Open-weight AI models are four months behind and 9x cheaper

Kimi K2.6 is the strongest of the open-weight AI models on Epoch AI’s Capabilities Index. It scores 151.60 and lists on OpenRouter at $0.95 per million input tokens and $4.00 per million output tokens. GPT-5.5 scores 158.22 and lists at $5 and $30. A lead of 6.6 index points costs 6.6 times as much on a blended basis.
A Gartner note published Nov. 12 explains that spread. Epoch AI measures the best open-weight models trailing the closed frontier by an average of four months and eight index points since January 2026. Gartner puts the lag at 12 months at the start of 2025 and under four months today, and it assumes one month by 2029. With the free model only four months behind, a per-token markup for raw intelligence gets hard to defend.
Gartner’s sharper point is about who pays once the license fee disappears. It calls free open models an illusion, because the host absorbs the compute bill, the MLOps payroll and all legal and compliance liability. For security and risk leaders, the open-weight decision is therefore a liability decision, and engineering teams often make it without pricing it that way.
I built this analysis from Epoch AI’s open versus closed capability data (May 29, 2026), OpenRouter list prices for 11 models on Nov. 14, 2026, and Gartner’s Strategies to Compete With Free Open-Weight AI Models (G00864526, Nov. 12, 2026). Blended prices, ratios and medians are my calculations.
Key findings
- The best open-weight AI models trail the frontier by four months and eight capability points. Epoch AI measures that average since January 2026, and calls eight points similar to the gap between GPT-5 and GPT-5.5.
- Gartner expects the lag to fall to one month by 2029. Gartner puts it at 12 months at the start of 2025 and under four months today, so about two thirds of it closed in under two years.
- Closed models cost nine times more at the median. Across 11 models priced on OpenRouter, the median closed model blends to $6.69 per million tokens and the median open-weight model to $0.74.
- Frontier output tokens cost up to 200 times more. GPT-5.5 lists output at $30 per million tokens, against $4.00 for Kimi K2.6 and $0.15 for gpt-oss-120b.
- Self-hosting moves all legal and compliance liability to the buyer. Gartner says OpenAI and Anthropic indemnify outputs and IP claims by contract, and neither yet guarantees agentic transactions.
- Gartner expects 55% of AI platform providers to sell outcome-priced tiers by 2029. Every outcome guarantee needs telemetry to prove it, and that telemetry is a data flow security teams have to review.
Open-weight AI models sit four months and eight points behind

Epoch’s index rolls many benchmarks into one capability score. The chart plots 16 models from that index, six closed and ten open-weight, released between May 2025 and May 2026. All six closed models arrived between February and May 2026 and score between 154.28 and 158.22. GPT-5.5 leads at 158.22, followed by Gemini 3.5 Flash at 156.31 and Claude Opus 4.7 at 156.18. By comparison, the strongest open-weight model, Kimi K2.6, scores 151.60. Kimi K2.5 follows at 148.22 and MiniMax-M2.5 at 147.44.
Averaged across the 16, closed models score 156.1 and open-weight models 145.9. The 16 are my selection, however, and they understate the top of the closed field. For example, the sample leaves out Epoch’s highest score, GPT-5.5 Pro at 159.35, and five other closed models released in 2026 that outscore Kimi K2.6. Epoch’s own average gap since January 2026 is eight points, which it calls similar to the gap between GPT-5 and GPT-5.5.
Gartner says the July 2026 launch of Kimi K3 signaled that open models could go neck and neck with closed ones on long-context reasoning, complex agentic workflows and tool execution. Epoch’s May 29 analysis predates that launch, so the four-month figure describes the field before Kimi K3 arrived.
Gartner says open-weight models closed two thirds of the lag in under two years

Gartner puts the average time for open-weight models to match closed ones at 12 months at the start of 2025 and under four months today. Epoch’s average since January 2026 is four months. Gartner’s strategic planning assumption is one month by 2029, and that figure is an assumption rather than a measurement.
The two sources do not measure the lag the same way. Epoch notes that its October 2025 analysis found an average of three months for January 2023 through October 2025, well below Gartner’s 12-month starting point. On Epoch’s figures, in other words, the lag has held between three and four months. The two-thirds decline is therefore Gartner’s estimate, and Epoch’s data does not show it.
A one-month lag changes procurement math. At that pace, any capability a closed provider ships would reach open weights before most enterprises finish a pilot. Gartner’s read is that competing on raw model intelligence stops being a sustainable strategy for providers, and the value moves to orchestration, context and trust.
Frontier output tokens cost up to 200 times more than open-weight models

