Recommendation for Budget / High volume
Best Cheap LLM for High Volume
Our top recommendation for Best Cheap LLM for High Volume, based on the public evidence we track, is OpenAI: GPT-5.6 Sol.[1][2][3][4] Complex agentic workflows requiring high reasoning quality: beats Claude Fable 5 by 11.4 points at medium reasoning for roughly one-quarter the estimated cost. Watch out: Cost excludes it from pure high-volume classification: Sol runs $5/$30 per 1M tokens versus Luna's $0.20/$1.20, a 25× gap that erodes at scale. DeepSeek: DeepSeek V4 Flash 0423 is the next-ranked alternative. Extreme cost sensitivity with off-peak scheduling: at $0.014/$0.44/$1.32 per 1M tokens during off-peak windows, it undercuts most alternatives for batch processing.
About this recommendation
- Updated
- Sep 25, 2026
- Evidence through
- Sep 25, 2026
- Sources
- 23
- Revision
- v77
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
67%
intended feed weight
Largest provider share
2 of 6
OpenAI
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| price weight | 40% | 99/100 | 20/20 |
| Route reliabilityunavailable | 30% | feed unavailable | 0/20 |
| LiveBench Instruction Following | 15% | #17 | 20/20 |
| OpenRouter usage | 10% | 97/100 | 20/20 |
| LMArena Text | 5% | #13 | 19/20 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- OpenAI2 models
- Anthropic1 model
- deepseek1 model
- Google1 model
- Qwen1 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 | GPT-5.6 SolOpenAI | 74 | 67% | 12 threads · 8 families · 4 cautions | #13 LMArena Text · #17 LiveBench Instruction Following |
| 02 | DeepSeek V4 Flash 0423deepseek | 73 | 67% | 23 threads · 11 families · 12 cautions | #40 LiveBench Instruction Following · #53 LMArena Text |
| 03 | Gemini 3.6 FlashGoogle | 71 | 67% | 7 threads · 6 families · 3 cautions | #8 LiveBench Instruction Following · #16 LMArena Text |
| 04 | Qwen3.8 27BQwen | 71 | 67% | 8 threads · 8 families · 3 cautions | #14 LiveBench Instruction Following · #52 LMArena Text |
| 05 | GPT-5.6 LunaOpenAI | 71 | 67% | 12 threads · 10 families · 5 cautions | #39 LMArena Text · #45 LiveBench Instruction Following |
| 06 | Claude Opus 4.6Anthropic | 68 | 67% | 7 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.
GPT-5.6 Sol ranks #13 on LMArena with strong reasoning efficiency.
Best when: Complex agentic workflows requiring high reasoning quality: beats Claude Fable 5 by 11.4 points at medium reasoning for roughly one-quarter the estimated cost.
Tips
- Complex agentic workflows requiring high reasoning quality: beats Claude Fable 5 by 11.4 points at medium reasoning for roughly one-quarter the estimated cost.
Watch out for
- Cost excludes it from pure high-volume classification: Sol runs $5/$30 per 1M tokens versus Luna's $0.20/$1.20, a 25× gap that erodes at scale.
DeepSeek V4 Flash offers aggressive peak/off-peak pricing as low as $0.014 cache hit / $0.44 cache miss / $1.32 output.
Best when: Extreme cost sensitivity with off-peak scheduling: at $0.014/$0.44/$1.32 per 1M tokens during off-peak windows, it undercuts most alternatives for batch processing.
Tips
- Extreme cost sensitivity with off-peak scheduling: at $0.014/$0.44/$1.32 per 1M tokens during off-peak windows, it undercuts most alternatives for batch processing.
Watch out for
- Silent stream failures: SSE endpoints intermittently stop sending chunks mid-generation without closing the connection, leaving requests hanging indefinitely.
- Broken prompt caching: cache hit rate drops to 0% or below 10% on repeated identical prompts via OpenCode Go endpoint, inflating real costs far above quoted rates.
- Uncontrollable reasoning: thinking/disabled parameter is not respected, forcing high-reasoning token burn even when explicitly disabled.
Gemini 3.6 Flash offers competitive $0.75/$3.75 pricing with Vertex MaaS availability, though its thinking-model architecture can silently return empty responses when output budgets are constrained.
Best when: Long-context companion agents with stable prefixes: prompt caching exists for AI Studio and is being added to Vertex, reducing repeated 50K-token system prompt costs.
Tips
- Long-context companion agents with stable prefixes: prompt caching exists for AI Studio and is being added to Vertex, reducing repeated 50K-token system prompt costs.
- Per-action routing economics: verified reachable on production keys with ~32% blended cost reduction when routing simple queries to Flash Lite and complex ones to Flash.
