# Best Value 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-value

Qwen3.8 27B ranks first for Best Value LLM as of October 20, 2026. It ranks #15 on LiveBench Instruction Following (72.66%) and #64 on LMArena (Elo 1438), fitting on a single H200 GPU at full 262,144 context length with 122.1 tok/s for one caller. Gemini 3.8 Flash scores within two points, so read the top two as close rather than as a measured gap. Watch out: Watch for understated vision token metrics where image tokens are excluded from usage.prompt_tokens, breaking cost comparisons. Gemini 3.8 Flash is the next-ranked alternative. It ranks #1 on LiveBench Instruction Following (81.41%) and #8 on LMArena (Elo 1496), with promotional pricing at $0.75 per 1M input tokens and $3.75 per 1M output tokens through 2026. [1,2,3,4,5,6,7]

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

Qwen3.8 27B ranks #15 on LiveBench Instruction Following (72.66%) and #64 on LMArena (Elo 1438), fitting on a single H200 GPU at full 262,144 context length with 122.1 tok/s for one caller.

**Best when:** Self-host on H200 for full 262K context without multi-node overhead, achieving 122.1 tok/s single-caller throughput and 3957.3 tok/s at 1024 concurrency. [3]

**More tips:**
- Deploy on RTX 3090 with 18GB weights where ~7GB remaining VRAM supports practical local inference. [8]

**Watch out for:**
- Watch for understated vision token metrics where image tokens are excluded from usage.prompt_tokens, breaking cost comparisons. [4]
- Avoid with prefix caching and speculative decoding together, as EAGLE/MTP causes 30-40% batch throughput loss on prefix-reusing workloads. [9]

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

Gemini 3.8 Flash ranks #1 on LiveBench Instruction Following (81.41%) and #8 on LMArena (Elo 1496), with promotional pricing at $0.75 per 1M input tokens and $3.75 per 1M output tokens through 2026.

**Best when:** Use for instruction-following tasks where benchmark-topping accuracy matters, as it scores #1 of 59 on LiveBench Instruction Following at 81.41%. [5]

**More tips:**
- Deploy for cost-sensitive batch workflows where confirmed results at $0.303 each outperform Claude Opus 5 at $0.609 and Claude Sonnet 5 at $0.726. [10]
- Route through Vertex for half the per-token price of Gemini 3.6 Flash with roughly twice the speed and no token cap overruns in testing. [11]

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

DeepSeek V4.1 Flash ranks #26 on LMArena (Elo 1475) and #25 on LiveBench Instruction Following (70.03%), with worker-tier evaluation showing $4.32 per solved task versus $9.05 on Opus high.

**Best when:** Deploy for cost-per-solve optimization where $4.32 per task undercuts Opus high by more than half. [12]

**Watch out for:**
- Watch for reliability issues on SM120 TP4 with engram offloading that caused reversion to GLM53 flash in production. [13]

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

Gemini 3.7 Flash ranks #2 on LiveBench Instruction Following (79.93%) and #13 on LMArena (Elo 1488), sharing the same promotional pricing structure as 3.8 Flash.

**Best when:** Select for high-accuracy instruction following at #2 rank on LiveBench with promotional pricing identical to 3.8 Flash. [14,7]

**More tips:**
- Use as a lower-latency alternative when 4.2s average response time is acceptable for the task. [15]

**Watch out for:**
- Expect price increases to $1.50/$7.50 per 1M tokens after January 1, 2027, reducing long-term value. [15]

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

GLM 5.3 ranks #21 on LMArena (Elo 1479) and #27 on LiveBench Instruction Following (69.3%), with monthly costs ranging from $106 to $342 across providers for 100M input + 20M output tokens.

**Best when:** Shop providers aggressively, as the same usage costs $106 at DeepInfra versus $342 at the most expensive option. [16]

**Watch out for:**
- Watch for missing published pricing on API endpoints despite provider announcements of availability. [17]

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

Kimi K3 ranks #14 of 152 on LMArena's overall text arena (Elo 1487), based on blind human preference votes.

**Best when:** Its strongest result for this task is on LMArena Text, which measures general reasoning and instruction following. [18]

## 7. Muse Spark 1.3
https://modelmatchbook.com/models/meta/muse-spark-1.3-20260902

Muse Spark 1.3 ranks #4 on LiveBench Instruction Following (78%) and #9 on LMArena (Elo 1494), with strong human preference scores indicating good practical utility.

**Best when:** Use in agent reviewer roles where $0.0861 for 147.1K tokens provides cost-effective validation. [19]

## 8. Gemini 3.6 Flash
https://modelmatchbook.com/models/google/gemini-3.6-flash-20260721

Gemini 3.6 Flash ranks #8 on LiveBench Instruction Following (75.37%) and #18 on LMArena (Elo 1483), with curated pricing at $1.50/$7.50 and thinking ladder support.