OpenRouter list prices show how wide the spread runs. GPT-5.5 lists at $5 per million input tokens and $30 per million output tokens. Claude Opus 4.7 lists at $5 and $25, and Gemini 3.5 Flash at $1.50 and $9. On the open side, Kimi K2.6 lists at $0.95 and $4.00 and GLM-5 at $1.00 and $3.20. Kimi K2.5 lists at $0.49 and $2.50, GLM-4.7 at $0.40 and $1.75, MiniMax-M2.5 at $0.27 and $0.95, DeepSeek-V3.2 at $0.29 and $0.43, and gpt-oss-120b at $0.03 and $0.15.
I then blended each price at three input tokens per output token. The median closed model comes to $6.69 per million tokens and the median open-weight model to $0.74, a 9x gap. Grok 4.20 breaks the pattern, though. It lists at $1.25 and $2.50, a blended $1.56 that undercuts Kimi K2.6 at $1.71, so closed does not always mean expensive.
These are list prices, so they overstate what many buyers pay. Caching and provider discounts lower the bill, and prices vary by hosting provider for the same open model. On Nov. 14, for example, OpenRouter showed temporary discounts of 59% on Kimi K2.6, 40% on GLM-5 and 28% on DeepSeek-V3.2, and I used the undiscounted rates. None of them include the cost of running the model yourself, which is where Gartner’s argument starts.
Free weights move the bill to the buyer

Gartner lists three costs that transfer to an enterprise hosting open weights. They are compute infrastructure to serve the model, MLOps talent to run it, and all legal and compliance liability for what it produces.
The third cost is the one security leaders should own. Gartner says OpenAI, through its service terms, and Anthropic, through expanded commercial API protections, provide contractual indemnification for generated outputs and IP claims. It also notes that neither provider yet financially guarantees agentic transactions, and treats today’s copyright indemnity as a sign of where risk absorption is heading.
Download the weights and the indemnity stays behind. If an agent built on a self-hosted model takes an action that triggers a claim, then the claim lands on the enterprise. Gartner adds that managed open-weight API providers such as OpenRouter offer far cheaper inference than frontier models and are adding premium compliance features, including sovereign AI and data localization.
Gartner’s advice to model providers follows directly. Stop selling model capability and sell enterprise risk absorption instead, bundling API use with explicit financial indemnification for agentic transaction errors, performance guarantees and compliance tooling. In effect, the premium buys the use of someone else’s balance sheet.
Routing becomes the control plane
Gartner expects standalone orchestration middleware to shrink as open models handle multistep reasoning and tool use natively. What survives is hybrid routing that weighs cost, latency and liability for each request. It names Portkey, Cloudflare AI Gateway and OpenRouter as platforms that send complex reasoning to a closed frontier model and routine background work to open models.
Palo Alto Networks agreed to acquire Portkey on April 30, 2026 and completed the deal on May 29, calling it a centralized control plane to manage and protect autonomous AI agents. The April announcement says Portkey already processes trillions of tokens a month and gives access to more than 3,000 LLMs and MCP tools. My read is that a security platform vendor bought the gateway because it sees every prompt, every model selection and every agent action in one place. I covered how the large platforms are pulling ahead in the vendor tape analysis.
For a CISO, that reframes the open-weight question. The choice is between model selection that happens inside a control plane you can audit and model selection buried in application code you cannot see.
Gartner’s strategy map runs across three layers

Gartner’s note maps eight pivots across three layers of the AI stack. Model providers get four of them, and three of the four depend on assets an open-weight model cannot download, whether a balance sheet to underwrite risk, custom silicon or live enterprise data. Orchestration platforms get one, hybrid routing. SaaS providers get three, each built on something the customer has and the model does not, such as outcomes, perimeter control or workflow telemetry.
Each pivot also carries a market dependency in Gartner’s telling. Risk absorption depends on whether nondeterministic failures can be priced actuarially. Customer-controlled deployments depend on buyers funding GPU clusters and MLOps staff. Telemetry moats last only as long as foundation models do not absorb the same features from below.
Outcome pricing needs a measurement you trust
Gartner’s second planning assumption is that by 2029, 55% of AI platform providers will offer outcome-priced tiers with a completion record as a product feature. Its live example is Cognition, which guarantees Devin will deliver at least as much engineering value as a customer consumes and issues usage credits of up to $10 million when it falls short. Cognition measures value in equivalent engineering hours, using an estimator that runs on every completed task.
Gartner also points to Harvey fine-tuning open models inside a client’s virtual private cloud for zero data retention, and to Cursor logging micro-interactions such as cursor position and tab-accept rates to build a telemetry moat. It is candid that defining success for nondeterministic output is hard, and that proving the value of something subjective is a different problem from counting engineering hours.
The security version of that problem is sharper. A vendor that guarantees an outcome has to instrument the work to prove it, and instrumentation means telemetry leaving the environment. Every outcome guarantee is also a data flow, and procurement should review it as one.
Sixteen open-weight and closed AI models on the index

