Recommendation for Best value
Best Value LLM
Our top recommendation for Best Value LLM, based on the public evidence we track, is Qwen: Qwen3.8 27B.[1][2][3][4][5] Run long-context workloads (60K+ retrieval) on single RTX 5090 or Apple Silicon with 16-27 GB RAM thanks to hybrid attention that uses only 17 full-attention layers and 48 SSM layers. Watch out: NVFP4 quantization incurs a measurable quality tax (92.5 vs 93.2 q_avg, 89.6 vs 94.5 HumanEval) versus GGUF Q4_K_M, so verify critical outputs when using compressed weights. Google: Gemini 3.6 Flash is the next-ranked alternative. Access strong human-preference-ranked quality at $0.75/$3.75 per million tokens (introductory through 2026-12-31), undercutting many tier-1 alternatives for high-volume text generation.
About this recommendation
- Updated
- Sep 25, 2026
- Evidence through
- Sep 25, 2026
- Sources
- 21
- Revision
- v78
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
20
live candidates
Evaluation feeds
5
task-weighted
Winner coverage
71%
intended feed weight
Largest provider share
1 of 6
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| price weight | 30% | 100/100 | 20/20 |
| Route reliabilityunavailable | 25% | feed unavailable | 0/20 |
| LiveBench Instruction Following | 20% | #14 | 19/20 |
| LMArena Text | 15% | #52 | 20/20 |
| OpenRouter usage | 10% | 89/100 | 20/20 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic1 model
- deepseek1 model
- Google1 model
- OpenAI1 model
- Qwen1 model
- xAI1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
| Rank | Model | Relative score | Coverage | Practitioner evidence | Strongest measured reason |
|---|---|---|---|---|---|
| 01 | Qwen3.8 27BQwen | 76 | 71% | 30 threads · 21 families · 8 cautions | #14 LiveBench Instruction Following · #52 LMArena Text |
| 02 | Gemini 3.6 FlashGoogle | 74 | 71% | 15 threads · 13 families · 6 cautions | #8 LiveBench Instruction Following · #16 LMArena Text |
| 03 | DeepSeek V4 Flash 0423deepseek | 74 | 71% | 27 threads · 23 families · 7 cautions | #40 LiveBench Instruction Following · #53 LMArena Text |
| 04 | GPT-5.6 SolOpenAI | 73 | 71% | 10 threads · 6 families · 4 cautions | #13 LMArena Text · #17 LiveBench Instruction Following |
| 05 | Grok 4.6xAI | 70 | 71% | 12 threads · 7 families · 7 cautions | #16 LiveBench Instruction Following · #34 LMArena Text |
| 06 | Claude Opus 4.6Anthropic | 67 | 71% | 3 threads · 3 families · 1 cautions | #2 LMArena Text · #38 LiveBench Instruction Following |
Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.
Qwen3.8 27B delivers strong quality at local inference costs through hybrid GDN architecture that reduces memory pressure and enables 64K context on consumer hardware.
Best when: Run long-context workloads (60K+ retrieval) on single RTX 5090 or Apple Silicon with 16-27 GB RAM thanks to hybrid attention that uses only 17 full-attention layers and 48 SSM layers.
Tips
- Run long-context workloads (60K+ retrieval) on single RTX 5090 or Apple Silicon with 16-27 GB RAM thanks to hybrid attention that uses only 17 full-attention layers and 48 SSM layers.
- Achieve 27-37 tok/s decode at 16K context on M3 Max with oMLX oQ4e-mtp, making it viable for interactive local agent workflows without cloud API costs.
Watch out for
- NVFP4 quantization incurs a measurable quality tax (92.5 vs 93.2 q_avg, 89.6 vs 94.5 HumanEval) versus GGUF Q4_K_M, so verify critical outputs when using compressed weights.
- Speculative decoding with MTP can trigger 1,648-token recomputes on prefix-cache hits in hybrid GDN layouts, cutting batch throughput 30-40% on reuse-heavy workloads.
Gemini 3.6 Flash ranks #17 on LMArena with competitive pricing at $0.75/$3.75 per million tokens, though catalog inconsistencies and parameter deprecation create integration friction.
Best when: Access strong human-preference-ranked quality at $0.75/$3.75 per million tokens (introductory through 2026-12-31), undercutting many tier-1 alternatives for high-volume text generation.
Tips
- Access strong human-preference-ranked quality at $0.75/$3.75 per million tokens (introductory through 2026-12-31), undercutting many tier-1 alternatives for high-volume text generation.
Watch out for
- Google deprecated temperature, top_p, and top_k for this model family, causing OpenRouter requests with provider.require_parameters: true to fail unless you strip these fields.
- Some provider catalogs misreport 3.7 Flash as zero-cost or omit 3.6 Flash entirely, so verify live pricing before committing to automated spend tracking.