**Best when:** Deploy for ask/propose actions in cost-optimized routing where verified reachability and HTTP 200 responses confirm reliable access. [20]

## FAQ

**What is the best model for Best Value LLM?**

Qwen3.8 27B ranks first as of October 20, 2026. Qwen3.8 27B ranks #15 on LiveBench Instruction Following (72.66%) and #64 on LMArena (Elo 1438), fitting on a single H200 GPU at full 262,144 context length with 122.1 tok/s for one caller. [1,2,3]

**What should I watch out for with Qwen3.8 27B?**

Watch for understated vision token metrics where image tokens are excluded from usage.prompt_tokens, breaking cost comparisons. [4]

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

Gemini 3.8 Flash ranks second. Gemini 3.8 Flash ranks #1 on LiveBench Instruction Following (81.41%) and #8 on LMArena (Elo 1496), with promotional pricing at $0.75 per 1M input tokens and $3.75 per 1M output tokens through 2026. [5,6,7]

**Can I self-host the top pick for Best Value LLM?**

Yes. Qwen3.8 27B is open-weight, so it can run on your own hardware as well as through hosted routes. [1,2,3]

## 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 152 catalog models matched to this evaluation: Ranks #64 of 152 on LMArena's overall text arena (Elo 1438), based on blind human preference votes." (LMArena text arena, https://lmarena.ai/leaderboard)
3. "Qwen3.8-27B is a 27B dense hybrid-attention model with a vision tower and an MTP draft head. It fits one H200 GPU at the full 262144 context, so it queues faster than any recipe that needs a whole node. | | Measured | | --- | --- | | One caller | 122.1 tok/s | | Concurrency 1024 | 3957.3 tok/s, still rising | | Context | 262144, the checkpoint maximum | | KV cache | 1,804,253 tokens, 6.88 full-length requests at once | | Hardware | 1 H200 GPU, TP1, no NCCL | Scope - New recipe at recipes/Qwen3.…" (mmshad, https://github.com/KempnerInstitute/hpc-agentic-recipes/issues/60)
4. "Related: #313. Problem For an OCR call on a full-size photo through LM Studio (qwen/qwen3.8-27b, image_url with detail: "high"), the response usage.prompt_tokens is ~230–260, while the server log for the same request shows uncached_tokens=3334. The sample's meta.promptTokens (and so totalTokens and costUsd) therefore excludes the image tokens. Token metrics for vision calls are understated, which breaks token comparisons between models in experiments (#313). Evidence - LM…" (abi83, https://github.com/abi83/prepify/issues/318)
5. "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/)
6. "Within the 152 catalog models matched to this evaluation: Ranks #8 of 152 on LMArena's overall text arena (Elo 1496), based on blind human preference votes." (LMArena text arena, https://lmarena.ai/leaderboard)
7. "Detected on 2026-10-03. Models removed - Copilot — Claude Opus 4.7 removed - Copilot — Gemini 3.5 Flash removed - Copilot — Gemini 3.6 Flash removed - Copilot — Kimi K2.7 Code removed - Copilot — Claude Opus 4.7 removed - Copilot — Gemini 3.5 Flash removed Changed notes - Copilot — Gemini 3.7 Flash and Gemini 3.8 Flash are available at the promotional pricing of $0.75 per 1M input tokens, $0.075 per 1M cached input tokens, and $3.75 per 1M output tokens throu…" (github-actions[bot], https://github.com/zrrbite/llm-price-watch/issues/15)
8. "Can't help much on the auto-approve side, but on a single 3090 the timeouts and empty replies are often a memory/context issue on the Ollama side, so here's the budget for qwen3.8:27b (estimates, not measured on your box): - Weights: the qwen3.8:27b tag is 18 GB (~16.8 GiB, vision tower included). That leaves ~7 GiB of the 24, minus ~1-1.5 GiB for the CUDA context and compute buffers. - KV cache: Qwen3.8 27B only keeps a KV cache in 16 of its 64 layers (the other 48 are linear attention wit…" (159753a52, https://github.com/cline/cline/discussions/14387#discussioncomment-18657911)
9. "EAGLE/MTP prefix-cache last-block drop causes a 1,648-token recompute per hit on one hybrid Qwen3.8 GDN layout — ~30-40% batch throughput loss on prefix-reusing workloads with speculative decoding enabled Environment - vLLM 0.26.1rc1.dev926+gb05ae5dc0, V1/V2 model runner, async scheduling - Dual RTX 5090 TP=2, FP8 weights + FP8 KV, prefix caching on - Model: Qwen3.8-27B (hybrid: GDN/Mamba layers + full attention → 15 MambaSpec KV groups + full-attention group; scheduler/hash block = 1648 t…" (Suppressor72, https://github.com/vllm-project/vllm/issues/53670)
