The meter moved from seats toward work

On July 6, 2026, OpenAI changed how ChatGPT Workspace Agent runs consume credits. The new rate is based on input tokens, cached input tokens, and output tokens. ChatGPT for Excel and, after its free period, PowerPoint use the same general agentic usage pool. A seat still gets someone through the door. The work now has a meter.

OpenAI's current rate card says a typical end-to-end Workspace Agent run using GPT-5.5 may consume between 5 and 25 credits. It lists 125 credits per million input tokens, 12.5 per million cached input tokens, and 750 per million output tokens. The same page gives a worked example in which 20,000 input tokens, 80,000 cached input tokens, and 5,000 output tokens total 7.25 credits.

This is useful transparency. It is not a unit of production. One five-credit run might produce a payment reconciliation that clears review. Another might produce a confident spreadsheet-shaped delay. The invoice sees tokens. The customer sees whether the job is done.

Agent architecture is now a margin decision

OpenAI's July 9, 2026 GPT-5.6 launch makes the new cost surface unusually explicit. The API prices Sol at $5 per million input tokens and $30 per million output tokens, Terra at $2.50 and $15, and Luna at $1 and $6. More important than the list prices is the claim that tool-heavy work can be reorganized to consume less of them.

The launch post says Programmatic Tool Calling lets a model filter intermediate data and coordinate tools without passing every result back through another model round trip. OpenAI quotes Clio reporting 38% fewer prompt tokens on multi-step document analysis with no quality loss. It quotes PlayCo reporting 63.5% fewer total tokens and 50.1% fewer model turns on scene-construction workflows while producing comparable visual results. These are customer-reported evaluations in a vendor launch post, not audited margin statements. They are still commercially useful clues.

The clue is that orchestration belongs in the P&L. Context filtering, caching, model routing, tool design, retry policy, and early stopping are not merely engineering neatness. At volume, they can reduce variable inference cost and may improve margin where inference is material. The prompt is not the cost center. The whole loop is.

A token is not a stable unit of work

Anthropic supplied a tidy warning on June 30, 2026. Claude Sonnet 5 launched at an introductory $2 per million input tokens and $10 per million output tokens through August 31, after which the announced rates become $3 and $15. A footnote says the model's updated tokenizer can map the same input to roughly 1.0 to 1.35 times as many tokens as Sonnet 4.6, depending on the content. Anthropic says the introductory rate was set to make that transition roughly cost-neutral.

In other words, even the token is vendor- and model-shaped. A lower token count can reflect a tokenizer, shorter output, better tool plumbing, or an agent that stopped before finishing. A higher count can reflect waste, but it can also reflect harder work. Anthropic's June 26 Economic Index found that conversations about building apps used more than three times the tokens of the median conversation. Across artifact types in chat and Cowork, average autonomy and median token use rose together with a correlation of 0.68.

The report does not establish business outcomes. It shows that token use tends to rise with task complexity and delegated autonomy, which is why a cheaper inference bill should not automatically be treated as a better business outcome. A weak model that needs three retries and a senior reviewer can be more expensive than a capable model with an alarming price per million and a clean first pass. Procurement tables rarely include the second half of that sentence.

Observability is arriving before accounting

OpenClaw's usage-tracking documentation, as accessed July 10, shows how quickly the operational layer is catching up. It can display provider-reported quota windows, balances, spend, budgets, daily cost history, and token or model attribution. Its Control UI can add provider-reported today, seven-day, and 30-day organization spend when the appropriate Anthropic or OpenAI admin credential is available. Separately, it derives transcript-based token totals and estimated costs when usage metadata and local pricing are available.

The docs make a careful distinction: OpenClaw does not merge provider-reported billing totals with local session estimates because they answer different questions. Sensible. Provider admin APIs report actual organization billing. OpenClaw's local view instead applies configured or catalog-resolved pricing to usage metadata recorded in available session transcripts; it is an estimate of that local history, not a provider invoice or lifetime billing ledger.

Neither says whether the customer accepted the output. Cost observability without an outcome ledger is a fuel gauge bolted to a vehicle with no odometer. Better than guessing. Not enough to run the route.

Build a job ledger, not a token dashboard

Choose one production unit the customer recognizes: reconciled account, qualified lead, patched vulnerability, completed renewal packet, resolved ticket, approved listing, shipped code change. Define acceptance before the agent runs. Then calculate fully loaded cost per accepted unit: model and cache charges, paid tools, runtime, retries, human review time, exception handling, and the cost of failures that escape downstream.

Track at least first-pass acceptance rate, average and 95th-percentile cost, reviewer minutes, retry count, exception rate, cycle time, and customer-visible defects. Segment the ledger by workflow version and model route. A blended monthly token bill is pleasant for finance and useless for diagnosis.

For the first 30 days, run two or three routes against a stratified replay set of the same representative jobs and the same acceptance test. Compare acceptance, review time, and cost within difficulty bands before changing live traffic. Then pilot conditional routing: a smaller model for routine cases, a stronger model for ambiguous ones, and a deterministic rule for any stable step, which will upset nobody except the mascot. Route work to the cheapest path that clears the quality and latency threshold, then review the exceptions that consumed the margin.

The operator takeaway

Founders should price the customer's finished job, not resell tokens with a tasteful markup. Keep enough margin to pay for model volatility, review, support, and the cases that refuse to resemble the demo. When efficiency improves, some of that gain can widen margin and some can improve the offer. Handing all of it to the customer immediately is a fine way to become a billing wrapper with opinions.

Buyers should ask for the last 90 days of accepted volume, first-pass yield, human review minutes, total model and tool spend, 95th-percentile cost per accepted unit, and the five most expensive failure modes. If a seller can only show seats, sessions, tokens, or hours saved, the denominator is still missing.

Operators should make this ledger part of release review. A new model is better only if it improves the frontier that matters for the workflow: accepted quality, total cost, cycle time, or risk. Benchmark wins are interesting. A gross-margin improvement that survives customer review is a business fact.

Token metering is good. It makes the machinery less mystical.

A token counter tells you what the model consumed. Cost per accepted job tells you whether the company can survive its own success.

sources
  1. ChatGPT Business - Release NotesOpenAI, accessed July 10, 2026
  2. ChatGPT Rate Card (Business, Enterprise/Edu)OpenAI, accessed July 10, 2026
  3. GPT-5.6: Frontier intelligence that scales with your ambitionOpenAI, accessed July 10, 2026
  4. Introducing Claude Sonnet 5Anthropic, accessed July 10, 2026
  5. Anthropic Economic Index report: CadencesAnthropic, accessed July 10, 2026
  6. Usage trackingOpenClaw, accessed July 10, 2026