By Evan Vega
Three AI models that charge exactly the same price per token can produce bills eight times apart for the same work.
OpenAI’s GPT-6.1 Sol, Moonshot’s Kimi K3.1 Pro and Anthropic’s Claude Sonnet 5.5 all list at $2 per million input tokens
and $10 per million output tokens. On Artificial Analysis’s independent test suite, one finished task cost Sol $1.27,
Kimi $3.68 and Sonnet $10.02, at each model’s top setting. Novel Cognition read every price on the vendors’ own pages and
every score on Artificial Analysis on 29 October 2026, to show what AI agents cost per finished task rather than per
token.
The check also found that the most-shared comparisons are already out of date. The full analysis, with charts and
sources, is at agentcost.novcog.us.com.
GPT-6.1 Sol, Kimi K3.1 Pro and Claude Sonnet 5.5 all list at $2 and $10 per million tokens, yet one finished task on Artificial Analysis’s test suite cost $1.27, $3.68 and $10.02. Novel Cognition read
Google’s pricing page lists Gemini 4 Argon at $2 input and $10 output per million tokens through 30 November 2026,
and $4 and $20 from 1 December; Artificial Analysis already uses the December rate. DeepSeek charges half price outside
its peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which leaves the US working day off-peak.
The same open model also sells at different prices from different hosts. DeepSeek V4.1 Flash 1017 costs $0.30 per
million input tokens at peak on DeepSeek’s own service, Together AI and Fireworks; DeepInfra lists $0.098 as a
promotional 51% discount on a regular $0.20. Fireworks charges 50% more for US-only versions of DeepSeek and GLM-5.3, and
Anthropic charges 1.1 times for US-only inference. The host price
sheet and the other cost considerations are on the
analysis site.
Artificial Analysis runs every major model through the same nine-test suite, covering agentic office work, terminal
coding and hard science questions, and records what each model spent to finish. On its Intelligence Index v5.0, read
29 October 2026, GPT-6.1 Terra scored 51.0 for $0.96 a task and GPT-6.1 Sol 54.5 for $1.27. Claude Opus 5.5 scored
58.4, the highest of the models compared, for $10.04; Claude Sonnet 5.5 scored 54.2 for $10.02. In the cheap tier,
Claude Haiku 5.5 scored 36.2 for $0.41, ahead of DeepSeek V4.1 Flash 1017 (32.5, $0.60) and MiniMax M3.1 Flash (28.9,
$0.49). Haiku’s $0.10 input price applies to prompts up to 100,000 tokens and rises to $0.50 above that.
The full table and score-against-cost chart are on the
analysis site.
The gap comes from how much each model writes. Output tokens cost several times input tokens, and reasoning models
can write pages before answering, so list prices alone do not predict the bill. Artificial Analysis’s figures are its
own test tasks, not quotes for any business’s workload.
Novel Cognition started from an AI-generated comparison dated 27 October whose headline chart divided each model’s
Artificial Analysis score by its per-task cost. Within two days Artificial Analysis had moved to Intelligence Index
v5.0, a new set of nine evaluations, and every score and cost changed. Kimi K3.1 Pro went from 55.1 to 46.0, DeepSeek
V4.1 Flash 1017 from 45.1 to 32.5 and MiniMax M3.1 Flash from 42.7 to 28.9; Claude Opus 5.5 rose from 57.6 to 58.4.
The model-by-model comparison is on the analysis
site.
Two newer models are not in that comparison. OpenAI released GPT-6.1 Terra on 26 October. Google released Gemini 4
Carbon on 28 October, and it now tops v5.0 at 59.4 for $7.76 a task; Artificial Analysis lists it at $4 and $20 per
million tokens, a price not yet shown on Google’s pricing page when Novel Cognition checked.
Novel Cognition’s reading is that agent builders should route routine steps to a cheap model, send judgment calls to
a mid-priced one such as GPT-6.1 Terra or Sol, and keep the most expensive models for the steps where failure costs most,
then measure dollars per accepted, finished task on their own work, retries included. The recommended stack is at
pick a stack, and every source is listed at
sources. The analysis is also an episode of the
Frontier Watch podcast.
This analysis was produced under the Hidden State Drift method, which
checks every claim against its primary source before publication. The ten pre-publication checks behind it are
published at hiddenstatedrift.com/verify. Novel Cognition is not
affiliated with any AI lab or host discussed.
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily Colorado News
