Skip to content
Primores AI wiki

Open-Weight Models: What They Cost and When to Route Work to Them (2026)

Open-Weight Models

TL;DR: An open-weight model is one whose trained weights are published, so you can run it on your own hardware or rent it from any host. Usage is growing fast: Menlo Ventures’ 2026 survey puts open models at 19% of enterprise tokens, up from 11% a year earlier, while their share of spending fell from 6% to 4%. The cheap ones cost a fraction of closed models per token (GLM-5.1 lists from $0.90 / $2.90 per million against $4 / $20 for Claude Opus 5.5). The best one doesn’t: Kimi K3.1 Pro lists at $2 / $10, the same as Claude Sonnet 5.5, and on Artificial Analysis’s index it cost $3.68 per task for a score of 46 where GPT-6.1 Sol at medium effort scored 46 for $0.36. Our read: use open models for routine, checkable steps behind a router, keep a closed model for the hard cases, and compare them on glossary/cost-per-completed-task, not on token price.

⚠️ Read 2026-11-11. Prices are list prices on OpenRouter and vendor pages that day and change often. The Gartner report is paywalled and quoted through The Register. Revolut’s figures are its own. The per-task comparison uses Artificial Analysis’s benchmark tasks, which are harder than the routine steps the savings case is about. Nobody has published a per-task comparison on routine business steps.

What open-weight means

A model’s weights are the numbers it learned in training. A closed model (GPT-6, Claude, Gemini) keeps them on the vendor’s servers and sells access. An open-weight model publishes them, which gives you three ways to use it:

Way to use itWhat you needWhat you pay
Hosted APIAn API key from a host or a marketplace such as OpenRouterPer token, at whatever that host charges
Self-hostedGPUs (rented or owned), serving software, people to run itThe hardware, busy or idle
Post-trainedThe above, plus your own data and training skillThe above, plus training runs

“Open-weight” is not the same as “open source”. The licence decides what you may do, and it varies by model: GLM-5.1 is published under the MIT licence and Qwen3.6-27B under Apache 2.0 (per their Hugging Face cards, as read for our 4 November article), while Kimi K3.1 Pro ships under its own licence, which Artificial Analysis summarises as “Commercial use is allowed with restrictions”. Read the licence for the model you pick.

How far behind they are

Two measurements, two answers:

  • Epoch AI (29 May 2026): from January to May 2026 the best open-weight models trailed the best closed ones “by an average of four months” on Epoch’s capability index, a gap of 8 index points. Epoch adds that the estimate “would grow to six months” under a stricter test of catching up.
  • Artificial Analysis (11 November 2026): the highest-ranked open-weight model, Kimi K3.1 Pro, scores 46 on its Intelligence Index. The top closed model, Claude Fable 5.2 at max effort, scores 65. Of 104 models ranked, 41 are open-weight.

Different indexes, different dates. Both say the best open models are behind. The November snapshot follows a run of closed releases: Claude Opus 5.5 in September, Gemini 4 in October, Claude Fable 5.2 on 9 November.

What they cost per token

List prices per million tokens, input / output, read 11 November 2026:

ModelTypePriceSource
GPT-6 AstraClosed, frontier$10 / $50OpenAI API pricing
Claude Fable 5.2Closed, frontier$8 / $40Anthropic pricing
Claude Opus 5.5Closed$4 / $20Anthropic pricing
Claude Sonnet 5.5, GPT-6.1 SolClosed, mid-tier$2 / $10Anthropic, OpenAI
Kimi K3.1 ProOpen-weight, strongest$1.80–$2.00 / $9–$10 at most hosts; $2 / $10 at Moonshot’s own endpointOpenRouter listing
GLM-5.1Open-weight$0.90 / $2.90 (cheapest host) to $1.40 / $4.40 (Z.ai and most others)OpenRouter listing
Claude Haiku 5.5, GPT-6 LunaClosed, small$0.10 / $0.50Anthropic, OpenAI

Two things in that table go against the usual story:

  • The strongest open model is priced like a mid-tier closed one. Kimi K3.1 Pro lists at Sonnet 5.5’s price.
  • The cheapest models on the list are closed. Haiku 5.5 and GPT-6 Luna list below every GLM-5.1 host. Menlo’s report describes the same thing: “The price war is at the low end. Frontier labs have cut small-model prices by 80% or more.”

What buyers actually pay is lower on both sides. Menlo estimates the average open-source token at $0.13 per million in September 2026, down 3.3 times in a year, against $0.70 for the blended closed-model token. Self-hosting can go lower still at scale: Revolut says running GLM-5.1 on its own GPUs cost about $0.30 per million tokens, a figure that leaves out hardware ownership and staff.

