LLM Cost per Task: Are Cheap Models Really Cheap?

Token prices are easy to compare, but they do not tell you what a completed task will cost. Two models can charge very different rates and still produce nearly the same bill for the same work.

Alibaba’s Qwen3.8-Max charges $6 per million output tokens and OpenAI’s GPT-5.6 Sol charges $20, according to their published price pages. On the same set of hard tasks, Qwen3.8-Max costs about $0.91 per task and GPT-5.6 Sol about $0.95.

The output-token rate differs by more than three times. The task cost differs by four cents. At 2,000 tasks a month, that is roughly $1,820 for Qwen rather than the $570 a token-price comparison might suggest, versus $1,900 for Sol.

The short version

• Price per million tokens is a unit price. It does not tell you how many tokens or agent turns a job will require.
• Token price does not reliably rank models by total cost. Some models preserve most of their apparent discount; others lose nearly all of it.
• Compare cost per successful task on your own workload. The method at the end of this article can be run with a small evaluation set.

Providers publish dollars per million tokens, so model comparisons use that figure. Our comparison of leading models uses it to build a shortlist. Cost per task shows what happens when those models do the work.

Why the sticker price stopped predicting the bill

Three things happened, and they multiply together.

Models decide how long to think. Modern models generate hidden reasoning tokens before answering, and those tokens are billable. Two models can return equally short answers while consuming very different amounts of reasoning. Artificial Analysis rates Qwen3.8-Max maximally verbose, describing it as “notably slow and very verbose.”

Models decide how many turns to take. Most serious AI work now runs as an agent: read, act, check, retry. One model may finish in a handful of steps while another keeps checking and revising. More round trips mean more tokens, even when the final deliverable is the same.

Every turn re-reads the earlier context. On turn 40, the model receives turns 1 through 39 again. More turns increase both generation and input costs, which is why providers price prompt caching separately.

A low token rate cannot compensate for inefficient reasoning and repeated turns in every workload.

Where the price page and the invoice disagree

The Cost per task column is Artificial Analysis’s measurement: they run every model through the same suite of hard tasks and total the real bill, counting input, cached, reasoning and answer tokens alike. The last column is our arithmetic on top of theirs.

Rows are sorted by Intelligence Index score, highest first. Models with the same score are grouped together.

Reading the table.

• Every price and measurement was re-checked against Artificial Analysis on September 1, 2026. Cost per task is its published measurement; Intelligence is its composite score on a 0–100 scale. Claude Fable 5.1, released that day, enters at 66 and $3.69 per task on the same $10 / $50 rate card as Fable 5. The other prices, scores, and task costs are unchanged from August 31. GPT-5.6 Sol remains the comparison point after its cut from $5 / $30 to $4 / $20, with a measured task cost of $0.95.
• One model, several figures. Artificial Analysis publishes a separate measurement for every reasoning-effort setting a model exposes, so a single model can carry several scores and costs. We quote the highest measured setting for each model and compare like with like below. Read the multiples as approximate because the benchmark and its costing method continue to evolve.
• The sticker side of the last column is output price only, because output is where the rate cards differ most. The cost-per-task side counts everything: input, cached input, reasoning and answer. Models with relatively flat cards - DeepSeek V4 Flash charges 3x its input price on output - therefore look like they lose more of their discount than they do.
• Open-weight pricing still depends on the endpoint. The table follows the provider and rate shown on each current Artificial Analysis model page. Another host, or your own hardware, can produce a different bill for the same weights.

Three things the table says that the price pages don’t

Most models lose some of their apparent discount. Cost per task includes input, cached input, reasoning, and answer tokens. GLM-5.3-Flash is 40x cheaper than GPT-5.6 Sol per output token but 11x cheaper per task. Kimi K3’s smaller discount moves from 1.3x to 1.1x.

A smaller discount can still matter. DeepSeek V4 Flash moves from 15x cheaper on output tokens to 8.6x cheaper per task and remains one of the least expensive capable models in the table.

Grok 4.6 reverses the comparison. It is 3.3x cheaper than Sol per output token but 1.3x more expensive per task. Claude Sonnet 5 is 2x cheaper per output token and 1.8x more expensive per task. Every current Claude reasoning tier costs more per task than its token rate suggests, while Haiku 4.5, which uses little reasoning, preserves more of the discount. Anthropic’s current tokenizer also bills roughly 30% more tokens for the same text than the models it replaced. Since OpenAI’s August cuts, Luna and Terra show the opposite pattern and cost slightly less per task than their rates suggest.

The gap, drawn

Each line compares the discount implied by the output-token price (hollow marker) with the discount in measured cost per task (solid marker). Longer lines show a larger disagreement.

Five things this number can’t see

Cost per task beats price per token, and it still comes off a benchmark.

