Six days after Slack sold none of the agents, Claude became its default model

TL;DR: On 26 August 2026 Salesforce and Anthropic announced Claudeforce. Claude becomes the default model across Slack — Slack AI, Slackbot, Claude Tag, Slack Code — and the reasoning model behind Salesforce’s Atlas Reasoning Engine, powering Agentforce Vibes and Agentforce Coworker by default. A Salesforce in Claude plugin ships with 37 prebuilt sales skills (meeting prep, deal health, pipeline review) that can execute governed actions, with single authentication and no per-user setup. Pilot now, open beta expected September 2026, more skills late 2026. For regulated industries, Claude runs inside the Salesforce Trust Boundary via Amazon Bedrock. No pricing was disclosed for any of it. The context that makes this sharp: six days earlier, Slack Code launched as a deliberately neutral room — four rival coding agents, none sold by Slack. The neutral room now has a house model. For you: find every place a default model is about to be set on your behalf, get the override path in writing, and ask who meters the inference before September.

What was announced

Salesforce and Anthropic published the Claudeforce announcement on 26 August 2026, framed as “the #1 AI meets the #1 AI CRM.” Behind the branding are four distinct changes, and they are not equally significant.

Claude becomes Slack’s default model. Claude is named as the default for Slack AI and Slackbot, and it powers Claude Tag — the channel-resident teammate Anthropic launched in June — as well as Slack Code.

Claude becomes the reasoning model inside Agentforce. Claude is now the model behind Salesforce’s Atlas Reasoning Engine, and powers Agentforce Vibes and Agentforce Coworker by default. It is also available in Agent Builder for teams composing their own agents.

Salesforce in Claude. A plugin putting Salesforce inside the Claude apps, shipping with 37 prebuilt sales skills covering the recognisable parts of a sales week — meeting preparation, deal health review, pipeline review — with the ability to update pipelines and execute what Salesforce calls governed actions. Onboarding is single-authentication with no per-user setup. It is in pilot with selected customers now, with open beta expected in September 2026 and further skills scheduled for late 2026.

A stated governance boundary. For regulated industries, Claude operates within the Salesforce Trust Boundary via Amazon Bedrock, with business rules enforced through Salesforce. Integrations named across the announcement include Headless 360, Data 360, Tableau and Slack.

Marc Benioff’s framing was that the partnership “fuses Claude’s extraordinary reasoning with the trusted data, workflows, and governance every enterprise runs on.” What was not published is any price, for any component.

Six days is the story

On 20 August, Slack launched Slack Code, and the commercially interesting thing about it was what Slack refused to sell. Code channels shipped supporting Claude, Devin, GitHub Copilot and Vercel, with ChatGPT listed as coming. Slack built the room and none of the agents. Every agent in it stayed a separate subscription bought elsewhere. That is an aggregator’s posture, and aggregators earn their position by being credibly indifferent to which supplier wins.

Six days later the same company named a default model across the same product.

These are not contradictory — a default coding agent was never claimed, and the Slack Code roster is untouched. But they point in opposite directions, and the second one is the one with durable commercial consequences. Neutrality at the room level with a house model at the assistant level is a specific and quite deliberate position: keep the multi-agent marketplace where developers notice and care, set the default where ordinary employees do not.

That is not a criticism. It is how defaults have always worked in enterprise software, and Claude at the Opus 5 and Sonnet 5 tier is a defensible choice on the merits — plausibly the same one a careful buyer would make. The point is narrower and worth being clear-eyed about: for most of the people this affects, the choosing is now done elsewhere.

The unpriced parts

An announcement that specifies a default model, an availability window and a compliance architecture while omitting every number is telling you the commercial terms are unsettled. Three questions have real budget consequences and none of them can be answered from what has been published:

Question	Why it matters
Do the 37 skills sit inside current Salesforce licensing, or above it?	Decides whether September is a configuration change or a purchase
Who meters inference when Claude powers Atlas and Agentforce Coworker?	Absorbed into existing Agentforce consumption, or a new Anthropic-metered line
Does Salesforce in Claude need a paid Claude seat per user?	Two seats per person instead of one, across a sales org

Ask all three before the beta. The answers are far more likely to shape your 2027 budget than the capability difference between one frontier model and another.