Watch out for
- Silent failures on token-constrained requests: with thinking enabled, small max_tokens values return HTTP 200 with empty choices and no error indication.
- Price uncertainty: LiteLLM's bundled price table was verified wrong for some Vertex models, requiring live rate validation.
Qwen3.8 27B is an open-weight model runnable on consumer hardware with strong throughput on RTX 5090 and GB10, though deployment requires careful handling of speculative decoding cache bugs and distributed reasoning budget controls.
Best when: Self-hosted high-volume processing at $0.38/$2.60 per 1M tokens with 140 tok/s speculative decoding on single RTX Pro 6000, eliminating per-request API costs entirely.
Tips
- Self-hosted high-volume processing at $0.38/$2.60 per 1M tokens with 140 tok/s speculative decoding on single RTX Pro 6000, eliminating per-request API costs entirely.
- Mac Studio deployment for moderate scale: M5 Ultra achieves 48 tok/s at 8K context, M3 Ultra 31 tok/s, enabling on-premise bulk processing without cloud egress.
Watch out for
- Prefix cache corruption with speculative decoding: EAGLE/MTP drops the last cached block, causing 1,648-token recomputes and 30-40% throughput loss on prefix-reusing workloads.
- Unbounded reasoning in distributed setups: default xhigh reasoning cannot be constrained via thinking_budget on TP2 clusters, risking token exhaustion before tool calls complete.
GPT-5.6 Luna delivers roughly 12× lower cost than GPT-5.4 at $0.20/$1.20 per 1M tokens, making it purpose-built for high-volume throughput workloads where per-token economics dominate.
Best when: Route classification and bulk document processing where sub-$0.001 per draft cost lets you process millions of records without quota anxiety.
Tips
- Route classification and bulk document processing where sub-$0.001 per draft cost lets you process millions of records without quota anxiety.
- Production environments needing predictable spend: Luna's 7/8 accuracy on structured extraction tasks matches Gemini 3.7 Flash at one-quarter the per-request cost.
Watch out for
- Aggressive rate limits hit hard: one 95K-token request consumed 48% of a 200K TPM budget, and production logs show sustained 503 errors during peak periods.
- Prompt caching barely functions: operational data shows 0.84% cache hit rate versus 20.5% on predecessor models, inflating real costs above headline pricing.
Claude Opus 4.6 ranks #2 on LMArena with exceptional quality.
Best when: Consider only after reviewing the cited caution.
Watch out for
- Prohibitive token economics: at $5/$30 per 1M tokens, a single heavy user processing 281B tokens could generate $1.4M+ in costs, and monthly quotas burn through in single days under agentic load.
- Opaque platform pricing: cost multipliers vary dramatically (1× to 27×) across platforms despite identical API pricing, making true cost unpredictable.
Frequently asked
- What is the top-ranked model for Best Cheap LLM for High Volume?
- OpenAI: GPT-5.6 Sol ranks first in the current evidence-weighted comparison. Complex agentic workflows requiring high reasoning quality: beats Claude Fable 5 by 11.4 points at medium reasoning for roughly one-quarter the estimated cost.[1]
- What should I watch out for with OpenAI: GPT-5.6 Sol?
- Cost excludes it from pure high-volume classification: Sol runs $5/$30 per 1M tokens versus Luna's $0.20/$1.20, a 25× gap that erodes at scale.[2]
Sources
- 1
“"On Agents’ Last Exam (opens in a new window), an evaluation of long-running professional workflows across 55 fields, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT‑5.6 Terra and GPT‑5.6 Luna outperform Fable 5 at around one-s…”
saberience · Hacker News · Jul 9, 2026 - 2
“## Motivation OpenAI shipped the GPT-5.6 family 2026-07-09 (Sol/Terra/Luna) and cut Terra/Luna prices 2026-07-30. Current `openai/pro` is **gpt-5.4 ($2.50/$15 per 1M)**. **Terra ($2/$12)** is 20% cheaper and one generation newer (reportedly GPT-5.5-class); **Luna ($0.20/$1.20)** is ~12× cheaper for throughput roles. Sol ($5/$30) is 2× our current cost — not a candidate without eval evidence. Secondary benefit: cheaper tokens directly reduce the month-end OpenAI credit-exhaustion risk that cause…”
rozavala · GitHub · Aug 1, 2026 - 3
“Two separate things, both in the DeepSeek rows of `data/model_data.ts`. ### 1. The prices are from the previous generation `deepseek-v4-pro` and `deepseek-v4-flash` carry a single flat rate each: DeepSeek's page (read 2026-08-23, ) gives per 1M tokens, USD: | model | band | cache hit | cache miss | output | | --- | --- | --- | --- | --- | | `deepseek-v4-pro` | peak | 0.044 | 1.32 | 3.96 | | `deepseek-v4-pro` | off-peak | 0.022 | 0.66 | 1.98 | | `deepseek-v4-flash` | peak | 0.014 | 0.44 | 1.32 |…”
