Open-Weight vs Frontier Models: Cost and Quality

Open-weight vs frontier models comes down to capability and cost. On general capability, the best open-weight models trail the frontier by about four months, by Epoch AI's measure in May 2026. On price, open-weight models cost several times less per token than the flagship frontier models as of October 2026. On one defined task, training an open model on that task can close the capability gap, and in our published QMSum study the trained models scored above all five hosted proprietary models we tested on ROUGE-1, within the limits set out below.

This post covers what separates the two kinds of model, how large the general gap is, why open weights cost less, which open models lead today, where a trained open model wins with our measured results, and what each costs per task on both cost dimensions.

What is the difference between open-weight and frontier models?

A frontier model is one of the most capable models available at a given time, and almost all of them are closed weight: you reach them through the developer's API, and the developer trains, serves, prices and retires them. An open-weight model publishes its trained parameters, so anyone can download it, host it, and train it further on their own data.

Artificial Analysis puts the line where most of the field does: models count as open weights "where their weights are accessible to download," which "allows self-hosting on your own infrastructure and enables customizing the model such as through fine-tuning" (read 29 October 2026). Open weights are a narrower claim than open source. The training data and training code usually stay private, and the licence decides what you may do commercially, which our guide to open model licences covers.

The UK AI Security Institute summed up the open side in July 2026: open weight models "can be hosted privately, with no data returning to the model providers, adapted to specific tasks, and run only at the cost of compute. Once in use, they offer a dependable base that providers can't change or deprecate."

How far behind frontier models are open-weight models?

About four months on general capability. Epoch AI found that since January 2026 "the most capable open-weight models have lagged frontier closed models by an average of four months" on its Epoch Capabilities Index, with an average gap of 8 index points, "similar to the gap between GPT-5 and GPT-5.5" (Epoch AI, 29 May 2026). Its earlier analysis put the lag at about three months between January 2023 and October 2025.

Measurements on narrower skills land in the same range. The UK AI Security Institute tested cyber capability and found that GLM-5.2 (June 2026) and DeepSeek V4-Pro performed similarly to frontier closed models released 4 to 7 months before them, a narrower gap than the 6 to 10 months it measured through most of 2025 (AISI, 17 July 2026).

Both measure capability across a broad set of tasks. A business running one task at volume asks a smaller question: how good is the model at this job, and what does each answer cost? The general gap is the starting point for that question, and training on the task changes the answer, as the QMSum results below show.

Why are open-weight models cheaper?

Open-weight models are cheaper because anyone can serve the same weights, so hosts compete on price per token, while a closed model has one price set by its developer. On OpenRouter on 29 October 2026, GLM-5.2 was listed on 34 endpoints from 25 providers, with output prices from $1.44 to $8.00 per million tokens. Claude Opus 5.5 was listed on 11 endpoints from 5 providers, every one at or above Anthropic's list price of $4 input and $20 output per million tokens.

Two more reasons sit behind the gap. Hosts pay the developer nothing for most open weights, so their price only has to cover serving. And you can choose the size: open models come in many sizes, and a smaller model costs less to run per token.

The last row is the complication. The frontier labs also sell budget tiers, and GPT-6 Luna undercuts the hosted open models on price per token. Against a budget tier, the case for a trained model rests on the second cost dimension, fewer tokens per task, and on quality: whether the budget tier holds your bar on your task.

Which is the best open-weight model?

For general capability, Artificial Analysis ranks DeepSeek V4.1 Flash 1017 and GLM-5.3, each at maximum reasoning effort, as the highest-intelligence open models, followed by MiMo-V2.6-Pro and Ling 3.1 Flash 1019, as of 29 October 2026 (Artificial Analysis). Rankings like this change often, so check the date on any you rely on.

For a business task, the best open-weight model is the smallest one that holds your quality bar after training on your own examples, under a licence that allows your use. A leaderboard measures broad capability. Your evaluation set measures the job you are paying for, and it can pick a model far smaller than the leaderboard's top entry.

When does a trained open model beat a frontier model?

A trained open model wins where one task runs at volume, the output can be checked, and the quality bar is set by that task. Our QMSum study (arXiv 2609.25028) shows the effect on query-focused meeting summarization, scored under one implementation on the full 281-query test split.

Three results matter for this comparison.

Training moved the open model above the hosted models on ROUGE-1. The untrained 1.2B open model scored 30.12 ROUGE-1, below the best hosted proprietary model at 32.43. Fine-tuning it on the same retrieved-span inputs it reads at inference added 5.29 ROUGE-1 (95% interval +4.02 to +6.56), and it finished at 35.41. Giving the untrained model those retrieved spans in place of the transcript's opening words had added 1.55 on the test split and 0.29 on validation, so most of the gain came from training.

Both trained models scored above all five hosted proprietary models. The hosted models ran zero-shot under one concise prompt and scored 27.89 to 32.43. The paper limits this ordering: the hosted models wrote longer answers, which contributes to the gap on a reference-overlap metric, and no human or factuality evaluation was run. Its conclusions are limited to QMSum and automatic metrics. The strongest published QMSum systems report higher scores in their own papers, up to 38.82 ROUGE-1, so the study's contribution is the comparison under one scorer and the cost result. The hosted versions tested were current when the study ran, and newer ones have shipped since.

The smaller trained model held the quality. The 406M specialist scored 36.33 against the 1.2B system's 35.41, a difference QMSum cannot statistically separate, with less than half the peak inference memory. For a task run at volume, that footprint sets serving cost, which is why we size each model to its task.

The same pattern shows up in where spend goes. In OpenRouter's public data for the 30 days to 12 October 2026, classification, extraction and transformation took 20.8% of classified model spend, and closed frontier models took 75% to 87% of the top ten models' spend on those tasks. These are tasks with checkable outputs, where a trained model can be evaluated precisely. For the wider argument, see is fine-tuning dead?.