Where this fits

Read Claudeforce next to OpenAI terminating model supply to Cursor on ownership grounds — announced two days later, in the same week — and a pattern comes into focus that neither story shows on its own.

In one, a tool that advertised every model lost one because its owner changed. In the other, a tool that advertised no particular model acquired a default because its vendor signed a deal. Different mechanisms, same underlying fact: which model reaches your desk is increasingly settled by contracts between vendors, above the level at which you are making choices.

The consequence for buyers is not that this is sinister. It is that the model selector — the interface element that made “we’re multi-model” feel like a guarantee — is a poor proxy for the thing it appears to represent. What actually determines your exposure is the supply agreement underneath it, and you cannot see that from the dropdown.

What to do before September

Inventory the defaults about to be set for you. Slackbot answers, Claude Tag, Atlas reasoning, Agentforce Coworker, Agentforce Vibes. For each, decide whether the default is acceptable or needs an override, and note that no override mechanism has been documented yet.
Get the override path in writing. Whether admins can change the model tenant-wide, whether users can override per channel, and whether any of that is contractual or merely the current configuration. This is the difference between “default” and “exclusive” and it is currently unpublished.
Ask the three pricing questions above before the open beta, not after. Governed actions against your CRM are not a thing to pilot broadly while the commercial terms are still moving.
Test one narrow skill against real work. Meeting preparation is the low-stakes entry point. What you are evaluating is not whether Claude Opus 5 is capable — it is — but whether a governed action fired from a chat surface produces CRM records your revenue operations team will accept.
If you are regulated, demand the region list and a data-flow diagram. “Within the Salesforce Trust Boundary via Amazon Bedrock” is a genuinely meaningful architecture, and the same mechanism that made Grok’s Bedrock arrival matter for data residency. But Bedrock availability is region- and version-specific, and a trust boundary describes where inference runs, not what leaves your CRM to get there.
Re-run your assistant comparison on defaults, not capability. If Claude is arriving by default in Slack while your organisation standardised on something else, that is a live inconsistency worth resolving deliberately — Claude vs ChatGPT and the best AI agent tools guide cover the field, and Claude Code and ChatGPT the individual products.

Honest caveats

Almost everything here comes from one side. The substance is a joint Salesforce and Anthropic announcement and its press coverage. No independent evaluation of the 37 skills, the Atlas integration or the Bedrock boundary exists, and this site has run no tests of its own.

“Default” is doing a lot of work and is not yet defined. Whether it means a changeable setting, a tenant-level policy or a contractual commitment materially changes what this story is worth to you, and the published material does not say.

Nothing here is live for most readers. Salesforce in Claude is in pilot; open beta is expected in September, with further skills flagged for late 2026. Expected dates in enterprise software move.

This announcement is four days old as of publication, and appears here because the buying decision it creates lands in September, not because it broke today.

Frequently asked questions

Does this mean Slack is now Claude-only?

No. Default is not exclusive, and the distinction is the whole practical question. Claude becoming the default model for Slack AI, Slackbot and the Claude Tag experience means it is what answers when nobody chooses, which in enterprise software is the setting that governs the overwhelming majority of usage. Slack Code, launched six days earlier, still lists Claude, Devin, GitHub Copilot and Vercel as taggable agents with OpenAI's ChatGPT flagged as coming, and nothing in the Claudeforce announcement retires that roster. What has not been published is the part that decides how much the default matters: whether administrators can change it tenant-wide, whether individual users can override it per channel or per thread, and whether any of it is contractual or simply the current configuration. Until Salesforce documents the override path, treat 'default' as 'what your organisation will actually use', because that is how defaults behave. If a specific model matters to your workflow for accuracy, cost or compliance reasons, ask your account team for the override mechanism in writing before the September beta rather than discovering it afterwards.