xyzs996 · GitHub · Aug 23, 2026 - 4
“## Summary DeepSeek updated the official API pricing for `deepseek-v4-flash` and `deepseek-v4-pro`, effective at 2026-08-17 00:00 Beijing time. The current pricing uses separate peak and off-peak rates, while the DeepSeek V4 entries in `genai-prices` still contain the previous flat USD rates. The official pricing page currently identifies the deployed model versions as DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813. ## Current official USD pricing Prices below are per 1M tokens. | Usage | V4 F…”
Imyukehan · GitHub · Aug 19, 2026 - 5
“## Summary DeepSeek endpoints intermittently stop sending SSE chunks mid-generation — typically mid-`reasoning_content` (thinking) — without closing the connection, without a terminal `finish_reason` chunk, and without `[DONE]`. The stream just goes silent. Observed on both the official `api.deepseek.com` and the `opencode-go` proxy, across multiple DeepSeek models (`deepseek-v4-flash`, `deepseek-v4-flash-vision-exp`). Other providers on the same transport (e.g. kimi-code/k3) never exhibit this…”
iacore · GitHub · Aug 25, 2026 - 6
“## Summary The DeepSeek OpenAI-compatible `/chat/completions` SSE endpoint (both `api.deepseek.com` and the `opencode-go` proxy) **intermittently stops sending chunks mid-generation without closing the connection** — typically mid-`reasoning_content` (thinking) — with no terminal `finish_reason` chunk and no `[DONE]`. The stream simply goes silent and stays open indefinitely. Observed across multiple models: - `deepseek-v4-flash` - `deepseek-v4-flash-vision-exp` This does **not** happen with ot…”
iacore · GitHub · Aug 26, 2026 - 7
“This issue might be a duplicate of existing issues. Please check: - #41125: OpenCode Go endpoint: no prompt caching on DeepSeek (same symptom — cache_hit always 0 on repeated identical prompts via the Go endpoint; you already referenced this one) - #42935: OpenCode Go quota exhausted in ~20 minutes after DeepSeek V4 Flash cache reads suddenly dropped to 0 (same root cause, quota impact) - #43218: Opencodego cache hit rate extremely low (<10%) when used via ccSwitch with Claude Code / Codex (bro…”
github-actions[bot] · GitHub · Sep 9, 2026 - 8
“[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 - 9
“### Summary Add explicit prompt caching (cachedContent) support to the google-vertex provider, matching the caching that already exists for the AI Studio (google-generative-ai) provider. ### Problem to solve The Gemini prompt-cache path (google-prompt-cache.ts) only works for AI Studio. Vertex users re-send their full stable context on every request. We run a long-lived companion agent on google-vertex/gemini-3.6-flash with a ~50K-token stable prefix (system prompt + memory) per message; withou…”
jbyolo29 · GitHub · Jul 30, 2026 - 10
“## 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 · GitHub · Jul 28, 2026 - 11
“## What `gemini-3.6-flash` is a thinking model. It spends output budget on reasoning tokens before emitting any text, so a request with a small `max_tokens` returns **HTTP 200 with `"choices": []`** — no error, no message, no indication anything went wrong. Measured against the live proxy on 2026-07-29, prompt `"Reply with exactly: OK"`: | `max_tokens` | `choices` | reasoning tokens | text tokens | | --- | --- | --- | --- | | 16 | **empty** | 13 | 0 | | 32 | **empty** | 29 | 0 | | 64 | `"OK"` |…”
ujjawalmisra · GitHub · Jul 29, 2026 - 12
“## What LiteLLM computes spend from `model_prices_and_context_window.json`, bundled in its container image. That table is wrong for Vertex MaaS models. Measured 2026-08-04 against the live deployment, via `GET /model/info`: | Model | LiteLLM | Actual Vertex rate | | | --- | --- | --- | --- | | `gemini-3.6-flash` | $1.50 / $7.50 | $1.50 / $7.50 | correct | | `gemini-3.5-flash-lite` | $0.30 / $2.50 | $0.30 / $2.50 | correct | | `qwen3-coder` | **$1.00 / $4.00** | **$0.22 / $1.80** | wrong | Actua…”
ujjawalmisra · GitHub · Aug 4, 2026 - 13
“served it the day after the drop, one rtx pro 6000, nvfp4+q5_k. 140 tok/s spec (p50), 0.156s first token, native 262k. image and video on the same endpoint. engine: https://github.com/avifenesh/memra try it: https://inference.tiyuvta.ai/app $0.38 in / $0.20 cache / $2.60 out”
Avifenesh · Hugging Face · Aug 15, 2026 - 14
“The numbers I was most interested in are tucked away in a chart towards the bottom - the speed comparison of the Mac Studios v.s. a RTX 5090: Qwen3.8 27B tokens sec generation speed Prompt size 8K 64K 128K 256K RTX 5090 PC 59 51 44 n a M5 Ultra 48 39 32 24 M3 Ultra 31 23.5 20 15 A whole bunch more comparison numbers in this section: https: www.macstories.net stories m5-ultra-mac-studio-revie...”
