AI tools for business, September 2026: what changed in prices, plans and features

The pattern: the AI line on the bill is now a unit of work

The most useful fact from September is how AI is charged, more than what it can do. On September 3 Salesforce replaced the lineup of its Sales Cloud editions with Core at $195, Advanced at $395 and Max at $550 per user and month, and put Flex Credits into all three: 500,000 a year per organization in Core, 1 million in Advanced and 2.75 million in Max. Salesforce states that existing customers keep their legacy prices. HubSpot's credits article, updated on September 9, sits next to Sales Hub prices of $7 (Starter), $90 (Professional) and $150 (Enterprise) per seat and month at annual billing, with 500, 3,000 and 5,000 credits included, and extra credits at $0.010 each.

On September 14 Workable made its four recruiting agents generally available: job brief, sourcing, screening and engagement. Every plan includes 3,000 AI credits, and extra credits cost $0.12 each in packs of 5,000, $0.10 in packs of 10,000 and $0.095 in packs of 50,000. Screening a candidate takes 1 credit, sourcing 2 and a candidate chat 10. According to its September 10 release the credit model started on July 13, and on that day Workable added a report that shows credit use and cost per hire. The release calls the agents an add-on to its Recruiting plans, so check what your plan includes.

Fin, the customer service agent formerly sold under the Intercom name, charges $0.99 per outcome, with a minimum of 50 outcomes a month, and $9.99 for a qualification. Salesforce completed its acquisition of Fin on September 10 and named no price. Our comparison of AI customer service agents shows how each vendor defines a billable resolution.

Our own data puts this in proportion. Usage-based billing is still a minority: 9 percent of the tools in our directory are billed by usage, against 9 percent per user, and 60 percent have no published model or quote prices only on request. In HR the usage share is lower, at 2 percent. Large vendors announce credit models, while the largest single group still publishes no price. The explorer below filters the tools that have a recorded pricing model by area.

Sources: [1][2][3][4][5][6][7][8][9]

One word, very different prices: what a credit buys

A credit is not a shared unit. At list price a Salesforce Flex Credit costs $0.005 ($500 per 100,000), and the Agentforce price page counts 20 credits, or $0.10, for one agent action. A HubSpot credit costs $0.010, and its rate sheet charges 100 credits, $1.00, for each lead the Prospecting agent recommends outreach for and 50 credits, $0.50, for each text conversation the Customer agent resolves. A Workable credit costs $0.095 to $0.12 and buys one candidate screening or half a sourcing action. The same word stands for very different amounts of work.

Notion and Atlassian use the same idea. Notion sells credits at $10 per 1,000 for Custom Agents, on its Business and Enterprise plans. Atlassian prices extra Rovo usage at $0.01 per credit and includes 25, 70 or 150 credits per user and month in Jira and Confluence Standard, Premium and Enterprise. Neither page carries a date, so we do not count them as September changes. They show that the credit is becoming the default unit.

For a budget this matters: a seat price is a ceiling you can predict, and a credit price is a meter you have to watch. HubSpot's credits reset every month and do not carry over, and its help article describes an automatic move to the next credit tier or pay-as-you-go overage when you run out. Ask each vendor which actions consume credits before you compare per-credit prices.

Sources: [2][10][3][11][4][5][12][13][14]

Some model prices fell, but seat prices did not follow

The model providers behind many of these tools lowered some prices in September. Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. On September 1 it released Fable 5.1 at the unchanged $10 and $50 and cut cache reads by 75 percent, from $1 to $0.25 per million tokens. Sonnet 5.5 followed on September 28 at $2 and $10, the same price as Sonnet 5.

OpenAI released GPT-6 Astra on September 3 at $10 input and $50 output per million tokens for prompts up to 272,000 tokens (higher above that), GPT-6 Sol and Luna on September 22, and GPT-6.1 Sol on September 29. Sol costs $2 and $10, and its changelog lists Luna at $0.10 and $0.50. Reports that OpenAI halved its prices compare against older prices we could not confirm at an OpenAI page, so we do not repeat that claim.

Price per token and price per seat are separate lines on a bill. A cheaper model lowers what a vendor pays, but in our review we found no cut to a seat or credit price at the application vendors covered here. The list changes we found for buyers were new price lists and bundles, not cuts.

Sources: [15][16][17][18][19]

Seat price lists are still reissued, and owners are changing

Not every change moved to credits. Lexware published a new price list valid from September 1, which shows Buchhaltung at 24.40 euros and Lohn+Gehalt at 31.90 euros a month. We found no earlier list to compare against, so we do not state the size of the change. Salesforce's new editions come with the statement that legacy prices stay unchanged, so the new list matters mainly for what you buy or upgrade to next.

Salesforce was also involved in two deals. It led a $166 million investment in HiBob on September 1 and closed its acquisition of Fin on September 10. Whoever owns a tool decides its subprocessors, hosting and the next price list, and our analysis of HR tech acquisitions explains what a change of owner gives customers a right to ask for.

If you are choosing between two CRMs, both price pages now show credits next to seats. Our HubSpot Sales Hub versus Salesforce Sales Cloud comparison works out year-one costs for 10 sellers.

Sources: [20][1][21][8]

Shipped, early access, announced: what you can use today

Feature news in September was full of agents, and the difference between available and announced mattered more than the names. MCP, the protocol that lets an AI assistant call another tool, was the common thread. Workable includes an MCP server at no extra cost on all tiers, Ashby lists one as available, Zendesk made its MCP client generally available, and Freshworks introduced an MCP Gateway in an update whose features are available to select customers under its Early Access program.

Read vendor claims with care. Paychex says WISE Hire is four times faster on average, Workday cites early adopters saving 75 percent of the time it takes to onboard a new benefit provider, and Ashby says 91 percent of its customers use AI features regularly. All three are the vendors' own figures, and none of them is audited.

Sources: [5][22][23][24][25][26][27][28][29][30]

What we could not confirm, and what to check before your next renewal

Several items did not survive a check at a primary page, and we left them out rather than print them. Harvey announced Contract Intelligence on May 21, and Artificial Lawyer reported general availability planned for the third quarter. Harvey's own pages still describe a waitlist. Reports that SAP planned a September go-live of SmartRecruiters inside SuccessFactors have no confirmation from SAP or SmartRecruiters that we could open. OpenAI's website pages blocked our automated fetch, so we confirmed OpenAI prices only in its developer documentation and print no ChatGPT seat prices. A Microsoft 365 Copilot Business promotion circulates in partner blogs with conflicting end dates, and we could not open Microsoft's page. Aggregator posts also named an OpenAI agent product called Dots, for which we found no primary source.

The checklist below turns this month's changes into questions for your next contract talk. How we collect and check tool data is described on our methodology page.

Sources: [30][31][18][8]