DeepSeek V4 Flash offers open-weight deployment with API pricing that undercuts closed alternatives, though prompt caching gaps and legacy alias confusion increase real costs.
Best when: Deploy on RTX 5080 16GB via community ports for local inference, or use the retired-alias pricing (deepseek-v4-flash requests now served by V4.1-Flash at Flash rates) for cost arbitrage.
Tips
- Deploy on RTX 5080 16GB via community ports for local inference, or use the retired-alias pricing (deepseek-v4-flash requests now served by V4.1-Flash at Flash rates) for cost arbitrage.
Watch out for
- OpenCode Go endpoint does not implement prompt caching for this model, so every request reprocesses the full prefix and burns metered limits 2-3× faster than cached estimates suggest.
- The thinking/disabled toggle is not respected by some providers, defaulting to high reasoning and billing extra tokens you cannot suppress in the UI.
GPT-5.6 Sol offers token-efficient reasoning at ChatGPT Pro credit rates, though 10× higher quota consumption than Sol and model substitution risks complicate value assessment.
Best when: Achieve verified token-cost reconciliation against published credit rates for uncached input (100 credits per unit) and cached input (25 credits), enabling predictable budgeting on ChatGPT Pro.
Tips
- Achieve verified token-cost reconciliation against published credit rates for uncached input (100 credits per unit) and cached input (25 credits), enabling predictable budgeting on ChatGPT Pro.
Watch out for
- Codex CLI 0.153.4 defaults to GPT-6 Astra (10× weekly quota per token), silently draining Pro allowances if you do not explicitly pin gpt-5.6-sol in harness configuration.
- Multi-account concurrency can trigger silent model substitution to GPT-4o level capability while charging advanced-model rates, with affected conversations showing June 2024 knowledge cutoff.
Grok 4.6 provides capable reasoning at competitive API rates with broad provider availability, though infrastructure timeouts and missing price catalog entries create operational uncertainty.
Best when: Access through multiple routes (xAI direct, OpenRouter, Bedrock, Vertex) with xhigh reasoning clamped to high, providing flexibility for multi-cloud failover strategies.
Tips
- Access through multiple routes (xAI direct, OpenRouter, Bedrock, Vertex) with xhigh reasoning clamped to high, providing flexibility for multi-cloud failover strategies.
- $20 Cursor Pro plan users report Grok 4.6 as good enough and cheap for daily coding workflows, suggesting viable price-performance for budget-conscious developers.
Watch out for
- Infrastructure timeouts on the caller side can kill generations mid-stream and record them as scored verdicts with zero cost, masking actual failure rates in evaluation pipelines.
- Missing price entries in default-model-prices.json cause agent-under-test calls to report $0.00 cost, making true spend invisible in run accounting until manually corrected.
Anthropic: Claude Opus 4.6 ranks #2 of 146 on LMArena's overall text arena (Elo 1505), based on blind human preference votes.
Best when: Consider only after reviewing the cited caution.
Frequently asked
- What is the top-ranked model for Best Value LLM?
- Qwen: Qwen3.8 27B ranks first in the current evidence-weighted comparison. Run long-context workloads (60K+ retrieval) on single RTX 5090 or Apple Silicon with 16-27 GB RAM thanks to hybrid attention that uses only 17 full-attention layers and 48 SSM layers.[1][2]
- What should I watch out for with Qwen: Qwen3.8 27B?