What does Claudeforce actually cost?

Nobody outside the two companies knows, and that is the single most important gap in the announcement. No pricing was disclosed for the Salesforce in Claude plugin, for the 37 prebuilt sales skills, or for how any of it interacts with existing Salesforce, Slack, Agentforce or Claude subscriptions. Three questions follow directly, and all three have material budget consequences. First, do the 37 skills sit inside current Salesforce licensing or above it as an add-on? Second, when Claude is the reasoning model behind the Atlas Reasoning Engine and Agentforce Coworker, who pays for the inference — is it absorbed into the Agentforce consumption model your organisation already meters, or does it introduce a separate Anthropic-metered line? Third, does the Salesforce in Claude plugin require a paid Claude seat for each user who touches it, on top of their Salesforce seat? An announcement that names a default model, an availability window and a governance boundary while omitting price is telling you the commercial terms are still being set. That is normal for a pilot, and it is also the reason not to build a September rollout plan on the assumption this is included in what you already pay.

Should we wait for the September open beta or move now?

Wait, and use the interval to prepare the questions rather than the deployment. Salesforce in Claude is in pilot with selected customers now, with open beta expected in September 2026 and additional prebuilt skills scheduled for late 2026. A beta with unpublished pricing, an unpublished override path and 37 skills that execute governed actions against your CRM is not a thing to roll out broadly on announcement momentum. The productive use of the next few weeks is to inventory where Claude will land by default in your stack — Slackbot answers, Agentforce Coworker, Atlas reasoning, Agentforce Vibes — and decide for each one whether the default is acceptable or needs an override you have not yet confirmed exists. Then pick one narrow, low-stakes skill from the sales set, such as meeting preparation, and evaluate it against whatever your team does today. What you are testing is not whether Claude is good; it demonstrably is at this tier. You are testing whether a governed action executed from a chat surface produces CRM data your revenue operations team will accept, because that is where these integrations usually fail.

Is the Amazon Bedrock detail important?

Yes, and it is the most concretely useful thing in the announcement for regulated buyers. Salesforce says Claude operates within the Salesforce Trust Boundary via Amazon Bedrock for regulated industries, with business rules enforced through Salesforce. That is a specific and checkable architecture rather than a marketing assurance: routing through Bedrock means the inference runs inside an AWS account and region under terms your organisation may already have reviewed, which is a materially different compliance conversation from calling a vendor API directly. It is the same mechanism that made Grok's arrival on Bedrock interesting for data-residency reasons earlier in August. Two caveats keep it from being a finished answer. Bedrock availability is region-specific and model-version-specific, so 'via Bedrock' does not automatically mean 'in your region on the model version you want'. And a trust boundary describes where inference happens, not what leaves your CRM to get there — the prompt assembly, the retrieved records and the tool-call payloads are the data-governance surface, and those are governed by Salesforce configuration rather than by Bedrock. Ask for the region list and the data-flow diagram, not just the boundary claim.

Why does one partnership deserve a buyer's attention when most do not?

Because it changes a default in a product almost nobody chose for its AI, which is a different category of event from a partnership press release. Most vendor tie-ups are distribution deals dressed as product news and change nothing you buy. This one alters what model answers inside Slack — a tool bought years ago for messaging, deployed to entire companies, and now the surface through which a growing share of AI work reaches ordinary employees. The people affected did not evaluate Claude, do not know they are using it, and in most organisations will never see a model selector. That is the real mechanism by which model market share is being decided in 2026: not procurement bake-offs, but defaults inside incumbent software. The same week produced the mirror image of this story, with OpenAI terminating model supply to Cursor on ownership grounds and removing a model from a tool that advertised having all of them. Read the two together and the pattern is clear. Which models you can reach is increasingly decided above your head, by contracts between vendors, and the correct response is to know where every default in your stack sits and whether you can move it.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.