What they cost per task

Per token is the wrong unit for a decision (see glossary/cost-per-completed-task). Artificial Analysis reports what it cost to run its index, per task, at list prices and with measured cache rates. Read 11 November 2026:

Model and settingTypeIndex scoreCost per index task
Claude Opus 5.5, mediumClosed48$2.45
Kimi K3.1 ProOpen-weight46$3.68
GPT-6.1 Sol, mediumClosed46$0.36
DeepSeek V4.1 Flash, maxOpen-weight33$0.60
GLM-5.3, maxOpen-weight32$5.01
Claude Sonnet 5.5, mediumClosed31$0.78
Claude Haiku 5.5, highClosed30$0.16

On these tasks the open models were not the cheap option at their score, with DeepSeek V4.1 Flash the closest. Kimi K3.1 Pro cost ten times what GPT-6.1 Sol did for the same score, and GLM-5.3 at max effort cost about thirty times what Haiku 5.5 did for a similar one.

Read that with its limits. These are hard multi-step benchmark tasks at list prices, the kind of work the routing advice below sends to a frontier model anyway. The savings case is about short routine steps, and Gartner’s analysts name the mechanism that erodes it: open models “can be more verbose and retry more often”. Our read: the per-token gap is real, the per-task gap has to be measured on your own steps, and on hard tasks it can run the other way.

Who is using them, and how much

Menlo Ventures, “2026: The State of Enterprise AI” (3 November 2026). A survey of nearly 400 US enterprise decision-makers, fielded 8–26 September 2026, plus Menlo’s own market model:

  • Open-source models’ share of tokens rose from 11% to 19% in a year. Their share of spending fell from 6% to 4%, which Menlo attributes to a 3.3x drop in their average token price.
  • Volume is up 22.5 times, to an estimated 2.9 quadrillion tokens a month.
  • Chinese models take 89% of open-source spend: Kimi 61%, DeepSeek 15%, GLM 11%, Qwen 3%. Meta’s Llama is the largest Western one at 7%.
  • “Buyers continue to prioritize capability over cost in model selection.” Frontier models’ share of spending rose from 35% to 63% over the same year, and near-frontier models (Grok and the open ones together) hold “just under 5% of LLM API spend”.
  • Post-training your own open model is still rare: 10% of AI builders, up from 6%.

Menlo marks Anthropic, OpenRouter, Fireworks and Modal as its portfolio companies, so it has money on both sides of this question.

Revolut (reported 3 November 2026). Its ops and infrastructure lead told a conference that Revolut made self-hosted GLM its default coding model for about 2,000 engineers, first GLM-5.1 and then GLM-5.3 on 56 Nvidia B200 GPUs. At about $0.30 per million tokens, GLM-5.1 “cost about a tenth of using OpenAI’s Codex or Anthropic’s Claude Code”, and AI coding inference spend fell by roughly 90% in under a year. GLM didn’t replace the closed models: it is text-only, and a router sends vision work to Gemini or GPT Luna. All of it is Revolut’s own account, with no hardware or staffing costs given.

The forecast

Gartner analysts Mark Petty and Raymond Paquet, in a late-October 2026 report titled “Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps”, as quoted by The Register:

  • “Nine in 10 agent workflow steps are routine, and open-weight models handle them at parity for a quarter of the price of the cheapest closed model.”
  • Against a frontier model on every step, an open model “costs as little as a twentieth”. After verbosity and retries, “the saving per task against a frontier model is closer to five or six times.”
  • “By 2030, 90% of enterprise operational agent workflow steps will execute on open-weight models.”
  • The saving is not their main argument: “Open-weights preserve your control. The cost saving is a bonus.”

It is a forecast, labelled by Gartner as a planning assumption. Set it beside Menlo’s survey: usage is moving the way Gartner expects, and spending is not.