- NVFP4 quantization incurs a measurable quality tax (92.5 vs 93.2 q_avg, 89.6 vs 94.5 HumanEval) versus GGUF Q4_K_M, so verify critical outputs when using compressed weights.[3]
Sources
- 1
“## 64K context: fits with room to spare — 16K was over-cautious Parsed the full GGUF metadata header (51 KV pairs) instead of estimating. The KV geometry makes long context far cheaper on this model than on a normal 27B. ### Architecture (from GGUF metadata, `general.architecture = qwen35`) | key | value | |---|---| | `qwen35.block_count` | 65 | | `qwen35.full_attention_interval` | **4** — so only every 4th layer is full attention → **17 attention layers**, 48 SSM | | `qwen35.attention.head_cou…”
changtimwu · GitHub · Aug 19, 2026 - 2
“# Benchmark: Full context ladder — oMLX (oQ4e/oQ6e/oQ8e, MTP+ANE) vs MTPLX (Speed/Quality) on M3 Max 128GB ## TL;DR On an **Apple M3 Max 128GB (40-core GPU, macOS 15.7.9)**, running **Qwen3.8-27B**: - **oMLX oQ4e-mtp (MTP on / TQ off / ANE on) is the best daily driver**: fastest decode at 16K (27.0 tok/s), fastest prefill (231.7 tok/s), lowest memory (16.6 GB), zero collapse risk. - **MTPLX Speed wins only in short-context (≤8K) decode**: 41.4 vs 37.5 tok/s at 1K (+10%), 31.6 vs 28.2 at 8K (+12…”
taozhiyuai · GitHub · Aug 18, 2026 - 3
“Hardware: 1xRTX 5090 32 GB (SM 12.0), driver 595.84, CUDA 13.2 toolkit Engine: vLLM 0.27.1 (pip) + FlashInfer 0.6.17, torch 2.13.0+cu130 Model: Qwen3.8-27B, compressed-tensors NVFP4 (W4A16, group 16), MTP speculative head in BF16 ### Quality functional correctness verified (math, tool calling, 60K context retrieval, multi-turn); NVFP4 carries a small quality tax vs GGUF Q4_K_M on evals (independent measurement: q_avg 92.5 vs 93.2, HumanEval 89.6 vs 94.5) — the trade for ~1.6x faster decode than…”
marco9899 · Hugging Face · Aug 18, 2026 - 4
“## 目的 benchmark に Gemini の最新 Flash モデル `gemini-3.7-flash` を追加して計測する。 ## 背景: Gemini Flash 系のラインアップと料金 公式ドキュメント (https://ai.google.dev/gemini-api/docs/models, https://ai.google.dev/gemini-api/docs/pricing) で確認した現行の Flash 系モデル。 | モデル | API ID | Input ($/1M) | Output ($/1M) | |---|---|---|---| | Gemini 3.7 Flash | `gemini-3.7-flash` | 0.75 (〜2026末) / 1.50 | 3.75 / 7.50 | | Gemini 3.6 Flash | `gemini-3.6-flash` | 0.75 (〜2026末) / 1.50 | 3.75 / 7.50 | | Gemini 3.5 Flash | `gemini-3.5-flash` | **1.50**…”
kiyoka · GitHub · Aug 16, 2026 - 5
“Ranks #17 of 146 on LMArena's overall text arena (Elo 1480), based on blind human preference votes.”
LMArena text arena · Benchmark · Sep 13, 2026 - 6
“# Benchmark: MTPLX (Speed/Quality) vs oMLX (oQ4e/oQ6e/oQ8e) on Qwen3.8-27B — full context ladder on M3 Max 128GB ## TL;DR On an **Apple M3 Max 128GB (40-core GPU, macOS 15.7.9)**, running **Qwen3.8-27B**: - **MTPLX Speed is the winner for short-context (≤8K) decode**: 41.4 vs 37.5 tok/s at 1K (+10%), 31.6 vs 28.2 at 8K (+12%) vs oMLX oQ4e. - **But MTPLX prefill collapses at long context**: 115.5 vs oMLX's 231.7 tok/s at 16K (50% slower). End-to-end at 16K, oMLX is ~2× faster. - **MTPLX Speed ≈…”
taozhiyuai · GitHub · Aug 18, 2026 - 7
“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 · GitHub · Aug 25, 2026 - 8
“### Description When using google/gemini-3.6-flash or google/gemini-3.5-flash-lite through OpenRouter, OpenCode automatically sends: Google documents temperature, top_p, and top_k as deprecated for these models and instructs clients to remove them: https://ai.google.dev/gemini-api/docs/generate-content/latest-model#migrate-to-gemini-3-6-flash This causes OpenRouter requests using provider.require_parameters: true to fail before inference because the Google AI Studio and Vertex endpoints do not…”
anandpant · GitHub · Jul 24, 2026 - 9
“## 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 · GitHub · Aug 15, 2026 - 10
“### Problem / Use Case Hi , it seems omniroute v3.8.48 obsoleted, Some Cookie not worked and can't connect , if connect , the model are unavailable. when i see list of models , the model are old version. for example in anatigravity gemini-2.5-flash-thinking is available but in reality it not available yet , also gemini-3.6-flash added but it's not in list. I can pretty much say that one of these issues exists in 90% of the providers. It seems like the Docker version is really outdated, or maybe…”
ImMahdi · GitHub · Jul 29, 2026 - 11
“(1) Use deepseek-flash as the model name. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price. (2) After extensive testing, V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed, and total time, so we plan to retire V4 Pro in an orderly manner. From 12:00 Beijing Time on September 14, 2026, and until V4…”
RainPPR · Hugging Face · Sep 10, 2026 - 12
