBench Instruction Following (#15 of 59), including paraphrasing, simplifying, story generation, and summarization. [1]

**What is the best alternative to Qwen3.8 27B?**

DeepSeek V4.1 Flash ranks second. DeepSeek V4.1 Flash scores 70.03% on LiveBench Instruction Following (#25 of 59) and ranks #26 on LMArena (Elo 1475). [2,3,4,5,6,7]

**Can I self-host the top pick for Cheapest Capable LLM?**

Yes. Qwen3.8 27B is open-weight, so it can run on your own hardware as well as through hosted routes. [1]

## Sources
1. "Within the 59 catalog models matched to this evaluation: Scores 72.66% on LiveBench Instruction Following (#15 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)
2. "Within the 59 catalog models matched to this evaluation: Scores 70.03% on LiveBench Instruction Following (#25 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)
3. "Within the 152 catalog models matched to this evaluation: Ranks #26 of 152 on LMArena's overall text arena (Elo 1475), based on blind human preference votes." (LMArena text arena, https://lmarena.ai/leaderboard)
4. "结果 上游钉住（provider.only / providerOptions.gateway.only）不生效，但插件报成功。 以 cline-pass/deepseek-v4.1-flash 为例，原因是网关按官方路由顺序执行、不采纳 only：该模型的官方顺序里 deepseek 排在第一位，所以请求永远由 deepseek 服务，钉别的渠道都会被回退。 网关自己的原话（响应里的 planningReasoning）： 对比 glm-5.3，它才是真的被限制住了： 也就是说：同样是 only，在 glm-5.3 上是限制，在 deepseek-v4.1-flash 上不起作用——后者由官方路由顺序决定。 实测钉 baseten： 插件只判断"请求是否成功"，从不回读实际服务的渠道，于是把这种失效报成了成功。同一条 cline_pass_validate 还把 15 条渠道全部标为 available，配置的 14 条 exclude 也没挡住 deepseek。 插件侧的问题 1. **…" (codeOct, https://github.com/yhshzh/dsh-cline-pass/issues/3)
5. "Description deepseek-v4.1-flash via OpenCode Go returns HTTP 402 insufficient_user_quota from an OrcaRouter upstream, while the OpenCode Go account quota is healthy and other Go models work with the same API key/session. This is the same September wave reported in #37231 (see the 2026-09-19 comment from @hyuri) and #49416. Steps to reproduce 1. Call OpenCode Go Chat Completions with deepseek-v4.1-flash: Response: HTTP 402 2. Check OpenCode Go usage with the same key: Result: HTTP…" (zfei2080, https://github.com/anomalyco/opencode/issues/49936)
6. "Live probes from the Claude VPS session, 2026-09-19 ~14:00 IST, echo 'reply with exactly OK' | pi --print --provider X --model Y: | seat | result | |---|---| | litellm worker-capable | two pi packets ended their turn after 15-69 tool calls with no deliverable (units 0509-d1-drill-weekly, -r2); both alias deployments show litellm_deployment_state 2 | | synthetic hf:zai-org/GLM-5.3-Flash | 429 subscription exceeded | | zenmux deepseek/deepseek-v4.1-flash | 402 no credits | | cline z-ai/glm-5.3-…" (nish3451, https://github.com/Nishfleet/fleet-ops/issues/7862)
7. "Separate ask: credit refund / account adjustment Beyond the technical fixes above, I'd like this reviewed as a billing matter, not only a bug: 1. My Cline Credits balance is negative (-0.26 USD) in the Cline provider, which silently breaks web_search with HTTP 402 insufficient_credits. 2. Part of that consumption happened while I was using models shown as FREE in the model selector (cline-free/muse-spark-1.3-contributor, cline-free/deepseek-v4.1-flash). 3. In the same pe…" (romemozu, https://github.com/cline/cline/issues/14256#issuecomment-5728160303)
8. "Prices per provider from OpenRouter's list, 23 Sep. Monthly cost assumes 100M input + 20M output tokens. | Provider | $/M input | $/M output | Monthly | |---|---|---|---| | DeepInfra | 0.56 | 2.50 | $106 | | Novita | 0.78 | 2.46 | $127 | | Reka | 0.76 | 2.57 | $127 | | Io Net | 0.81 | 2.73 | $135 | | Phala | 0.84 | 2.64 | $137 | Same model, same usage: $106 at the cheapest provider, $342 at the most expensive. Live prices, precision and uptime for all 31, with a calculator for your own usag…" (yash-711, https://huggingface.co/zai-org/GLM-5.3/discussions/26#6ab3afcc6587c2152c2278d9)
9. "Within the 59 catalog models matched to this evaluation: Scores 81.41% on LiveBench Instruction Following (#1 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)
10. "Within the 59 catalog models matched to this evaluation: Scores 71.36% on LiveBench Instruction Following (#20 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)
11. "Within the 152 catalog models matched to this evaluation: Ranks #14 of 152 on LMArena's overall text arena (Elo 1487), based on blind human preference votes." (LMArena text arena, https://lmarena.ai/leaderboard)
12. "Within the 59 catalog models matched to this evaluation: Scores 79.93% on LiveBench Instruction Following (#2 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)
13. "What omp models on 17.3.0 lists google/gemini-3.7-flash with cost.input/output = 0 and reasoning: false / thinking: null. google/gemini-3.6-flash is curated ($1.50 / $7.50, thinking ladder). Session usage therefore under-reports 3.7 Flash spend, and :medium / :high suffixes are not in the catalog even though the API accepts thinkingLevel. Official rates (Gemini API, introductory through 2026-12-31) From [the 3.7 Flash announcement](https://blog.google/innovation-and-…" (john-aigora, https://github.com/can1357/oh-my-pi/issues/8605)
14. "Within the 59 catalog models matched to this evaluation: Scores 63.14% on LiveBench Instruction Following (#44 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)
15. "Within the 59 catalog models matched to this evaluation: Scores 67.2% on LiveBench Instruction Following (#31 of 59), including paraphrasing, simplifying, story generation, and summarization." (LiveBench Instruction Following, https://livebench.ai/)