FOR CFOS · UPDATED OCT 10, 2026
Cut the AI line item, not the output.
Most AI budgets leak in three places: premium tokens spent on routine tasks, stacks of SaaS wrappers, and per-seat licenses. In October 2026, a million output tokens costs $10 on Claude Sonnet 5.5 and $1.20 on DeepSeek V4.1-Flash. Our teardowns show outbound, video and office stacks rebuilt for 80% to 97% less.
By Lucas Page, The ShanghAI Guy · Updated October 10, 2026
Free. One before/after cost teardown every Monday.
LOCALIZED VIDEO ADS
-97%
$1,500 a month on HeyGen vs $34 a month with two Chinese video models, in our teardown.
Source: ShanghAI teardown, August 2026.
THE BEFORE / AFTER LEDGER
What the same work costs on open weights
LINE ITEM
BEFORE
AFTER
SAVING
1M output tokens
$10.00 (Claude Sonnet 5.5)
$1.20 (DeepSeek V4.1-Flash)
-88%
8-tool outbound sales stack
$550+ / month
~$90 / month
-84%
Localized video ads
$1,500 / month
$34 / month
-97%
One coding-agent PR
$5.12
$0.31
-94%
Token prices: official list prices, October 10, 2026. Stack and PR figures: ShanghAI teardowns from August and September 2026, some on older model versions.
BEFORE YOU RENEW ANYTHING
Three questions for every AI line
Is it routine work?
Extraction, classification and drafting rarely need a frontier model. Route them to a model at $1 or less per million output tokens.
Is it a wrapper?
Many $50 to $100 a month tools are simple logic on top of a model. The same workflow runs on a $10 server.
Is it priced per seat?
Seat licenses scale with headcount, not usage. Usage-based open models usually win past a few dozen users.
THE SUITE THAT KILLS SEATS
ByteDance's Doubao Work vs your SaaS stack
We priced ByteDance's Doubao Work line by line against a typical $150 a month SaaS stack, and showed which subscriptions it makes redundant.
Read the teardown, then decide what to cut at your next renewal.
Cost teardowns for the balance sheet
The CFO reading list
The $600 SaaS Tax: replacing an 8-tool outbound stack for $90/mo
ByteDance just liquidated your $150/mo SaaS stack with Doubao Work
HeyGen wants $1,500/mo for localized video ads. Two Chinese models do it for $34.
The 2026 AI workspace tier list nobody in San Francisco wants you to see
Related guides: AI inference cost · Chinese AI models for business · For CTOs
CFO FAQ
What finance leaders ask us
How much can we realistically save on AI?
On routine work, 85% to 95% per token by moving to open-weight models. Whole tool stacks in our teardowns dropped 80% to 97%. Your hardest reasoning tasks should stay on a premium model, so the blended saving depends on your mix.
Where should we start?
List every AI and SaaS line, then flag what is extraction, classification or drafting. That work moves first, usually without any visible change in quality.
What are the hidden costs of self-hosting?
GPUs, storage, monitoring and on-call engineering. Start on a hosted API, then self-host only once volume or data rules justify the infrastructure.
What about compliance and data risk?
Private deployment keeps data inside your perimeter with zero egress. China UnionPay runs DeepSeek-V4 air-gapped and passed PCI DSS. Have legal review each model license before production.
Find the leak before your next renewal.
Free every Monday: one before/after cost teardown for finance leaders.
Want your own numbers? LeanAI audits your token and SaaS spend and shows the saving.
Get a LeanAI cost audit