The AI Agent Pricing Ledger: Every Dated Switch From Seats to Tokens to Outcomes

Support AI agents have settled on a price for their work: between 50 cents and two dollars per billed outcome or conversation. Coding agents are moving the other direction, dismantling the flat monthly seat into token meters. The six major products this ledger tracks record five pricing-model switches in the 22 months from August 2024 to June 2026, and every switch splits by vertical: support bills outcomes, code bills inputs. No vendor has switched meters since June 2026, but two additions kept the logic moving: Cognition attached a productivity guarantee worth up to $10 million to Devin in June, and in September a ChatGPT subscription became a way to pay Devin's model bill. The voucher underneath Copilot keeps drifting too: since May 12, GitHub's paid seats carry more credit than they cost, and the $100 Max plan carries twice its price in monthly model credit 1.

Four prices anchor the support side. Intercom Fin charges $0.99 per outcome 2. Zendesk charges $1.50 per automated resolution 3. HubSpot's Breeze Customer Agent dropped to $0.50 per resolved conversation effective April 14, 2026 4. Salesforce's Agentforce Service Agent has been priced from $2.00 per conversation since its October 29, 2024 general availability 56. By ProvenBrief arithmetic, the median price of one billed support interaction now sits near $1.25, and four of the six products in this ledger bill per outcome or conversation: an Outcome-Pricing Share of about 67 percent at first reading.

Four prices, four different definitions of "outcome"

The price list looks like a converged market. The billing triggers underneath it do not agree. Salesforce's original meter is the conversation itself, priced from $2, and since May 15, 2025 it has also sold per-action Flex Credits (20 credits, $0.10 per action) alongside it 56. Zendesk charges only when a ticket is handled with no human intervention and stays closed through what Zendesk typically defines as a 72-hour quiet period 3. Its current guide sharpens the gate: outcomes grade into assisted escalations, contained resolutions and verified resolutions, and only the verified tier bills; the quiet window varies by channel, two hours for messaging by default with an option to extend to 72 hours, and 72 hours after the first email for email 3. Zendesk also extends the model with a pricing structure that combines unlimited seats and a flexible outcome budget spanning conversational outcomes and operational ones, meaning AI tasks completed outside customer conversations 3. Intercom charges once per conversation when the customer confirms the issue is resolved, stops asking for more help, or when Fin completes a workflow 2. HubSpot halved its unit price while tightening its trigger, from $1.00 per conversation to $0.50 per resolved conversation, paid as 50 credits per resolution; its chief customer officer compresses the pitch to "You pay when it works, full stop" 4. A buyer comparing these four quotes is comparing four different meters wearing the same label, and Salesforce has sold a fifth, per-action meter alongside its own since May 2025.

Coding agents turned the seat into a prepaid token voucher

GitHub switched Copilot's billing model twice in under a year. Billing for premium requests began June 18, 2025 across paid plans on GitHub.com 7; on April 27, 2026, GitHub announced that premium request units would be replaced on June 1, 2026 by GitHub AI Credits, consumed per token, input, output and cached alike, at each model's listed API rates, with one credit worth $0.01 89. The seat survives as a voucher, and since May 12, 2026 an overstuffed one: base credits still match each seat price one to one, but a variable flex allotment rides on top. Copilot Pro at $10 a month now includes 1,500 credits worth $15 ($10 base plus $5 flex), Pro+ at $39 includes 7,000 worth $70 ($39 base plus $31 flex), and the new Max plan at $100 includes 20,000 worth $200, with flex amounts free to move as the economics of AI evolve 110. The business side pools rather than stuffs: Copilot Business at $19 per user contributes 1,900 credits ($19) and Enterprise at $39 per user contributes 3,900 ($39) to a shared pool at the billing-entity level, while the Copilot access plan for eligible enterprises includes no allowance at all and is billed from the first use 11. Code completions stay unlimited and unbilled, and through November 30, 2026, paid plans receive a 30 percent discount on model costs when using Copilot Fusion 810. GitHub's stated reason is that agentic usage made the old model unsustainable, since a quick chat question and a multi-hour autonomous coding session could cost the user the same 8.

Devin made the same move in the same month. On April 14, 2026, Cognition announced it was retiring Devin's Core and Team plans for a new lineup: Free, Pro at $20 a month, Max at $200, and a usage-based Teams plan with an $80 monthly minimum replacing the old Team plan's $500 entry point 12. For self-serve plans, overage beyond included quota is now billed in dollars at the underlying model cost instead of the earlier ACU accounting, and enterprise billing continues on ACUs 12. Previously free products, including Ask Devin, DeepWiki and Devin Review, began charging so that pricing reflects the real compute they consume 12. The current documentation holds that structure in place: Free, Pro at $20, Max at $200, Teams at an $80 monthly minimum with full seats at $40 inside it, and Enterprise on ACUs 1314.

