The Token Explosion
In 2023, a typical LLM completion used 2,000 to 5,000 tokens. A customer asked a question. The model answered. The cost was fractions of a cent.
In 2026, an agentic workflow looks different:
- Planning agent breaks the task into sub-goals. (2,000 tokens)
- Tool-use agent selects and executes API calls. (5,000 tokens)
- Reasoning agent evaluates results and reflects. (8,000 tokens)
- Validation agent checks the output for accuracy. (3,000 tokens)
- Synthesis agent composes the final response. (4,000 tokens)
Total: approximately 22,000 tokens for a task that would have cost 3,000 tokens in 2023. That is a 7x increase. And that is a conservative example — some multi-agent pipelines burn 50,000 to 100,000 tokens per task.
The Pricing Lag
Most AI companies priced their products in 2023 or early 2024, when a "request" meant a single completion. They set flat monthly fees based on the number of requests or users.
Those prices have not changed. The token volume per request has.
The result: companies that ship agentic features are effectively subsidizing their customers' inference bills. Every new agent in the pipeline is a cost increase that the pricing model does not capture.
The Companies Shipping Agents Now
Several companies in our target segment have recently pivoted to or expanded agentic capabilities:
- Basis raised $100 million at a $1.15 billion valuation in February 2026 for full-workflow AI agents in accounting. Full-workflow agents mean exploding tokens per client.
- Nabla pivoted from ambient scribe to agentic workflows in healthcare, raising $70 million in Series C funding. The pivot means rising tokens per clinical encounter.
- Hebbia processes over 1 billion pages with multi-agent reasoning. Their SEC filing names "model dependency" as a risk factor. They run GPT-5, Gemini, and Claude simultaneously.
- Spellbook is on track for approximately $100 million ARR by end of 2025 and raised $40 million in debt for M&A. AI contract drafting with multi-agent review chains is token-intensive.
- Liberate raised $50 million at a $300 million valuation for "reasoning AI agents" in insurance quoting and claims. Reasoning agents are the highest token burn category.
- Dust has 300,000 agents running on company data, backed by a $40 million Series A from Sequoia. Inference is becoming the dominant cost line.
The Margin Math
Here is what happens when a company ships an agentic feature without adjusting pricing:
| Period | Tokens per task | Cost per task | Monthly tasks | Monthly COGS | Revenue | Gross margin |
|---|---|---|---|---|---|---|
| 2024 (single completion) | 3,000 | $0.01 | 500,000 | $5,000 | $50,000 | 90% |
| 2026 (5-agent workflow) | 22,000 | $0.08 | 500,000 | $40,000 | $50,000 | 20% |
Same revenue. Same customer count. Gross margin dropped from 90 percent to 20 percent because the product got smarter and the pricing did not change.
What Needs to Happen
Measure before you ship. Before launching an agentic feature, instrument the token cost. Know the delta between the old completion cost and the new workflow cost.
Attribute by customer. Not every customer uses agentic features equally. Some will trigger the full pipeline. Some will not. Without per-customer attribution, you cannot see who is driving the cost increase.
Price on consumption, not features. If a feature burns 10x more tokens, the customer who uses it 100 times a day should pay more than the customer who uses it once a week. Usage-based pricing is the only model that scales with agentic cost.
Track margin by workflow. Not just by customer — by workflow. The planning agent might be cheap. The reasoning agent might be expensive. You need to know which step in the pipeline drives the cost.
The Board Will Ask
When a board member asks "why did our gross margin drop 15 points this quarter," the answer "we shipped agentic features" is not sufficient. The answer they need is "we shipped agentic features, and here is the per-customer margin impact, and here is our plan to adjust pricing."
The companies that have the attribution data will have that conversation from a position of control. The companies that do not will have it from a position of surprise.
TokenOps tracks per-workflow, per-customer token cost and margin. But the fundamental requirement — knowing what each agentic task costs and who triggered it — is something every AI company must solve before scaling agent-based products.
William Min is the creator of TokenOps and a Technical Product Manager at Lovie. He has 12+ years of experience building payment infrastructure and fintech products. View his LinkedIn profile.
Sources: Multi-agent token consumption estimates based on Anthropic Claude and OpenAI GPT-4o pricing models. Funding data from company press releases and Crunchbase (Basis $100M Feb 2026, Spellbook $40M RBCx debt Mar 2026, Hebbia Series B, Liberate $50M Oct 2025).