Salesforce built its own AI. That's a pricing signal.

Two threads run through this week's pricing news. Salesforce spent Dreamforce showing that it will no longer let Anthropic or OpenAI price its inference layer — the Koa reveal is the biggest single margin story in enterprise SaaS this year. And across the rest of the market, vendors split cleanly into two opposite responses to the same AI margin squeeze: retreat upmarket, or barbell around a $20 middle. Both stories rhyme. Both change how you should think about your 2027 renewals.

Salesforce trained its own AI. That's a pricing signal.

On September 15 at Dreamforce, Salesforce announced Koa — its first proprietary CRM reasoning model, built by post-training NVIDIA Nemotron 3 Super on synthetic data modeled from 27 years of Salesforce CRM deployments. Salesforce controls the weights. It runs inference inside its own trust boundary. GA is winter 2026 in US regions. Pilot customers include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.

Koa is the third foundation-model move Salesforce has made in thirty days. On August 26, Claudeforce committed $5B to Anthropic. On September 15, Missionforce brought NVIDIA and OpenAI into the mix for government and air-gapped deployments. That same day, Salesforce unified its stack on Google Cloud and expanded the AWS collaboration. Koa closes the loop: Salesforce is now committed to Anthropic, OpenAI, Google, AWS and its own model. That is not a bet. That is a model portfolio.

The pricing lens is where this gets interesting. Koa is a hedge against Anthropic's and OpenAI's API pricing power. When Claude Sonnet doubles next year — and it will — Salesforce shifts inference to Koa for the workloads that fit and captures the delta. Frontier labs' enterprise ARR growth slows because their biggest customers now have alternatives that they own.

For Salesforce customers, the implication runs the other way. The Flex Credit rate ($500/100K, roughly $0.005/credit) will not drop just because Salesforce's underlying inference cost falls. Salesforce captures the margin, and Agentforce gross margin expands over the next two to three quarters. If you are renewing any Agentforce tier — Core ($195), Advanced ($395), Max ($550) — negotiate on volume commit, not credit rate. The credit rate is the vendor's margin lever, not yours.

The most telling detail is Xero on the pilot customer list. Xero competes directly with Sales Cloud in the SMB CRM overlap. Xero using Koa signals Salesforce positioning as the CRM AI infrastructure — even for its own competitors. It is the AWS playbook translated to the AI stack: sell picks and shovels to the people mining the same gold you are.

The "own the weights" playbook is coming to every enterprise SaaS

Salesforce is the loudest example. It will not be the only one. Any enterprise SaaS vendor with the customer data and the ARR to justify training a specialized model is going to do it, and every one that does will run the same margin play.

The list of vendors positioned to follow through in the next 12 months is short and predictable. HubSpot has the data and has publicly signaled AI ownership intent. ServiceNow already runs Now Assist on top of frontier models — the next step is post-training its own. Workday announced the OpenAI partnership recently; expect an owned-model equivalent to follow. Atlassian's Rovo is deliberately model-agnostic, which reads as scaffolding for a proprietary layer. Rippling and Gusto sit on vertical HR data that is one of the highest-value training corpuses in the SaaS market.

When enterprise SaaS vendors own their inference layer, three things happen. Frontier lab API growth slows because their biggest enterprise customers stop escalating usage. SaaS vendor gross margins expand — Salesforce is the first case study of many. And customer pricing does not drop, because the savings accrue to the vendor, not the buyer. This is not a technology story. It is a margin-expansion story. And every enterprise SaaS renewal in 2027 will happen against this backdrop.

The rest of the market is splitting into two opposite pricing responses

The story below the enterprise line is the mirror image. We tracked fifteen structural pricing changes across the directory in the last two weeks. They sorted into two clean camps.

Retreat playbook — kill the free tier, raise the entry, quietly move upmarket. Tabnine retired all individual and free plans in April 2025; the new floor is $39/user/mo Code Assistant, annual-only. Sourcegraph Cody discontinued Free, Pro, and Enterprise Starter on July 23, 2025 — the product is now enterprise-only at $59/user/mo with a $16K team minimum. Zoom's acquisition of Common Room closed with a repriced ladder: Essential jumped from $625 to $2,500/mo, annual-only, no free plan. Krisp swapped its free-forever plan for a 7-day trial. Consensus doubled its Pro entry from $10 to $20 and added a $65 Deep tier above it.

Barbell playbook — add a free tier, crush the entry to ~$20, layer a $200+ premium on top, collapse the middle. Cognition retired Devin's $500/seat Team plan; the new ladder is Free / $20 Pro / $200 Max. Windsurf (formerly Codeium) restructured on March 19 with Free / $20 Pro / $200 Max, Pro moving $15 to $20 to match Cursor exactly. Cursor Pro has held this shape for a year and every AI coding tool has converged on it.

Both playbooks are responses to the same problem: seat-based pricing does not work when AI usage varies 100x per user. The Retreat vendors give up the variable-use middle to protect enterprise margin. The Barbell vendors capture free users cheaply and monetize the top-1% at $200. Neither is stable at the middle. Full analysis at The 2026 SaaS pricing divide: retreat or barbell?.

Also this week

• Cursor acquired Continue.dev around June 16 in a quiet acqui-hire; the paid hub was retired, cloud data deleted July 15, continue.dev/pricing now returns 404. Open-source repo lives on under Apache 2.0 but read-only. Salesforce's absorption of Momentum.io in March 2026 also surfaced in this week's audit trail — the AI-native SaaS consolidation wave is now a countable pattern (Cursor / Continue, Zoom / Common Room, Salesforce / Momentum, Salesforce / Airtable, Cognition / Windsurf, Adobe / Semrush — six deals in six months).
• Salesforce edition restructure — the September 3 rebase to Core $195, Advanced $395, Max $550 is now the standing baseline on new customer contracts. Existing Enterprise/Unlimited holdouts are still on legacy pricing; expect migration nudges through Q4.
• SPH directory milestones — the tracker crossed 220 tracked SaaS tools plus 91 AI-native additions this week (Cursor, ElevenLabs, Vapi, Otter, Harvey, Runway and dozens more). The daily refresh cadence stepped up to 30 tools/day; the full directory now recycles every 7-8 days. Every /tools/[slug] page carries a QAPage JSON-LD block for AI Overview citation eligibility.
• Anthropic Claude Sonnet 5 — the scheduled September 1 introductory-to-standard rate step (from $2/$10 to $3/$15 per million input/output tokens) still has not shipped. Introductory pricing is now the de facto standard.

Catch you next week.

— SaaS Price Hub