Dollars per Task: The New Metric That Makes the On-Premises Case

The Metric That Prices What You Actually Buy

Per-token pricing is a unit cost, not a bill. Two models listed at the same price per million tokens can cost wildly different amounts to finish the same job, because the smarter one burns fewer tokens to get there and the chattier one burns more. For two years buyers have been comparing frontier models on a metric that does not map to what they pay. The capability leaderboard told you which model was smartest. It did not tell you which model was cheapest to put to work.

Cost per Task fixes that. It multiplies a model's per-token price by the tokens that model actually consumes to complete a benchmark task, and reports the result in dollars. One number. The real cost of one unit of work. Artificial Analysis rolled it into Intelligence Index v4.1 on June 16, 2026, alongside Time per Task and Tokens per Task. The index itself weights nine evaluations toward agentic work — Agents 34% / Coding 24% / Science 24% / General 18%, with a 95% confidence interval of ±1. Add the cost dimension and the leaderboard stops being a capability chart and becomes a buying chart, which is what a buyer needed all along.

That reframing matters because the open-weights frontier has caught up on capability. Once you measure both axes at once — how smart, and how much per task — the case for running these models yourself stops being ideological and becomes arithmetic.

What the Leaderboard Showed: v4.1 (June 2026)

The v4.1 table reads top to bottom in index points, with the new cost and time columns laid bare. The proprietary models sit above the open ones on raw intelligence; the open models sit below them on price by an order of magnitude:

Read it carefully. Fable 5 topped the index at 60 — but on June 12, 2026, four days before this index shipped, a US Commerce Department order forced Anthropic to disable it. (Commerce lifted the order on June 30, and access was restored the next day.) Strip away the model nobody could use at the time, and the best model you could actually buy was Claude Opus 4.8 at index 56, costing $1.78 per task. GPT-5.5 (xhigh) sat one point back at 55 for $0.99. Then the open models: DeepSeek V4 Pro, MiniMax-M3, and Kimi K2.6 clustered at 43 to 44 — within roughly twelve index points of the available flagship — and DeepSeek V4 Pro did it for $0.04 per task.

The honest read is that the open-weights frontier is no longer a capability compromise. It is within a dozen points of the best model you can buy, and it costs between twenty-five and forty-five times less per task. The capability gap that once justified proprietary pricing has collapsed. On the one metric that maps to a bill, the open models win — and that is before you run them on hardware in your office, where the cost-per-task column does not shrink but disappear.

GLM-5.2 and Kimi K2.7 Code shipped in June 2026, after this index was published. Both published vendor benchmarks at the open-weight coding frontier (GLM-5.2: SWE-bench Pro 62.1%, FrontierSWE 74.4%; Kimi K2.7 Code: beats Opus 4.8 on MCP Mark Verified). Artificial Analysis later scored GLM-5.2 at 51 on v4.1, which made it the open leader in June 2026 — above the cluster in this table, as we expected. See the side-by-side: Open Source AI Models Compared.

The Leaderboard in October 2026: v4.3

Artificial Analysis rescaled the index on September 7, 2026. Version 4.3 swapped in Terminal-Bench 4.0 and added AutomationBench-AA, so every model’s score dropped and every cost per task rose. These numbers can’t be compared with the v4.1 table above. Here is the v4.3.2 picture, per Artificial Analysis:

Two things are true at once. The gap to the very top widened: Claude Opus 5.5 at max effort scores 58, twelve points above the best open model, MiMo-V2.6-Pro at 46. And the cost gap grew: MiMo-V2.6-Pro runs at about $0.13 per task against $1.99–$7.63 for the 53-point closed tier, and GLM-5.3-Flash matches Opus 5.5 at low effort (42) for less than half its cost per task. One caution: Kimi K3, an open model, still costs about $2.00 per task on Artificial Analysis’s numbers. Open weights don’t make an API cheap. Running them on hardware in your office does.

The Catch With Even the Cheapest Cloud Model

DeepSeek V4 Pro at $0.04 per task is the best value in the table by a wide margin — and it is still a cloud API. That matters for three reasons that have nothing to do with price, and they are the reasons the cost-per-task number is not the whole story.

A cloud API is a meter. Every task is $0.04 out of pocket, forever, scaling linearly with usage. At a thousand tasks a day that is $40 a day; at ten thousand it is $400. The rate is set by a vendor that can change it without your say — flagship prices have climbed with each new tier, and promotional launch rates expire.

The data leaves your network. DeepSeek's API runs on infrastructure subject to China's National Intelligence Law, which can compel cooperation with state intelligence work — the same sovereignty problem that hangs over Z.ai's GLM API. For a business handling customer data, proprietary code, or anything compliance-sensitive, that architecture is not acceptable no matter how cheap the model is.

Access is a lease, not ownership. A vendor can pull an endpoint for regulation, geopolitics, or business reasons — as Anthropic had to with Fable 5 on June 12, 2026. When the endpoint goes, your workflow breaks overnight, and the $0.04-per-task efficiency you built a process around is gone. Fable 5 came back when the order was lifted on June 30; until then, any workflow built on it was simply off.

The open weights are the escape hatch. DeepSeek V4 Pro, GLM-5.2, MiniMax-M3, and Kimi K2.7 Code all ship downloadable weights under permissive licenses. The cheapest way to use them is also the way that solves all three problems at once: run them yourself.

On-Premises: Where Cost per Task Goes to Zero

On hardware in your office, the cost-per-task column does not shrink — it disappears. There is no per-token price to multiply by tokens used. The marginal cost of a task is the electricity to run hardware on a fixed monthly plan. The capability is the model's; the economics are yours.

A Faraday Scale plan — one Mac Studio with an M5 Ultra chip and 512 GB of unified memory, $999 US a month with hardware and support included for the full 36-month term — runs DeepSeek V4.1 Flash (as a 1- to 3-bit build of about 250–350 GB, which trades some quality on long or precise work; a 4-bit build of about 400–445 GB needs a larger setup), MiniMax M3, GLM-5.3-Flash, MiMo-V2.6-Flash, and Qwen3.8-27B, all at zero per-token cost. Kimi K3, GLM-5.3, and MiMo-V2.6-Pro are available on larger clusters. That is $999 a month, $35,964 over the term, fixed regardless of how many tasks you run. That is the plan that turns the cost-per-task leaderboard's most important number — the $0 — from a slogan into your operating model.

The Math for a Real Team

Put a 10-person engineering team behind the cloud numbers. Claude Code Max at $200 a seat is $2,000 a month. GPT-5.5 Codex for the complex tasks the seats can't handle adds roughly $2,500. And the per-task meter that Artificial Analysis v4.1 just made explicit — call it about 800 tasks a month on Claude Opus 4.8 at $1.78 each — adds another $1,400. Cloud stack: roughly $5,900 a month, every month, climbing with usage and with whatever the vendors charge next.

Those are June 2026 prices. On the v4.3 index, the same 800 tasks on Claude Opus 5.5 come to about $440 a month at low effort ($0.55 each), $2,770 at xhigh ($3.46), and $4,780 at max ($5.98). The meter now has a dial on it — and the more capability you ask for, the faster it spins.

The same team on a Faraday Scale plan: $999 a month, fixed. Same frontier-class open models. Zero per-token meter. Data stays in the building.

That is about $61,200 a year you stop paying cloud vendors — for infrastructure that runs the cost-per-task winners at $0 per task and keeps your data under your control. The metric Artificial Analysis just published is the same math a buyer does in their head; on-premises is the answer that has been hiding in plain sight behind the per-token price list.

References