What do open-weight and frontier models cost per task?

Cost per task is the price per token multiplied by the tokens each task uses, and a trained open model lowers both. The price falls because open weights are served at competing rates and a right-sized model needs less hardware. The token count falls because the instructions, protocol and examples a general model needs in every prompt are trained into the model, so they stop being sent.

The table uses an illustrative task: 6,000 input tokens, of which 4,500 are repeated instructions and examples, and 500 output tokens. The trained row keeps the 1,500 tokens of per-request data and writes a shorter answer. Prices are the vendors' list prices on 29 October 2026.

The trained row holds the rate at GLM-5.2's to isolate the token effect. A model trained for one task can be much smaller than a general model (the trained summarizers in our QMSum study have 406M and 1.2B parameters), so its rate per token can sit well below GLM-5.2's. The GPT-6 Luna row shows where the per-token case runs out: against a budget tier, the savings come from the token count, and the deciding question is quality on your task.

Agents with heavier prompts save more. When every request carries guardrails, operating procedures, tool definitions and worked examples, the starting prompt can reach 10,000 tokens, and only the per-request data changes from one request to the next. Train all of that into the model and the same 1,500 tokens of data are all that is sent:

Input tokens fall by 85%. The trained model costs less than a quarter of the same open model running the full prompt, and about a thirteenth of Claude Opus 5.5. GPT-6 Luna still costs less per task at these list rates; closing that gap takes the lower rate of a smaller trained model, and quality on your task decides the rest.

Calibrate the token dimension to your own traffic. It scales with how much of each prompt is repeated boilerplate rather than per-request data, and with whether your current model bills reasoning tokens. Prompt caching already discounts repeated input: Anthropic bills cache hits on Claude Opus 5.5 at $0.20 per million tokens, a twentieth of the input rate, as of 29 October 2026. Against a well-cached setup the input saving is smaller, and output and reasoning tokens carry more of the case.

Run your own numbers in the cost-per-task calculator, which takes price per token and tokens per task as separate inputs. Our guides to AI cost per task and when a custom model beats a frontier API on cost work through the break-even with training and hosting included.

When should you use a frontier model instead?

Rent a frontier model while the task is broad, new or low in volume, and train an open model once it is narrow, stable and repeated.

The two work together. Prompting a frontier model is the fastest way to learn what a task needs, and the examples and evaluation set you build while prompting become the training data and the quality bar for the trained model. Hard or unusual requests can stay on a frontier model while the trained model takes the volume.

Ertas, an applied AI lab, runs the adaptive model endpoint for this: a model trained on your task, served through an OpenAI-compatible API. New versions train on your agents' traffic in the background, are scored against your evaluations, and go live only with your sign-off. If a new version falls short, the current model stays in service. It runs Ertas-managed, in your own cloud or on-premise, and your data stays private. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.

Frequently Asked Questions

Is it "open weights models" or "open-weight models"?

Both appear. "Open-weight model" takes a hyphen because "open-weight" describes the noun, the same way as "open-source model". "Open weights" without the hyphen works as a noun phrase, as in "the model has open weights". Epoch AI writes "open-weight models"; the UK AI Security Institute writes "open weight models".

Is ChatGPT a frontier model?

ChatGPT is OpenAI's chat product, and the models behind it are OpenAI's frontier models. Those models are closed weight: you reach them through ChatGPT or OpenAI's API, and their weights are private.

Is ChatGPT open or closed source?

Closed. OpenAI keeps the weights of the models behind ChatGPT private. It also publishes open-weight models under the gpt-oss name, which anyone can download and host; Epoch AI lists gpt-oss-120b as unrestricted.

What is an example of an open-weight model?

GLM-5.2 from Z.ai, DeepSeek V4 Pro, Kimi K2.6 from Moonshot, Gemma 4 from Google DeepMind and gpt-oss from OpenAI are all open-weight models. Licences differ: Epoch AI lists Kimi K2.6 as unrestricted and Gemma 4 31B as restricted use, so check the terms before commercial use.

Why are open-weight models cheaper to run?

Many hosts serve the same open weights and compete on price, and you can pick a model sized to your task. A closed model has one list price set by its developer. As of 29 October 2026, GLM-5.2 costs $4.40 per million output tokens on Together AI against $20 for Claude Opus 5.5.

Can an open-weight model compete with a frontier model?

On general capability the best open models trail by about four months on average, per Epoch AI in May 2026. On a single defined task, training changes the comparison: in our QMSum study, trained open models scored above five hosted proprietary models on ROUGE-1, with the caveats that the hosted models wrote longer answers and no human or factuality evaluation was run.

The Short Version

Frontier models lead on general capability by a few months, and open-weight models cost several times less per token than the flagships. For one task at volume, a right-sized open model trained on that task can pass the hosted models on the task's own measure and send far fewer tokens per request. Use a frontier model while the task is broad or new, and train a model once it is stable and repeated. Put your own prices and token counts into the cost-per-task calculator to see where your crossover sits.

For when one task earns its own model, see foundation models vs task-specific models. For how a trained model keeps improving once it is live, see continual learning for AI agents.

---

Sources, read 29 October 2026: Epoch AI, "Open models lag state-of-the-art closed models by 4 months" (29 May 2026); UK AI Security Institute, "How Far Behind the Frontier are Leading Open Weight Models on Cyber?" (17 July 2026); Artificial Analysis open-weights comparison; pricing from OpenAI, Anthropic and Together AI; provider listings from OpenRouter. Our research: the QMSum study and the OpenRouter structured-task report.