“Running DeepSeek V4 Flash on an RTX 5080 with 16GB VRAM Under Linux/WSL2 via DS4”
peppe200175 · Hacker News · Aug 12, 2026 - 13
“## Summary The OpenCode Go endpoint (`https://opencode.ai/zen/go/v1/chat/completions`) does not perform prompt caching for `deepseek-v4-flash` or `qwen3.8-max`, while `glm-5.2` on the same endpoint does. Every request re-processes the full prefix from scratch, so Go sessions on DeepSeek/Qwen burn the $12/$30/$60 metered limits much faster than the Go docs' request estimates assume (those assume ~50-80k cached tokens per request). Related: #41125 (same symptom for DeepSeek, reported earlier). ##…”
csalvarado · GitHub · Sep 9, 2026 - 14
“[BUG] DeepSeek 官方模型「思考」开关无效：请求未透传 thinking/disabled，默认 high 推理持续烧 token 环境 - Agora v2.1.0 - Provider：DeepSeek 官方（base_url = api.deepseek.com） - 模型：deepseek-v4-flash / deepseek-v4-flash-vision-exp 问题描述 设置里把「思考」关掉，但每次请求仍返回 reasoning_content（界面显示「思考 xx 秒、调用 x 工具」）， 推理 token 被持续计费，且 UI 上无法关闭。 影响 - 每次调用都会产生额外 reasoning token（计入输出上限与费用） - 用户无法通过「思考」开关关闭，体验上是「关不掉」 - 存在多轮工具调用时输出 token 被进一步放大 复现逻辑（简单） 1. Agora 新建对话，Provider 选 DeepSeek 官方，模型选 deepseek-v4-flash(或任意模型) 2. 设置 →「思考」设为 关闭 3. 发送任意一条消息 4. 观察：即使…”
Eason-hub-114 · GitHub · Sep 6, 2026 - 15
“### What version of the Codex App are you using (From “About Codex” dialog)? 26.905.11957 ### What subscription do you have? ChatGPT Pro ### What platform is your computer? win11 x64 ### What issue are you seeing? I built an auditing system that reconciles Codex's locally recorded token usage telemetry against the ChatGPT Pro weekly allowance meter. When using GPT-5.6 Sol, the observed weekly allowance consumption reconciles very closely with the published credit rates: - Uncached input: 100 cr…”
2203476 · GitHub · Sep 16, 2026 - 16
“## What happens On codex CLI **0.153.4**, a turn opened through the `--control` path (`codex app-server`) runs under codex's **default model** rather than the `model` oneharness resolved from `[harness.codex]`. On 0.145.0 the same path honoured it. The default in 0.153.x is `gpt-6-astra`, which on a ChatGPT Pro plan draws roughly **10× the weekly quota per token** of `gpt-5.6-sol` (fit from the `rate_limits.primary.used_percent` snapshots codex writes into its own rollouts: ~1.0 pp/M uncached i…”
nickderobertis · GitHub · Sep 11, 2026 - 17
“### What issue are you seeing? # OpenAI Charges for GPT-6 Astra and GPT-5.6 Sol, but Delivers GPT-4o — This Is Stealing from Paying Users OpenAI’s interface says **GPT-6 Astra xhigh** or **GPT-5.6 Sol**, and users’ premium allowance is deducted as though those advanced models were actually being delivered. But under multi-account, multi-agent concurrency, the capability delivered by the cloud suddenly collapses to **GPT-4o level**. Affected conversations have even claimed a **June 2024 knowledg…”
Accademia · GitHub · Sep 13, 2026 - 18
“## Client or integration Codex App / Codex CLI / Direct HTTP ## Provider or upstream service xAI (`xai/grok-4.6`) ## OpenCodex version `dev` at `c306fb53c` (also reproduces on published 2.14.0 static seed behavior) ## Endpoint or capability `/v1/responses` and `/v1/chat/completions` reasoning effort mapping / Codex catalog picker ## Current behaviour Selecting Codex `xhigh` for `xai/grok-4.6` is clamped to upstream `high`. `dev` already seeds `grok-4.6`, but the xAI registry still advertises th…”
olddonkey · GitHub · Aug 13, 2026 - 19
“$20 cursor pro plan only, grok 4.6 is good enough and cheap”
armadillo1354 · Hacker News · Aug 14, 2026 - 20
“## Summary A run in which most model calls **errored** is recorded as a clean scored verdict, with the errored calls' cost and tokens counted as `$0.00` / `0`. Three grok-4.6 runs came back as "the model is bad at this task" (2/77, 2/51, 9/65) when in fact their generations were being killed mid-stream by an infrastructure timeout on the caller's side. apo already receives everything it needs to know this: those spans arrive with `gen_ai.response.finish_reasons: ["error"]` **and** OTel span sta…”
samik-bind · GitHub · Aug 19, 2026 - 21
“## Summary `grok-4.6` has no entry in `backend/apo/data/default-model-prices.json`, so every agent-under-test call on a grok run is unpriced. The run's reported cost then consists of **only the harness's own spend** — the `claude-sonnet-4-6` judge and sim-user calls — while the model actually being evaluated contributes `$0`. This is the surviving half of #103. That issue's fix gave `apo runs show` an honest `(partial — N unpriced calls)` annotation, but the number it annotates is still present…”
samik-bind · GitHub · Aug 14, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.