Epoch publishes an uncertainty range with each score, which the table leaves out. Several models appear in Epoch’s data at more than one reasoning-effort setting, and the table uses the setting Epoch lists in its comparison. Five of the 16 are not priced here, so the price medians rest on 11 models.
What security leaders should do about open-weight AI models
- Find every self-hosted open-weight model already running in your environment. Then find who signed off on absorbing its liability. Often nobody did, because nobody framed it as a liability decision.
- Treat the AI gateway as a security control. Routing decisions about which model sees which data belong in a control plane with an audit trail, and Palo Alto Networks’ Portkey acquisition shows where the market is putting that control. Gartner’s AI Governance Hype Cycle groups AI gateways with AI agent identity under agentic AI oversight, which I covered in 34% of enterprises govern AI with policies they only partly follow. Gartner’s first AI Governance Hype Cycle shows CISOs what to fund first..
- Read frontier contracts for what is actually indemnified. Gartner’s reading is that output and IP claims are covered today and agentic transaction errors are not. That line matters most once agents start taking actions.
- Price open weights on total cost. A 9x list-price advantage at the median shrinks once compute, MLOps staff and uninsured liability go into the comparison.
- Ask every outcome-priced vendor what it measures and where that data goes. The guarantee is only as safe as the telemetry behind it.
Frequently asked questions
How far behind closed AI models are open-weight models?
Epoch AI measures the best open-weight models trailing the closed frontier by an average of four months and eight points on its Capabilities Index since January 2026. Gartner expects the lag to fall to one month by 2029.
How much cheaper are open-weight models?
On OpenRouter list prices from Nov. 14, 2026, the median open-weight model in this sample blends to $0.74 per million tokens and the median closed model to $6.69, about nine times more.
Are open-weight AI models free?
The weights are free to download. Gartner says the enterprise hosting them pays for compute and MLOps talent and absorbs all legal and compliance liability.
Do AI providers indemnify customers?
Gartner says OpenAI and Anthropic provide contractual indemnification for generated outputs and IP claims, and that neither yet financially guarantees agentic transactions.
What is outcome-based AI pricing?
Pricing tied to delivered results rather than seats or tokens. Gartner expects 55% of AI platform providers to offer outcome-priced tiers by 2029, and cites Cognition’s guarantee of up to $10 million in credits for Devin.
How I built this analysis
Capability scores and release dates come from Epoch AI’s open versus closed data insight, published May 29, 2026 by Jack Edwards and Luke Emberson. Open or closed status follows Epoch’s release-access labels, except for Gemini 3.5 Flash, which carries no label in Epoch’s data and which I treat as closed. The 16 models are my selection of six closed and ten open-weight releases, so the sample averages are illustrative. The four-month and eight-point figures are Epoch’s own. The 12-month, under-four-month and one-month lag figures and the 55% outcome-pricing assumption are Gartner’s, from G00864526.
Prices are OpenRouter list prices per million tokens captured Nov. 14, 2026 from each model’s page, before caching and discounts. Where a page showed a temporary discount, as the Kimi K2.6, GLM-5 and DeepSeek-V3.2 pages did, I used the undiscounted list rate. Blended price weights input three to one over output. Medians, ratios and averages are my calculations, and the 11 priced models are a sample rather than the full market. Adding or removing a model moves the median ratio.
More analysis from this blog
- Security vendors are outgrowing Gartner’s 12.7% forecast (Nov. 16, 2026)
- Gartner’s secure AI Hype Cycle predicts twice the incidents without preproduction offensive GenAI testing (Nov. 5, 2026)
- Gartner’s $4.05T AI forecast and Forrester’s Predictions 2027 put the AI spending slowdown in 2028 (Nov. 2, 2026)
- AI revenue gap forecasts compared across Gartner, Goldman, Bain and JPMorgan (Oct. 29, 2026)
- AI infrastructure spending forecasts compared: every major firm raised its 2030 number in 2026 (Oct. 22, 2026)
- Gartner’s $239B AI forecast: Agentic workflows take half of GenAI model revenue (Sept. 21, 2026)
- 34% of enterprises govern AI with policies they only partly follow. Gartner’s first AI Governance Hype Cycle shows CISOs what to fund first. (Aug. 10, 2026)
- Gartner’s $5.95 trillion AI forecast puts the chatbot era on a 2027 deadline (Aug. 5, 2026)
This post is my personal analysis of Gartner’s research and public sources and does not represent my employer.
Sources
- Gartner, Strategies to Compete With Free Open-Weight AI Models (G00864526), Nov. 12, 2026.
- Epoch AI, Open models lag state-of-the-art closed models, Jack Edwards and Luke Emberson, May 29, 2026.
- Palo Alto Networks to Acquire Portkey to Secure the Rise of AI Agents, Palo Alto Networks, April 30, 2026.
- Palo Alto Networks Completes Acquisition of Portkey, Palo Alto Networks, May 29, 2026.
- The AI Productivity Guarantee, Cognition.
- Kimi K3, Moonshot AI.
- OpenRouter model pages for GPT-5.5, Claude Opus 4.7, Gemini 3.5 Flash, Grok 4.20, Kimi K2.6, Kimi K2.5, GLM-5, MiniMax-M2.5, DeepSeek-V3.2, GLM-4.7, gpt-oss-120b, accessed Nov. 14, 2026.











































