Cheaper Tokens Won’t Cut Your AI Bill Or Improve The Quality Of The Output

By Kevin Gregson, CEO and Co-Founder, SalesAssistIQ (SAIQ) | November 2026

Summary: Anthropic’s S-1 confirms that frontier AI is a consumption business. The price of a token is falling fast, but the number of tokens each task consumes is rising, and models still hallucinate, overthink and take shortcuts while the meter runs. Enterprise value sits in the application layer that controls context, verifies output and caps cost. For sales teams, that layer is SAIQ.

What does Anthropic’s S-1 tell enterprise buyers?

Anthropic filed its public S-1 on October 26. Revenue reached $16.3 billion in the first half of 2026, up about 1,300% from a year earlier. Roughly 1,900 customers pay Anthropic more than $1 million a year, and about 250 pay more than $10 million. The company spent $5.6 billion on compute in the second quarter against $11.6 billion in revenue.

Anthropic has built one of the fastest-growing businesses in history, and it discloses more about its models’ behavior than most of its peers. Read the filing as a buyer, though, and one fact stands out. Frontier model revenue grows with consumption. Every lab that bills by the token, Anthropic, OpenAI and Google included, earns more when customers use more tokens.

That means the incentives of the model provider and the enterprise buyer don’t line up on their own. The buyer wants the right answer at the lowest total cost. The meter rewards volume.

Why are AI costs rising when token prices are falling?

The price of a token is collapsing. Andreessen Horowitz measured a 10x annual decline in the cost of equivalent model capability. Epoch AI found a median decline of 50x a year across benchmarks, ranging from 9x to 900x depending on the task. OpenAI’s GPT-6.1 Terra launched on October 26 at $1 per million input tokens and $5 per million output tokens.

Consumption is growing faster. Weekly token volume on OpenRouter, a marketplace that routes traffic across models, rose from 0.5 trillion tokens in January 2025 to 128.4 trillion in September 2026, roughly 250x in 20 months. Reasoning models drive much of it. Artificial Analysis reports that frontier models running at maximum reasoning settings generate between 52,000 and 173,000 output tokens to complete a single task in its Intelligence Index, most of them reasoning tokens the user never sees. Artificial Analysis states the consequence directly: more tokens per task raise the cost per task even when the price per token stays flat.

A cheaper unit price doesn’t produce a cheaper bill when every task uses more units. Enterprises that budgeted for falling AI costs are learning this at renewal.

Do AI models burn tokens on wrong answers?

Yes, and the labs document it themselves. The meter runs whether the output is right or wrong. So is it a design flaw or a business model? You decide.

Some models drift toward approval at the expense of accuracy. In April 2025, OpenAI rolled back a GPT-4o update after the model became noticeably sycophantic. OpenAI explained that a new reward signal built on thumbs-up feedback weakened the checks that had kept flattery in line. A model tuned toward user approval tells people what they want to hear.

Reasoning doesn’t guarantee accuracy. OpenAI’s system card for its o3 and o4-mini models reported hallucination rates of 33% and 48% on its PersonQA benchmark, against 16% for the older o1.

Models overthink. A study of 4,018 software engineering agent runs found that selecting the runs with less overthinking improved performance by nearly 30% and cut compute costs by 43%.

Models take shortcuts. Anthropic’s system cards describe models that special-cased tests to make them pass instead of solving the underlying problem. Anthropic cut that behavior by 67% to 69% between Claude 3.7 and Claude 4. METR, an independent evaluator, reported in June 2025 that reward hacking appears in frontier models from several labs.

None of these failures are free. Every flattering answer, fabricated fact, redundant reasoning loop and gamed test consumes billable tokens. The labs are working to reduce them. The buyer pays for them in the meantime.

What does this mean for a sales organization’s AI spend?

Sales is where these economics bite hardest, because sales work runs on context. A seller preparing for a meeting needs the account’s history, the stakeholders and their positions, open issues and the right solution for the buyer’s problem.

A general agent has no persistent memory of the account. Every session starts by rebuilding context from emails, notes, files and the CRM. As the account file grows, the agent faces a bad choice. It reads too little and misses what matters, or it reads everything and dilutes the answer into generalities while burning tokens on material it has already seen.

General-purpose Skills don’t fix this. A Skill is a set of instructions the agent loads into a session. It improves how the agent approaches a task. It doesn’t give the agent a memory of the account, a verified view of who makes the decision, or a way to know which facts changed since yesterday. The agent still rebuilds context every time, and the quality of the answer still depends on how well each seller frames the question.

Across a sales force, cost scales with the number of sellers, times the number of sessions, times the size of the context each session reloads. Quality varies by seller. Consistency across the team is accidental. Some sellers get a strong brief, others get a confident answer built on a hallucinated fact, and leadership can’t tell which is which.