Where they fail

  • Vision and UI work. Many open models are text-only, GLM among them, which is why Revolut routes vision elsewhere. Some do take images. Check per model.
  • Long, multi-step planning. The AgentFloor benchmark (May 2026: 16 open models up to 32B parameters against GPT-5) found small open models “already sufficient for much of the short-horizon, structured tool use work”, with the gap showing “most clearly on long-horizon planning tasks”, where “neither side reaches strong reliability.”
  • Tool calls with strict JSON output. A June 2026 study found that with tool calling and a JSON schema both switched on, several open-weight models stopped calling tools while still producing valid JSON.
  • Delegated coding on small local models. An October 2026 report by one engineer, with Claude Opus 5.5 as the parent model: in one suite, direct coding passed 45 of 45 checks and delegation to a small local worker passed 12 of 45 (37 and 8 if sessions stopped by a spend limit count as failures); a different worker passed 0 of 45. In a separate study delegation passed 12 and 14 of 15 where direct coding passed 15 of 15. In the studies with a direct comparison, direct coding came out cheaper.
  • Idle hardware. A rented GPU bills the same busy or idle. One break-even analysis found self-hosting a 70B model on a single GPU 52 times more expensive than a hosted version of the same model, with “no break-even at any volume”. GPU rental prices also rose in 2026 (see glossary/ai-compute-costs).
  • Buyer and regulator scrutiny. With 89% of open-source spend on Chinese models, expect questions from government and regulated customers. Two US House committees wrote to Airbnb and to Cursor’s parent company about their use of Chinese models in April 2026 (Semafor, as cited in our article). Revolut’s position is that the concern is about APIs run from China, not weights run on your own or regional servers. Ask your buyers before you choose.

How to route work

The move most businesses can make is routing: a cheap model by default, a stronger one on exception. It is the same idea as the glossary/advisor-strategy, applied across vendors.

Kind of stepWhere it goesWhy
Routine and checkable: classification, extraction, summaries, ticket triage, first-pass draftsA hosted open model, or a small closed one, with a stronger model as fallbackShort structured steps are where small models are sufficient
Hard reasoning, long multi-step work, vision, anything customer-facing where an error is costlyA closed frontier modelWhere open models still trail, and where per-task cost favoured closed models in the data above
Data that must stay in a region or on your serversA regional host first; your own GPUs only if that isn’t enoughControl is the reason, not price
Very high, steady volumeConsider self-hostingOnly pays when the hardware stays busy

Before moving a step, test it: run about 50 real requests through the candidate with the same prompts and have someone grade the outputs blind against your current model. That is our rule of thumb from the article, not a published standard.

FAQ

What is an open-weight model?

A model whose trained weights are published, so anyone can run it on their own hardware or rent it from a hosting company. Kimi, GLM, DeepSeek, Qwen and Llama are examples.

Are open-weight models cheaper than ChatGPT or Claude?

Per token, the cheap ones are far below mid-tier and frontier closed models, but not below the smallest closed models (Claude Haiku 5.5 and GPT-6 Luna at $0.10 / $0.50). Per task it depends on the work: on Artificial Analysis’s benchmark tasks in November 2026, the best open model cost more per task than closed models at the same score.

How much of AI usage runs on open models?

About 19% of enterprise tokens and 4% of enterprise model spending in September 2026, per Menlo Ventures’ survey.

Is open-weight the same as open source?

No. Open-weight means the weights are downloadable. The licence, which ranges from MIT and Apache 2.0 to custom licences with commercial restrictions, decides what you may do with them.

Should a small business self-host a model?

Usually not. Hosted open models give most of the per-token saving with no hardware, and an idle rented GPU costs the same as a busy one. Our article Should you self-host an LLM? works through the thresholds.

Do not cite

ClaimWhy not
”Open models are 20 times cheaper”Gartner’s own realistic figure is five or six times per task; twenty is the best case against a frontier model on every step
”Open models are always cheaper than closed ones”Haiku 5.5 and GPT-6 Luna list below every GLM-5.1 host; Kimi K3.1 Pro lists at Sonnet 5.5’s price and cost more per benchmark task
”Enterprise use of open-source models is falling”That was Menlo’s December 2025 reading (19% to 11%). Its November 2026 report has open models at 19% of tokens, up from 11% a year earlier, with spend share down to 4%
“Revolut cut AI costs 10x” as a verified resultSelf-reported; hardware and staff costs not disclosed
”Open models are four months behind” as a current figureEpoch’s window is January to May 2026, and the estimate is six months under its stricter test
”90% of agent steps will run on open models” as a factA Gartner forecast for 2030

Key Takeaways

  • Open-weight means the weights are published. The licence still decides what you can do.
  • Usage is rising (19% of enterprise tokens) and spending isn’t (4%): buyers still pay for capability.
  • Cheap open models undercut mid-tier and frontier closed models per token, not the smallest closed ones.
  • Per task, on hard benchmark work, open models were not the cheap option in November 2026. Nobody has measured routine steps.
  • The practical use is routing: open or small models for routine checkable steps, a frontier model on exception.
  • Self-hosting is for data control or very steady volume, not for small-team savings.

Sources

By Andrej Ruckij