LLM Cost per Task: Are Cheap Models Really Cheap?

A lower token price looks like an easy saving. It stays a saving only if the model can finish the work without spending the difference on extra reasoning and retries.

The bill depends on the whole attempt: how much context the model reads, how long it reasons, which tools it calls, and whether you have to run the job again. A rate card gives you the price of each token. It cannot tell you how many the job will need.

Take Qwen3.8-Max (0902) and GPT-6 Astra. Their output rates are $6 and $50 per million tokens. On Artificial Analysis’s current benchmark, however, Qwen costs $10.02 per task and Astra at max effort costs $6.24. At today’s prices, Qwen’s output rate is 88% lower, while its measured task cost is 61% higher. Qwen also scores lower: 34 against Astra’s 57.

The point is to measure how a model works through the job. Some low-price models retain a substantial cost advantage. Others consume enough tokens to lose it.

What cost per task measures

Artificial Analysis calculates a weighted average cost across its Intelligence Index evaluations. For each evaluation, it accounts for input, cache reads and writes, reasoning, and answer tokens, divides by the task count, and applies the evaluation’s index weight. It counts only the model’s own calls, not the cost of grading them.

That is more useful than token price alone, but it still measures benchmark attempts. It does not divide spending by the number of results your team would accept, or include your staff’s review time.

For a model shortlist and input capabilities, use our LLM comparison. This article focuses on the bill.

The November 2026 cost snapshot

The table uses Intelligence Index v5.0.1, checked November 3, 2026, and the same models and settings as the comparison article. Rows are sorted by rounded score, highest first; ties go to the lower published cost, with unpriced rows last within a tie. Token prices are US dollars per million uncached input and output tokens.

We show the highest published effort, or the single published reasoning row where no effort is named. Version 5.0 changed the evaluations behind both columns, and most of its weight is now on tasks where the model uses tools. Most scores fell and most task costs roughly doubled, so neither column can be compared with earlier snapshots. Wonder Full has no public API token price or published task cost; its missing values do not mean free usage. Google announced Gemini 4 Argon on September 30 and made it generally available through its API on October 13. It appears at the only setting Artificial Analysis publishes, at the $4 / $20 that Google charges from December 1. These are reference settings; the largest effort label does not guarantee the best result or establish a maximum bill.

Eleven models from our October 19 table are left out. Artificial Analysis has not scored nine of them on v5.0.1 at the setting that table used, and now lists most of them as deprecated: GPT-5.6 Sol and GPT-5.6 Luna, Claude Opus 5, Fable 5 and Sonnet 5, Grok 4.6, Kimi K3, DeepSeek V4.1 Flash and Qwen3.8-Flash-Next. Their scores on the earlier index are not comparable with this table. The other two, GPT-6 Sol and GPT-5.6 Terra, now have v5.0.1 scores at max, but Artificial Analysis lists both as deprecated, and the table includes their successors, GPT-6.1 Sol and GPT-6.1 Terra.

The table uses endpoint rates recorded by Artificial Analysis: Standard pricing for Muse Wonder and Spark 1.3, reference endpoint pricing for Spark Open, and peak token rates for DeepSeek. Spark Open’s rates are not a self-hosting cost. For Wonder Full, N/A means no published API rate or benchmark task cost, not free usage. Cached input, time-of-day discounts, batch processing, long prompts, and tool fees can change your bill.

Gemini 3.8 Flash’s $0.75 / $3.75 is introductory through December 31, 2026 and rises to $1.50 / $7.50 on January 1, 2027. Gemini 4 Argon appears at $4 / $20, the rate Google lists from December 1 and the one Artificial Analysis costs it at; until November 30, Google charges an introductory $2 / $10. Claude Haiku 5.5’s rates apply only to prompts up to 100,000 tokens; a longer prompt pays $0.50 / $2.50 on the whole request. Check the provider terms for the requests you will send; MindsHub’s rate card is separate.

GPT-6.1 Sol: a high score at a low task cost

OpenAI released GPT-6.1 Sol on September 29 in its API as gpt-6.1-sol, listed at $2 input / $10 output per million tokens for prompts up to 272K input tokens, with higher rates above that. Its cached input costs $0.10 per million. Artificial Analysis’s entry scores it 55 at $1.27 per task at max. That is the lowest task cost among the four models that score 55 or more; the other three cost $6.24 to $10.04.

