Best Open-Weight Frontier Model 2026: Ranked by Cost
The Short Answer
Cheapest frontier-tier: Tencent Hy4 preview (~$0.834/$2.501 per MTok). Best coding claim: GLM-5.3. Cheapest overall: DeepSeek V4 Flash ($0.22/$0.66 off-peak) for volume, GLM-5.3-Flash ($0.15/$0.50) for multimodal. Largest: Qwen3.8-Max (2.4T parameters). Most permissive licence: GLM-5.3-Flash (MIT) and Qwen3.8-27B (Apache 2.0).
August 2026 was the densest month the open-weight field has had. Three frontier-tier models landed in four weeks.
The Ranking
Costs use a reference task of 30,000 input and 5,000 output tokens.
| # | Model | In / Out per MTok | Task cost | Context | Licence |
|---|---|---|---|---|---|
| 1 | GLM-5.3-Flash | $0.15 / $0.50 | $0.007 | 1M | MIT |
| 2 | DeepSeek V4 Flash | $0.22 / $0.66 † | $0.0099 | 1M | Open weights |
| 3 | DeepSeek V4 Pro | $0.66 / $1.98 † | $0.0297 | 1M | Open weights |
| 4 | Tencent Hy4 preview | ~$0.834 / ~$2.501 | $0.038 | >1M | Published on HF |
| 5 | Muse Spark 1.2 | $1.25 / $4.25 | $0.059 | 1M | Meta terms |
| 6 | GLM-5.3 | $1.40 / $4.40 | $0.064 | 1M | Custom (revenue-gated) |
| 7 | Qwen3.8-Max | $2.00 / $6.00 | $0.090 | 1M | Custom (2.4T) |
| 8 | Kimi K3 | $3.00 / $15.00 | $0.165 | — | Open weights |
† DeepSeek prices are off-peak. Peak (01:00–04:00 and 06:00–10:00 UTC) doubles them.
Closed frontier for reference: GPT-5.6 Sol $4/$20 = $0.22. Claude Opus 5 $5/$25 = $0.275.
Last verified: August 30, 2026.
1. Cheapest Volume Workhorse — GLM-5.3-Flash
$0.15/$0.50 per MTok · MIT licence · 320B total / 18B active · 1M context · image + video input
Released August 26, 2026, this is the value leader and it is not close. At $0.007 per reference task it is roughly 31x cheaper than GPT-5.6 Sol, and it is the only model in that price bracket that accepts video input.
The MIT licence is the other headline. In a year when licences got more restrictive, Z.ai shipped its cheap multimodal model under one of the most permissive terms available.
Use it for: high-volume classification, extraction, summarisation, document and video pipelines, long-context retrieval. Not for: hard multi-file agentic coding. It is a workhorse, not a frontier model. Artificial Analysis put it at an Intelligence Index v4.1.1 score of 57 — excellent cost-per-point, not frontier intelligence.
2. Cheapest Overall — DeepSeek V4 Flash and Pro
Flash $0.22/$0.66 · Pro $0.66/$1.98 · both off-peak · 1M context · 384K max output
DeepSeek repriced on August 16, 2026 to a peak/off-peak model. This is the single most common costing error in the field right now: never quote a flat DeepSeek price.
Off-peak covers 17 hours a day and is exactly half peak. If your load runs in US or EU business hours you get off-peak rates most of the time. Asia-Pacific business hours land largely in peak, which erases much of the advantage — at peak, V4 Flash costs $0.0198 per task, nearly 3x GLM-5.3-Flash.
Use it for: cost-sensitive volume where you control scheduling. Batch work overnight into off-peak windows.
3. Best Frontier-Tier Value — Tencent Hy4 Preview
~$0.834/$2.501 · 770B total / 49B active · >1M context · weights on Hugging Face
Released August 28, 2026. At $0.038 per task it is about one-seventh of GPT-5.6 Sol while claiming genuine frontier-adjacent capability.
Two things distinguish it. Tencent built training data with internal domain experts — software engineers, game developers, finance analysts, security experts — around work they actually ship. And it is marketed on agent orchestration: Hy4 coordinating several Codex sessions in parallel and evaluating their results.
That orchestration angle is underrated. A cheap model that competently supervises expensive models is worth more than a cheap model that merely codes.
Caveat: it is a preview with a thin independent benchmark record. Test before you commit.
4. Best for Coding — GLM-5.3
$1.40/$4.40 · ~743B base, post-trained · 1M context · weights published, revenue-gated licence
Z.ai’s claim is specific: the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench. In-house benchmark, so discount it — but no other model in this list makes a sharper coding claim.
The licence is the story. Z.ai published the weights on Hugging Face but dropped MIT, adding a requirement that providers above $10 billion in revenue undergo a security review. That is aimed at hyperscalers. If you are below that threshold it changes nothing practical; read the file and confirm.
Use it for: hard multi-file agentic coding, long-horizon engineering tasks, cyber defence work.
5. Largest — Qwen3.8-Max
$2/$6 · 2.4T parameters · 1M context · cached input $0.25
Released August 3, 2026 (announced July 19 at WAIC), built for autonomous multi-day coding and research. Alibaba’s framing was that it ranks “second only to Fable 5” — a claim it notably did not back with a published benchmark table at launch.
The open-weight release split: the 2.4T checkpoint uses a custom licence and is text-only — no vision, no native 1M context — while Qwen3.8-27B is genuine Apache 2.0. The hosted endpoint gets capabilities the downloadable weights do not.
Use it for: long-horizon autonomous runs where raw capacity beats per-token price.
The Licence Warning
“Open-weight” stopped meaning one thing in 2026:
- MIT — GLM-5.3-Flash. Genuinely permissive.
- Apache 2.0 — Qwen3.8-27B. Genuinely permissive.
- Revenue-gated custom — GLM-5.3. Fine for most, a gate for hyperscalers.
- Custom, reduced capability — Qwen3.8-Max 2.4T. Text-only weights versus a multimodal endpoint.
Read the licence file, not the press release.
Should You Self-Host?
Usually no, and the parameter counts explain why.
Hy4 preview is 770B total. Qwen3.8-Max is 2.4T. Even at 4-bit quantisation these need multi-node GPU serving. GLM-5.3-Flash needs around 192GB of GPU memory at 4-bit — server-class hardware, roughly $2,000/month of continuous MI300X rental, against a crossover near 14 million tokens per hour of sustained load.
Most teams asking this question are two orders of magnitude below the crossover.
Published weights are worth most as portability insurance and compliance evidence, not as an instruction to run the model yourself. Self-host for data residency, air-gapping, regulatory constraint, or genuinely enormous throughput.
The Recommendation
Route, don’t choose. Send bulk volume to GLM-5.3-Flash or DeepSeek off-peak. Send hard coding to GLM-5.3. Escalate the genuinely difficult 5–10% to Claude Opus 5 or GPT-5.6 Sol. Blended cost typically lands 60–80% below single-model frontier deployment, because most tokens in most production systems are easy.
The August 2026 releases make that pattern cheaper than it has ever been.