10. "The runs Four models measured under the current protocol, in one batch: | Model | Total | Cost | $/confirmed | Confirmed / reported | | --- | --- | --- | --- | --- | | google/gemini-3.8-flash | 20m12s | $3.33 | $0.303 | 11 / 11 | | anthropic/claude-opus-5 | 27m36s | $10.35 | $0.609 | 17 / 17 | | anthropic/claude-sonnet-5 | 39m51s | $5.81 | $0.726 | 8 / 8 | | anthropic/claude-haiku-4.5 | 10m03s | $0.64 | $0.320 | 2 / 7 | Each cost reconstructed exactly at OpenRouter's published pr…" (using-system, https://github.com/using-system/oddyssey/issues/499)
11. "Summary Switch the Cloudflare Worker's LLM from google/gemini-3.6-flash to vertex/gemini-3.8-flash on Requesty. Same output quality on the live batch prompt, half the per-token price, roughly twice as fast, and it never overran the token cap in testing (3.6 Flash did once in four refills, forcing a full strict-prompt retry). Changes (all in worker/) 1. wrangler.toml: set UR_MODEL = "vertex/gemini-3.8-flash". 2. wrangler.toml: bump UR_GENERATION_VERSION from 2 to 3 so dev…" (alexsiri7, https://github.com/alexsiri7/un-reminder/issues/434)
12. "Completed and merged into main in pull request https://github.com/Wladefant/super-board/pull/342 (commit 3c360188711ef811452a2643c9c4265c4ce5db5f). Authoritative deliverable: https://github.com/Wladefant/super-board/blob/3c360188711ef811452a2643c9c4265c4ce5db5f/docs/research/agent-benchmarks-2026-09.md Comprehensive benchmark research delivered: Artificial Analysis Terminal-Bench 4.0 analysis, DeepSeek V4.1 Flash worker tier evaluation ($4.32/solved vs $9.05 on Opus high), cost-per-solve me…" (Wladefant, https://github.com/Wladefant/super-board/issues/211#issuecomment-5892105103)
13. "Same here, and I kind of had this with all ds 4 flash models as well... also running the stock snapshot on SM120 TP4 with engram offloading. This really makes it unreliable to use for us, and we reverted to GLM53 flash now, although I really like the much faster architecture of ds 4/4.1 flash!" (opt-dev, https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/discussions/72#6ad5248c118a080c497d160d)
14. "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/)
15. "追加評価: gpt-5.6-luna（2026-08-17） | | gpt-4o-mini | gemini-3.7-flash | gpt-5.6-luna | |---|---|---|---| | 精度 | 5/8 | 7/8 | 7/8 | | 平均レイテンシ | 3.1s | 4.2s | 4.3s | | 単価（in/out per 1M） | $0.15/$0.60（legacy） | $0.75/$3.75 → 2027年に$1.50/$7.50へ値上げ | $0.20/$1.20（値上げ予定なし） | | draft 1回概算 | 〜$0.0007 | 〜$0.004 | 〜$0.0011 | - Luna は gemini-3.7-flash と同じ 7/8（落ちたのは全モデル共通の conflict warnings ケースのみ）。曖昧期限・前日準備も正しく分解 - priority 判定は gemini がやや上（Luna は起床/出社/帰宅に mandatory を付ける過剰傾向。ただし商談=external/…" (devacxx01, https://github.com/devacxx01/voice-nudge/issues/2#issuecomment-5312741497)
16. "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)
17. "Observed Z.AI announced GLM-5.3 to all API users on 2026-08-22: same base model as GLM-5.2, same wire shape, "one parameter change" to adopt, and — per the launch note — "available at the same price". #6988 catalogued the route the same day but deliberately left it unpriced ("Z.AI has not added it to the public per-token rate card"). Two problems follow from the state that landed: 1. zai/glm-5.3 carries no pricing. docs.z.ai's rate card now lists GLM-5.3 at $1.40 in / $4.40 out / $…" (kennethsinder, https://github.com/burin-labs/harn/issues/7024)
18. "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)
19. "Corrección: los costos reales, medidos en DevPass Primera facturación real de las lanes de este issue. Mis proyecciones estaban 6-10× por encima. | Rol | Modelo real | Req | Tokens | Costo | | --- | --- | ---: | ---: | ---: | | explorer | azure/gpt-5.6-luna | 5 | 127.2K | $0.0104 | | reviewer | meta/muse-spark-1.3 | 5 | 147.1K | $0.0861 | | (auxiliar) | google-vertex/gemini-3.7-flash | 2 | 3.0K | $0.0023 | | | | 12 | 277.3K | $0.0988 | Las dos lanes completas de…" (ingamartinez, https://github.com/ingamartinez/personal-financial-dashboard/issues/922#issuecomment-5632309311)
20. "Context Unified spec item 3b — per-action Gemini model routing (economics doc E-2: ~32% blended cost cut, "one URL constant plus a per-action switch, no product change"). Routing: ask/propose → gemini-3.6-flash, hint → gemini-3.5-flash-lite. Both targets are VERIFIED REACHABLE on the prod key — see the 2026-07-28 curl record in #838's comments (AIE-04: all three candidates HTTP 200; AIE-05: thinkingBudget: 0 verified honored with a control). Note the route's line-69 comme…" (thomgabriel, https://github.com/solanabr/superteam-academy/issues/868)