Claude Sonnet 5.5 lists the same $2 / $10 rates and the same $0.10 price for cached input. It also scores 55, but costs $10.02 per task: about 7.9 times as much. Artificial Analysis’s cost breakdown shows where the difference comes from. GPT-6.1 Sol writes about 68k output tokens per task against about 285k, so its output costs $0.68 against $2.85. Its cache reads cost $0.25 against $4.88, which means Sonnet 5.5 re-reads about 19 times as much cached context.

Claude Haiku 5.5 and GPT-6 Luna: same rates, different bills

Claude Haiku 5.5 and GPT-6 Luna list identical token rates: $0.10 per million input tokens and $0.50 per million output tokens. At max effort, Artificial Analysis measures Haiku 5.5 at 36 and $0.41 per task, and Luna at 28 and $0.15. The same token rates produce about 2.7 times the benchmark cost.

The difference is how many tokens each model uses. Haiku 5.5 writes about 226k output tokens per task at max, against about 75k for Luna, and re-reads about two and a half times as much cached context. Artificial Analysis’s breakdown puts its input and cache cost at $0.30 per task against $0.11, and its output cost at $0.11 against $0.04. It also scores eight points higher, so the useful question is whether those points change the results you accept. Effort narrows the gap: Haiku 5.5’s xhigh row scores 35 at $0.25, and its high row scores 30 at $0.16, two points above Luna at max for about the same cost.

Anthropic’s $0.10 / $0.50 also applies only to prompts up to 100,000 tokens. A longer prompt pays five times as much on the whole request, and cached tokens count toward that limit, so long-context work needs its own estimate.

Muse: compare the exact model and access route

Muse Wonder scores 44 at $6.95 per task at max effort, using Meta’s $4 / $20 Standard token rates. GPT-6.1 Sol scores 55 at $1.27 per task, so on this benchmark Wonder costs about 5.5 times as much for a lower score. Test it on the work you would give it before paying the difference.

Muse Wonder Full scores 50, but has no published task cost or public API token rate. It is available through qualifying Muse Code subscriptions. Do not use Wonder’s $6.95 as an estimate for Full or divide a subscription price by an assumed task allowance.

Muse Spark Open scores 25 at $2.02 per task at xhigh, using $1.25 / $4.25 reference endpoint rates. Spark 1.3 scores 37 at $2.39 at max with those same headline rates. They are distinct checkpoints and configurations. Spark Open’s downloadable weights give you another hosting option; the benchmark cost is not the cost of running your own GPUs.

Where the newer low-cost models fit

MiMo-V2.6-Pro reaches 30 at $0.26 per benchmark task, using the rates recorded by Artificial Analysis. MiMo-V2.6-Flash reaches 27 at $0.13, a point below GPT-6 Luna’s 28 at $0.15. Artificial Analysis calls it slower than average and very verbose, so its low task cost comes from its $0.14 / $0.28 rates, not from short answers. These are useful candidates for a budget evaluation, especially when their supported input types match the work. DeepSeek V4.1 Flash and Qwen3.8-Flash-Next have no v5.0.1 score, so this snapshot does not compare them.

Two details keep the comparison honest. First, a model name can hide a release change: the current Qwen3.8-Max (0902) row scores 34 at $10.02, while the older 0803 row had different measurements. Flash-Next is a separate model, not a lower effort setting. Second, the DeepSeek row is the V4 Pro 0813 checkpoint. DeepSeek’s API names can move to a newer release, as its older V4 Flash names did when V4.1 Flash launched, so record the checkpoint behind an alias, not only its name. See the DeepSeek section of the comparison for the routing details.

A token discount may not reach the bill

Sonnet 5.5 shows how little of a token discount can reach the bill. Its $2 / $10 rates are half Opus 5.5’s, but its max-effort task cost is about the same: $10.02 against $10.04. It scores 55 against 59. Both results include the default fallback.

GPT-6.1 Terra shows a smaller version of the same effect. Its $1 / $5 rates are half GPT-6.1 Sol’s, but its max-effort task cost is only about a quarter lower: $0.96 against $1.27. It scores 51 against 55.

Equal rates do not promise equal bills either. Grok 4.7 and Qwen3.8-Max (0902) both list $2 input / $6 output per million tokens. Grok scores 35 at xhigh and Qwen 34, yet Qwen costs $10.02 per task against Grok’s $7.77, about 29% more.