Where does enterprise AI value sit if the model is a commodity?

It sits in the layer that knows the account. The model is becoming interchangeable. The durable asset is knowledge about your customers, held in a form any model can use, verified before anyone acts on it, and delivered at a predictable cost.

What should buyers ask before their next AI renewal?

• What does a correct answer cost us, counting the answers we had to throw away?
• How many tokens do our sellers burn each week rebuilding context the system has already seen?
• Who owns our account knowledge if we change model providers?
• Can we trace every claim our AI makes to a source?
• Is our AI spend capped per account, or does it float with consumption?

How does SAIQ solve token burn and Skills inadequacy?

SAIQ solves token burn by never rebuilding context, and it solves Skills inadequacy by holding what a Skill can’t: account memory, verified knowledge and a commercial model of the seller’s business. SalesAssistIQ (SAIQ) is an AI deal intelligence platform built as that layer for sales. Four of the five largest global insurance brokerages use it, and it won the 2026 ServiceNow AI Innovation Award for CRM.

Token burn

SAIQ keeps persistent deal memory for every account and uses delta reasoning. It reads each document once, at arrival, and processes only what’s new, so tokens go to new information instead of material the system has already seen. SAIQ pushes roughly 90% of what a seller needs before a meeting without a single query, which removes the long prompting sessions a general agent needs to reach a usable answer. It reads inbound email and documents and writes updates back, so sellers don’t key in data. Engineering out hallucination, fabrication and wasted tokens takes deep engineering and continuous tuning as frontier models change. SAIQ does that work so its customers don’t have to. SAIQ prices per account with no per-seller charge, which caps the cost of intelligence while token consumption swings underneath it.

Skills inadequacy

A Skill tells an agent how to approach a task. SAIQ supplies what the agent needs to know. It assembles the firm’s existing data and adds what the firm doesn’t have: public-record data such as regulatory filings and procurement activity, and proprietary data on how a client’s leaders and employees actually make decisions. Every recommendation runs against a commercial ontology that maps the seller’s specific solutions to the buyer’s specific problems, along with the ideal customer profile, buyer personas, ROI and deal breakers. A governed verification layer traces every claim to its source. Conflicting facts are flagged for a person to reconcile, and gates fail closed when stakeholder information can’t be verified. Insurance specialty skill packs sit on top of that knowledge, so industry expertise is applied to verified account data instead of a blank session.

Quality, efficiency and consistency

The result is quality, efficiency and consistency that general agents and their Skills don’t provide. Every seller on an account works from the same verified knowledge, delivered in the same structure, regardless of how well they prompt. When a seller leaves, the account knowledge stays. SAIQ runs inside Claude as SAIQ for Claude and connects to any CRM through MCP. It writes back to whichever data repository the customer chooses, so the customer can switch AI providers without losing what it knows about its accounts.

The models will keep getting cheaper and more capable, and every enterprise benefits from that. Buy the layer that knows your accounts, and let the labs compete for the tokens underneath it.

Frequently asked questions

What did Anthropic disclose in its S-1? Anthropic reported $16.3 billion in revenue for the first half of 2026, up about 1,300%, with about 1,900 customers paying more than $1 million a year.

Why do AI bills rise as token prices fall? Reasoning models use far more tokens per task. Artificial Analysis reports 52,000 to 173,000 output tokens per benchmark task at maximum reasoning, so cost per task rises even when the price per token falls.

Do newer AI models hallucinate less? Not always. OpenAI reported hallucination rates of 33% for o3 and 48% for o4-mini on PersonQA, against 16% for the older o1.

Where does enterprise AI value sit? In the application layer that holds persistent customer knowledge, verifies model output and prices predictably.

What is SAIQ? SalesAssistIQ (SAIQ) is an AI deal intelligence platform used by four of the five largest global insurance brokerages. It provides persistent account memory, verified intelligence pushed to sellers and per-account pricing, and it runs inside Claude and any CRM.

Sources

• Fortune, Anthropic’s S-1 (October 26, 2026)
• IPOScoop, Anthropic files its S-1 (October 2026)
• Andreessen Horowitz, Welcome to LLMflation (November 2024)
• Epoch AI, LLM inference price trends (March 2025)
• OpenRouter, The Token State of Play (September 2026)
• Artificial Analysis, Benchmarking Methodology and Model Comparisons (October 2026)
• OpenAI, Expanding on what we missed with sycophancy (May 2025)
• OpenAI, o3 and o4-mini system card, hallucination evaluations (April 2025)
• Overthinking in agentic tasks, arXiv 2502.08235 (February 2025)
• Anthropic, Claude 4 System Card (May 2025)
• METR, Recent Frontier Models Are Reward Hacking (June 2025)