Two additions since June, and neither is a switch

The switching stopped on June 1, 2026. The logic did not stop with it. On June 4, 2026, Cognition attached the AI Productivity Guarantee to enterprise Devin deals: an estimator agent reviews each completed session, counts a session as useful output only when the work lands, since a session ending in unmerged pull requests is considered not useful, then converts hours of productive output into dollars at a standard global rate and compares that total against what the customer actually consumed near the end of the annual contract; if the value falls short, Cognition issues credits to cover the gap, up to $10 million 15. The vendor's own compression: if Devin delivers less engineering value than you are paying for, Cognition will fund your usage up to $10M until it does 15. Outcome pricing "shifts part of the performance risk to the seller," in Zendesk's own definition of the model 3, and the guarantee does the same thing on the coding side without changing what the meter reads: quota, then on-demand credits.

The second addition crosses vendor lines. Starting September 29, 2026, a ChatGPT Go, Plus, or Pro subscription can be linked to Devin so the GPT-model share of each session is billed to the ChatGPT plan instead of the Devin quota, which Devin's documentation calls token sharing; it is available on Devin's self-serve Pro, Max and Teams plans, not on Free or Enterprise, and when the linked plan runs out of usage, sessions keep running on Devin quota 1617. GitHub turned its own seat into a voucher for its own tokens in June. Three months later, a seat issued by one vendor pays another vendor's meter. The seat is drifting away from the vendor that printed it.

The dated ledger, every switch with its page

• August 28, 2024: Zendesk announces outcome-based pricing for AI agents, billed per automated resolution 18.
• October 29, 2024: Agentforce reaches general availability with Service Agent pricing from $2.00 per conversation 5.
• May 15, 2025: Salesforce introduces per-action Flex Credits (20 credits, $0.10 per action, $500 per 100,000) alongside $2-per-conversation pricing, an addition, not a switch 6.
• June 18, 2025: Copilot billing for premium requests begins across paid plans on GitHub.com 7.
• April 14, 2026: HubSpot moves Breeze Customer Agent to $0.50 per resolved conversation and Prospecting Agent to $1.00 per recommended lead 4.
• April 14, 2026: Cognition announces the retirement of Devin's Core and Team plans for seat-plus-quota billing with dollar-metered overage 12.
• April 27, 2026: GitHub announces the June 1 replacement of premium requests with per-token AI Credits 8.
• May 12, 2026: Three weeks before the switch lands, GitHub pads the individual lineup: flex allotments lift Pro's included usage from $10 to $15 and Pro+'s from $39 to $70 at unchanged prices, and a new Max plan at $100 a month includes $200 of monthly credits; an amendment to the June 1 switch, not a new meter 1.
• June 1, 2026: AI Credits take effect; premium requests survive only as legacy billing for annual Copilot Pro and Pro+ subscribers 19.
• June 4, 2026: Cognition introduces the AI Productivity Guarantee for enterprise Devin customers, covering shortfalls between delivered engineering value and actual consumption with credits up to $10 million, an addition, not a switch 15.
• September 29, 2026: Devin begins letting customers link a ChatGPT Go, Plus, or Pro subscription so GPT-model usage in Devin sessions is billed to the ChatGPT plan instead of Devin quota, an addition, not a switch 1617.

Two directions, one logic: bill the thing you can verify

The split is not a contradiction. It is accounting. A closed support ticket is a verifiable event, so support vendors can sell the result and define it contractually, down to a 72-hour quiet period 3. Nobody has agreed on what a verified unit of coding work is, and the vendors say so plainly: GitHub repriced because agentic sessions vary enormously in compute 8, and Cognition repriced so pricing reflects real compute 12. So coding bills the input. April 14, 2026 compresses the whole story into one day: the same morning HubSpot tightened its outcome trigger, Cognition bolted a meter onto Devin's seats 412. June and September sharpen the same split from the other side: Cognition now measures useful output, with unmerged pull requests scoring zero, and backs it with credits, but still bills inputs underneath the guarantee 15, while token sharing lets a subscription seat pay for agent work it never bundled 16. May's flex allotments put a hand on the same dial: the seat price is now a floor on bundled compute, not a ceiling 1.

For buyers, the consequence is two budget shapes. Support automation is becoming a cost per ticket that falls as resolution rates rise. Coding agents are becoming utility bills that rise with ambition. The ledger's open question: which coding vendor first dares to bill per merged change.

Every figure above was read from the vendor's own pricing, newsroom or documentation page, and every ledger entry carries the date the vendor stated, whether announcement or effect. Figures verified November 4, 2026.