Cache prices can move the bill on their own

On October 7, Anthropic halved Sonnet 5.5’s cache-read price, from $0.20 to $0.10 per million tokens, and left its input and output rates unchanged. Even at the lower price, cache reads make up $4.88 of its $10.02 task cost, nearly half, because agentic tasks re-read a growing context at each step. Cache writes can matter as much: they make up $7.03 of Gemini 3.8 Flash’s $7.77. If your workload caches heavily, compare cache prices as carefully as the headline rates.

Try a lower effort setting

Opus 5.5 max scores 59 at $10.04; xhigh scores 57 at $6.05, and high scores 52 at $3.46. GPT-6 Astra max scores 57 at $6.24, against 55 at both xhigh and high, for $3.66 and $3.24. GPT-6.1 Sol max scores 55 at $1.27; high scores 50 at $0.59.

One model scores higher one step below the top: Claude Fable 5.1 scores 50 at xhigh for $10.85, against 49 at max for $12.99. Grok 4.7 scores the same 35 at high as at xhigh, for $6.18 against $7.77.

Sonnet 5.5 makes an especially large move: max scores 55 at $10.02, xhigh scores 50 at $4.09, and high scores 42 at $1.64. The cheaper setting gives up measured capability, so judge the saving against your acceptance criteria.

Those tradeoffs give you a useful experiment: run your tasks at both settings. Keep the extra reasoning where it improves accepted results, and lower it where it only adds time and cost.

Read discounted tiers as contracts

Gemini 4 Argon is the clearest recent case. Artificial Analysis scores it 50 at $7.68 per task, at the $4 / $20 that Google lists from December 1. That is about 23% more than GPT-6 Astra’s $6.24, and Astra also scores higher: 57 against 50. Until November 30, Google charges an introductory $2 / $10, at which the same run would cost half as much, about $3.84.

Argon’s standard rates are 60% below Astra’s, yet it averages about 116k output tokens per task against about 52k, so its task cost ends up higher than Astra’s, not lower. Even at the introductory rates, 80% below Astra’s, its task cost would be only 38% lower. Its token use, not its rate card, decides the bill. GPT-6.1 Sol lists $2 / $10, the same as Argon’s introductory rates, scores 55, and costs $1.27 per task. Artificial Analysis lists a single setting, high, so there is no measured cheaper effort level to compare. Check the billing terms before planning around Argon’s price.

Meta’s Muse Contributor option charges $0.10 per million input tokens and $0.20 per million output tokens for supported Spark variants in exchange for permission to use prompts and completions in training. It is not available for Wonder’s API, and Spark 1.3 Contributor does not support max effort. That can change which data you are allowed to send. The $2.39 Muse result above uses Standard pricing; it is not a measured Contributor result.

DeepSeek’s peak, off-peak, and cache rates are another reason to record billing conditions with an evaluation. A discount matters only when your actual requests qualify for it.

MindsHub free tier

The MindsHub free tier includes selected models at no charge. Currently included: the MindsHub Air model and Jev. No credit card is required; fair use and daily rate limits apply.

Measure the cost of results you can use

Start with ten representative tasks as a pilot. Include the documents, instructions, and tool access the model will have in practice, along with a few cases you know are difficult.

1. Set acceptance criteria before running the models. For extraction, that might mean every required field is correct. For a code change, it may mean the tests pass and a reviewer accepts the patch.
2. Run two or three candidates. Record the model version, reasoning setting, prompts, and billing conditions. Repeat tasks where inconsistent results would matter.
3. Count the whole workflow. Include failed attempts, retries, tool charges, and any follow-up model calls. Record human review time separately so a cheap but demanding model does not appear free to operate.
4. Divide total spend by accepted results. If you spend $12 across all attempts and accept eight results, the cost is $1.50 per accepted result. If none pass, the model has not established a usable cost for that workflow.

Ten tasks can reveal a large mismatch; they cannot establish a dependable success rate. Expand the evaluation before a consequential rollout, using the process in our model selection guide.

Re-run it when the model, alias, prompt, or pricing changes. Keep the evaluation set and scoring rules outside any one agent workspace; the workspace migration checklist covers the context and workflow details to preserve.

---

MindsHub Agents provides a workspace with a choice of models and Anton, an open-source agent harness. MindsHub Inference gives developers access to the MindsHub model catalog through one API. Check model availability and MindsHub pricing for the service you plan to use.