• It prices attempts, not successes. The metric counts what a model spent, never whether the answer was usable. A model that fails fast and cheap scores beautifully. In production you re-run the failures, and often a person reads the output before it ships, so a model that’s 10% cheaper per attempt and wrong 30% more often is more expensive in every way that matters. No public leaderboard measures cost per finished task, because nobody but you can judge finished.
• The “task” is Artificial Analysis’s, not yours. The average blends long agentic runs, terminal coding, graduate science questions and tool-calling exercises, with roughly half of it agentic work. That’s the right call for a general-purpose number in 2026, and it’s also why your own ranking will come out much flatter if your job is summarizing support tickets.
• It’s list-price, text-only, and blind to the fine print. Caching is modeled at each model’s typical hit rate rather than the rate you’ll get. Grok roughly doubles its rate above ~200K-token prompts, and the whole GPT-5.6 line does the same above 272K. Batch endpoints can halve the bill for anything that isn’t time-sensitive, and open-weight pricing changes from one host to another.
• Some models have no cost per task at all. Motif 3 has a published intelligence score and no published cost per task, and probably never will: its weights are free and there is no priced first-party endpoint to measure, so the only bill is your own hardware. Inkling and NVIDIA’s Nemotron 3.5 Lightning were in the same position when this post first went up and now have measured figures, $0.34 and $0.08, because a host started charging for them. That is not a gap in the data so much as a different question. Where a model has no price, “what does it cost” becomes a capacity-planning exercise rather than a lookup.
• It moves for reasons that have nothing to do with the model. GPT-5.6 Luna’s cost per task fell from $0.21 to $0.05 in July because OpenAI cut its price 80%. Open-weight figures can also move when the benchmark changes which provider endpoint it reports. Claude Sonnet 5’s figure would have risen by half in September, until Anthropic canceled the increase on August 10. Any cost-per-task number needs a date attached, or it means nothing. That includes ours.

The dial matters more than the model

Modern models expose a reasoning effort setting, which controls how long the model may think before answering, and it moves cost further than swapping models does. Claude Opus 5 is published at five effort settings, and across them its token consumption on the same benchmark suite spans roughly eight-fold, from 12 million tokens at the lowest setting to 100 million at the highest. Same model, same rate card, same tasks.

Claude Fable 5.1 makes the same point in dollars. At its top setting it scores 66 for $3.69 a task, the dearest figure in the table. Two notches down, at high, it scores 62 for $1.43: the same score Fable 5 posts at $3.14, and one point below Claude Opus 5’s 63 at $2.34. At medium it scores 60 for $1.00, level with Kimi K3, GLM-5.3 and Grok 4.6 at their top settings. One dial on one model, and the run costs 3.7x less than the figure in the table.

Labs usually submit models at maximum effort because it produces the highest score. Published task costs therefore tend to reflect the most expensive setting rather than the one a team would use for routine work. Lowering reasoning effort is often the fastest way to reduce cost, and a frontier model at a lower setting may beat a mid-tier model on both quality and price.

Work out your own, in an afternoon

Every API already reports what you need.

1. Collect ten real tasks, with the messy context attached. Not clean samples: the actual ones.
2. Run each through two or three candidates, at the reasoning effort you would actually use in production.
3. Read the token counts off the response rather than estimating. Every provider returns a usage object with input, output and cached counts, and reasoning tokens sit inside the output number. Multiply by the rate card separately for input, cached input and output, because those three prices can differ by an order of magnitude.
4. Divide by the results you’d actually ship, not the number you ran. If eight of ten came back usable, your cost per finished task is the total divided by eight. That division is the one no leaderboard can do for you.

Ten tasks are not statistically robust, and a larger evaluation may produce a different winner. They are still enough to expose large cost or quality differences. For a more rigorous process, read our guide to choosing and testing a model.

What to do with this number

Token prices remain useful for estimating a single request. They become less useful when models choose how long to reason, how many turns to take, and how much context to read again.

Before changing models to reduce cost:

1. Run the same real tasks through both models.
2. Use the reasoning settings you would deploy.
3. Count only results you would accept.
4. Compare total cost per accepted result.

The cheapest model depends on the job, effort setting, and date measured.

MindsHub treats the model as a choice rather than a fixed dependency. MindsHub Cowork helps people complete knowledge work with open-source agent harnesses. Unified Inference lets developers compare and switch models through one API without rebuilding the integration. View current pricing.

Frequently asked questions

What is cost per task for an LLM? The average total bill to finish one real task, counting every token the model consumed: input, cached input, hidden reasoning, and the visible answer. Artificial Analysis publishes it for its Intelligence Index suite, and it regularly ranks models in a different order than their price pages do.

Does a cheaper model per token ever cost more per task? Yes, and it isn’t rare. Claude Sonnet 5 is twice as cheap as GPT-5.6 Sol per output token and costs 81% more per task, $1.72 against $0.95. Grok 4.6 is the starker case: 3.3x cheaper per output token, 1.3x dearer per task. Anthropic’s reasoning tiers all behave this way relative to their sticker price, because they’re built to think longer and persist through more steps; Haiku 4.5, which barely reasons, does the opposite.

Can the same model get more expensive without a price change? Yes. A version update can change how long a model reasons or how many turns it takes without changing the headline token rate. Run the new version through your evaluation set before moving production traffic, because the rate card alone will not show that difference.

Does reasoning effort change cost more than switching models? Often, yes. Claude Opus 5’s token consumption spans roughly eight-fold across its five published effort settings at one unchanged price per token. Leaderboards typically report the highest setting, so the published figure is usually a worst case, and turning the dial down is the fastest saving most teams have available. Claude Fable 5.1 shows the same thing in dollars: $3.69 a task at its top setting, $1.43 at high for a 62 that matches Fable 5 at $3.14, and $1.00 at medium for a 60 that ties Kimi K3, GLM-5.3 and Grok 4.6. It can even be free: Grok 4.6 at its top setting costs $1.23 a task and scores 60, while one notch down it costs $0.94 and scores 61.

Why do published cost-per-task numbers keep changing? Three reasons that aren’t the model: providers change prices, benchmark suites get revised, and the costing method itself gets updated. GPT-5.6 Luna went from $0.21 to $0.05 per task on a price cut. Always check the date on the number, and for an open model, check which host the measurement uses.

---

MindsHub opens access to leading AI intelligence without locking people into one model or provider. Cowork helps people complete knowledge work with open-source agent harnesses. Unified Inference helps developers build intelligence into their own products through one API.