simonw · Hacker News · Sep 21, 2026 - 15
“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 - 16
“**Is your feature request related to a problem? Please describe.** On a distributed TP2 cluster (oMLX 0.6.0.dev1, 2× M4 Max 48 GB, ring), Qwen3.8-27B's default reasoning is xhigh and there's no way to bound it: `thinking_budget` is rejected by the distributed engine, and even when forwarded the budget is not enforced. Agentic sessions either run unbounded xhigh (burning the whole max_tokens budget in reasoning, cut off at finish_reason=length before a tool call) or run with reasoning fully off.…”
xunlinkx · GitHub · Aug 16, 2026 - 17
“## 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 · GitHub · Aug 17, 2026 - 18
“## Context Split out of the TPM rate-limit fallback work (see the linked issue). That change keeps the athlete served when `gpt-5.6-luna` hits its tokens-per-minute ceiling, by falling back to another model's TPM bucket. **This issue is about the underlying pressure**, not the failover. From Sentry issue [7588764791](https://large-language-music-llc.sentry.io/issues/7588764791): One request consumed **95,794 tokens ≈ 48% of the entire 200,000 TPM budget**. Confirmed to originate from bulk file-…”
zew1me · GitHub · Jul 27, 2026 - 19
“## Problem Statement 用户希望 Storyflow 在托管模型频繁失败时继续完成当前工作，而不是反复等待同一个故障模型、手工重发消息或重新选择模型。系统应能自动切换到访问协议和任务能力兼容的备用模型，同时清楚说明实际使用的模型，并避免重试放大、历史损坏和工具重复执行。 2026-09-16 的线上只读调查给出了以下基线；这是调查时的历史快照，不是实施时仍然成立的服务状态： | 窗口/现象 | 观测 | | --- | --- | | 最近 24 小时，截至北京时间 17:10 | 约 14,769 次模型 HTTP 尝试，815 次上游 HTTP 错误，约 5.52% | | gpt-5.6-luna | 24 小时 334/713 失败；16:10–17:10 的 93/93 失败 | | 最新复核，17:08:44–17:18:44 | Luna 16/16 返回上游 503 | | 跨模型 403 | 16:41–16:58 集中发生，约 465 次，涉及 14 个调用身份、7 个模型 | | gemini-3.1-flash-lite | 七天 790 次…”
littlewwwhite · GitHub · Sep 16, 2026 - 20
“## Problem High-volume conversation-processing routes were moved to `gpt-5.6-luna`, but their callers still use the pre-GPT-5.6 prompt-cache contract: a global `prompt_cache_key` plus `prompt_cache_retention="24h"`, without GPT-5.6 request-wide cache options or explicit cache breakpoints. **Operational observation from a privacy-safe closed production analysis (2026-07-23 17:00–23:00 UTC):** - the Luna population read **0.84%** of input tokens from cache, versus **20.5%** for the preceding matc…”
Git-on-my-level · GitHub · Jul 24, 2026 - 21
“>281 billion tokens. Using the least expensive version of Claude Opus 4.6, which costs $5 for every million tokens, that one user alone could have cost Meta more than $1.4 million. And that is based on the pricing that they sell a dollar for less than a dollar so the costs may be much much more.”
anshumankmr · Hacker News · Jul 14, 2026 - 22
“Yeah, how many agents can you people even run at once and how much does it cost you? In company we used the monthly token quota and nowadays it's basically unusable with claude opus 4.6 on high reasoning. You can basically burn through 100% usage through a single day. How does it even scale for you with N agents and which magical plans or models do you use, where tools like this are even viable?”
Tooster · Hacker News · May 22, 2026 - 23
“It's interesting that the cost multiplier for Claude Sonnet 4 4.5 4.6 varies so much (1 6 9), while the API cost is exactly the same for all three models. Also, the multiplier of 27 for Claude Opus 4.6 4. is way higher than the increase in API price would suggest. I wonder why that is.”
johndough · Hacker News · Apr 27, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.