Two frontier AI labs cut prices on the same day — about 90 minutes apart, by TechCrunch's count. The same day, Salesforce Ben reported that Salesforce will start charging Flex Credits for the calls AI agents make into your CRM.
Read together, those headlines tell a CFO something a price chart can't: AI cost isn't disappearing. It's moving from the model layer, where prices keep falling, to the platform layer, where new charges and new work are showing up.
On September 22, 2026, Anthropic released Claude Opus 5.5 at $4/$20 per million input/output tokens, 20% below Opus 5, and OpenAI launched GPT-6 Sol and Luna at half their GPT-5.6 promotional API prices. Model prices are falling, but that is only one line of an AI budget. Platform costs are rising: CRM metering of agent calls, agent identity, governance, and data work. CFOs and CIOs at RIAs, banks, and insurers building 2027 budgets should plan for both lines: budget per workflow, not per token, and keep model choice swappable. Vantage Point helps firms build that plan at the platform layer, where the durable cost sits.
Every price below is vendor-reported.
Anthropic released Claude Opus 5.5. According to Anthropic's announcement, input and output tokens now cost $4 and $20 per million, down from $5 and $25 on Opus 5. Cache reads fall to $0.20 per million tokens, 60% less than Opus 5. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and, based on its own tests, costs 40% less than Opus 5 on typical workloads at default settings. The Claude pricing page matches.
OpenAI launched GPT-6 Sol and GPT-6 Luna. Per OpenAI's announcement, Sol now costs $2 input and $10 output per million tokens, down from $4 and $20, and Luna costs $0.10 and $0.50, down from $0.20 and $1.20. OpenAI frames the cut as 50% "compared with their GPT-5.6 promotional pricing." An OpenAI spokesperson told VentureBeat the new rates are permanent, not promotional.
Salesforce moved in the other direction. Per Salesforce Ben and Salesforce's help article on AIforce impact, successful MCP and API calls made by registered AI agents will consume Flex Credits as "Headless Platform Interactions." The multiplier hasn't been published, and Salesforce says it will give 30 days' notice before metering begins. For budgeting, the point is simple: a new platform-layer line item now exists.
This wasn't a one-off. On September 1, Anthropic cut cache-read pricing on Claude Fable 5.1 by 75%, to $0.25 per million tokens.
Both, depending on what you measure. That gap is this week's most useful lesson for anyone who owns an AI budget.
| What you measure | Opus 5 → Opus 5.5 (per Anthropic) | What it tells you |
|---|---|---|
| Output price per million tokens | $25 → $20 (20% lower) | The list-price change on generated text |
| Input price per million tokens | $5 → $4 (20% lower) | The list-price change on prompts and context |
| Cache-read price per million tokens | $0.50 → $0.20 (60% lower) | What reused context costs; Anthropic says cache reads are most of agentic work costs |
| Estimated cost on typical workloads | 40% lower at default settings | Price cut plus fewer tokens per task, based on Anthropic's tests |
A 20% per-token cut and a 40% cost-to-run estimate are different claims. The second depends on your prompts, context size, caching, and agent steps. Anthropic says Opus 5.5 "costs less per token than Opus 5 and uses fewer tokens per task."
The same caution runs across vendors. MacRumors notes that GPT-6 Sol costs half as much per token as Opus 5.5, "but Anthropic says Opus 5.5 uses fewer tokens per task than Opus 5, so it's not a straight comparison." Price sheets compare tokens. Your finance team pays for workflows.
An AI workflow has two cost layers. The model layer is what you pay a lab to reason: tokens in, tokens out. The platform layer is everything that makes that reasoning safe and useful: the CRM the agent reads and writes, the identity it runs under, the governance around it, and the data it depends on.
The model layer is commoditizing; labs now compete on price within hours. The platform layer is going the other way, because that's where systems of record and accountability live.
| Cost line | Layer | Direction | What drives it |
|---|---|---|---|
| Model tokens (input, output, cache) | Model | Falling | Lab competition, efficiency gains, caching |
| CRM platform interactions | Platform | New and rising | Salesforce metering of registered agents' MCP and API calls; multiplier not yet published |
| Platform licensing and editions | Platform | Rising | Agentic features tied to higher editions, as with the Claudeforce premium-edition decision |
| Agent identity and access | Platform | New work | Registering agents, scoped permissions, OAuth credentials, access reviews |
| Governance and oversight | Platform | Recurring | Evaluation, logging, supervision, vendor and model risk review |
| Data readiness | Platform | Front-loaded, then ongoing | Clean records, permissions that hold at agent scale, integration upkeep |
For RIAs, banks, and insurers, governance and identity are not optional. An agent that reads client records needs its own identity, a defensible permission scope, and a record of what it did. Salesforce ties registration and billing together: the step that gives an agent its own identity is what makes its activity "visible, governable, and billable."
Consumption also grows as agents take on more steps, a separate dynamic covered in why AI bills rise as tokens get cheaper. This post is about where cost lands, not how fast it grows.
A multi-year AI business case built on today's per-token rates models the falling line and misses the rising one.
Consider a hypothetical $2B RIA planning three agent workflows for 2027: meeting preparation, account-opening document review, and service-request triage. If the business case is a token forecast times this week's price sheet, it looks better every quarter. Meanwhile, the costs that decide whether those workflows run in production sit outside the forecast: a registered identity for each agent, a possible platform charge on each CRM call, evaluation and supervisory review for each workflow, and client data clean enough to trust.
None of those lines is priced per token, and none falls when a lab cuts prices. In this hypothetical, the model line could shrink every year while total cost per workflow holds flat or rises.
Repricing also runs both ways. OpenAI's cut was measured against GPT-5.6 promotional pricing. VentureBeat reports that Google's Gemini 3.8 Flash introductory rates rise on January 1, 2027. Salesforce's platform charge has no published rate yet. A plan that assumes prices only fall is a single-scenario plan.
Four moves keep a 2027 AI budget honest, whichever model you pick.
If prices can drop 20% to 50% in a day, locking a workflow to one model is a pricing bet. Put an abstraction layer between workflows and model APIs, keep prompts model-agnostic where you can, and write contracts that let you move. Maintain an eval harness: a fixed set of real test cases, scored the same way every time. In a regulated firm, it also documents why a model change was safe.
Platform investments carry over when you change models. Agent identity, permission scoping, logging, and data quality stay useful whether the model underneath is Claude, GPT, or something released next quarter. Our compliance and security solutions and system integration and data migration work focuses on these durable lines.
For each workflow, total the full cost to run it: model tokens, platform interactions, identity and access upkeep, evaluation and review time, and data maintenance. Compare that to the value of the work. Per-token prices belong inside that calculation, not at the top of the slide.
Run at least three scenarios: model prices keep falling, model prices hold, and platform charges arrive at a rate you don't yet know. Don't guess the Salesforce multiplier; show which workflows stay viable across a range, and set a review trigger for when it's published.
Vantage Point is an official Claude partner (Member, Claude Partner Network) that implements Salesforce and HubSpot for regulated firms, and our advice on model cost stays model-neutral. The durable work is the platform layer: agent identity and permissions, governance, clean data, and per-workflow business cases. Through our Claude and AI implementation services, we design swappable architectures and eval harnesses so pricing changes become routine decisions. Across 400+ engagements and 150+ clients, Vantage Point holds a 95% client retention rate and a 4.71/5.0 average engagement rating. Senior consultants only — no junior handoffs; the experts you meet are the experts who deliver.
Model prices will change again before your budget is approved. Vantage Point can map your AI workflows by layer, separate model and platform costs, and build per-workflow business cases with repricing scenarios in both directions. Talk to Vantage Point about your 2027 AI budget.
Anthropic priced Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, with cache reads 60% lower at $0.20 per million. OpenAI priced GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50, which it describes as 50% below GPT-5.6 promotional pricing. All figures are vendor-reported (September 22, 2026).
Anthropic estimates Opus 5.5 costs 40% less than Opus 5 on typical workloads at default settings, based on its own tests. The per-token price cut is 20%; the rest comes from Opus 5.5 using fewer tokens per task. Your saving depends on your own prompts, caching, and agent steps, so test it on your workflows.
Because model tokens are only one layer of AI cost. Platform-layer costs such as CRM metering of agent calls, agent identity, governance, and data readiness are growing, and none of them falls when a lab cuts token prices.
The platform layer is everything around the model that makes an AI workflow safe and useful: the CRM the agent uses, its identity and permissions, logging and oversight, evaluation, and data. For regulated firms, it carries the durable cost because it holds the controls and records that supervisors expect.
Budget per workflow, not per token. Total each workflow's model tokens, platform interactions, identity upkeep, review time, and data maintenance, then run scenarios where model prices fall, hold, and platform charges arrive at an unknown rate. Vantage Point helps firms build these business cases.
Not automatically, but you should be able to. Keep workflows behind an abstraction layer and run every candidate model through the same eval harness before switching. That turns a price cut into a measured, documented decision.
According to Salesforce Ben and Salesforce's help documentation, successful MCP and API calls by registered AI agents will consume Flex Credits as Headless Platform Interactions, tracked in Digital Wallet. The multiplier is unpublished, with 30 days' notice promised before metering begins. Carry it as a platform-layer line with a range, not a guess.
Vantage Point is a boutique CRM consulting firm helping businesses transform with Salesforce, HubSpot, and AI — 150+ clients, 400+ engagements, and a 4.71/5 average engagement rating. Learn more at vantagepoint.io.