For buyers watching AI reshape software.
The workbench is the product: three ways frontier labs are buying verticals
Week 27 of 2026 · July 4, 2026
Big read
On June 30 Anthropic shipped Claude Science — and pointedly did not ship a new model. It is an AI workbench for scientists: the tools, packages, databases, and compute researchers already use, integrated around Opus 4.8 and producing auditable artifacts, in beta for every paid tier with discounted plans for academic and nonprofit labs and up to 50 funded 'AI for Science' projects. TechCrunch's framing is the week's most useful lens: three frontier labs are now attacking the same scientific market with three different architectures, making this the cleanest natural experiment yet in how labs buy verticals.
The three plays: Anthropic goes wide — a workflow wrapper on an unchanged model, distributed through subscriptions anyone can turn on. OpenAI goes narrow — GPT-Rosalind is a gated specialist model behind enterprise trusted-access, and per The New Stack the company has disbanded OpenAI for Science as a broad effort. Google DeepMind bundles owned proprietary models — AlphaFold and Gemini for Science — into its own surface. The buyers are refusing to choose: Novo Nordisk and the Allen Institute appear on both Anthropic's customer list and OpenAI's early-access list, confirming that multi-vendor is the pharma default from day one.
This is the editorial thesis playing out in public. The model is the runtime; whoever owns the workflow, the artifacts, and the audit trail owns the application. Claude Sonnet 5, launched the same day, makes the point from below: agentic capability that recently required Opus-class models is now the baseline at mid-tier prices ($2/$10 per million tokens introductory, $3/$15 after August 31), which means raw capability is commoditizing and the durable margin migrates up to the workbench. Snowflake made the point from the side, shipping Sonnet 5 same-day inside its Cortex AI perimeter as a launch partner — the governed data layer is becoming day-zero model distribution, a real procurement alternative to direct API contracts.
The runner-up story is what happens after the workbench wins: pricing follows the work. Salesforce's Agentforce Help Agent reaches GA in July with pay-per-resolution — charged only when the agent resolves an issue end-to-end autonomously, no charge on human escalation or negative feedback, with Salesforce absorbing token-cost risk on failures. Combined with Microsoft's Service Agent GA (an action-taking agent with 70+ MCP tools inside licensing enterprises already own) and Salesforce's pending ~$3.6B Fin acquisition, support has become the first vertical where seat pricing visibly dies — because resolution is the rare outcome vendors can actually measure.
What to do with this week: CIOs should treat the science fight as the template — the same three plays (workflow wrapper, gated specialist model, proprietary-model bundle) will replay in law, finance, and engineering, and the evaluation question is who owns workflow state, artifacts, and audit, not whose model benchmarks best. Vertical-function heads negotiating support renewals should demand resolution-rate telemetry and a contractual definition of 'resolved' before outcome-priced SKUs land. SaaS investors should discount model-adjacent capability claims and price workbench ownership. And vertical-AI founders should note the uncomfortable part: when the lab decides your vertical is next, its distribution move is a subscription toggle, not a sales cycle.
Vertical movements
2026-06-30 · Anthropic · Research · Frontier Lab · Seat Based
Claude Science
R&D and IT leaders should read this as the template frontier labs will apply to law, finance, and engineering next: own the workflow layer of a vertical, not just the model, and distribute it through subscriptions already in place. Pharma is hedging — Novo Nordisk and the Allen Institute appear on both Anthropic's customer list and OpenAI's GPT-Rosalind early-access list — so plan for multi-vendor science stacks, not a single winner.
2026-06-30 · Anthropic · Other · Frontier Lab · Usage Based
Claude Sonnet 5
Buyers should benchmark cost-per-completed-task, not per-token rates: agentic capability that recently required Opus-class models is now baseline at $2/$10 per million tokens (introductory through Aug 31, then $3/$15 vs Opus 4.8 at $5/$25). A single-sourced Artificial Analysis read warns the new tokenizer raises token counts up to ~35% and max-effort runs burn ~40% more output tokens per task than Sonnet 4.6, so rate cards understate true cost.
Sources Anthropic; TechCrunch
2026-06-30 · Microsoft · Support · Incumbent Saas
Service Agent in Microsoft 365 Copilot
Support and IT leaders should re-baseline every support-agent RFP: an action-taking agent is now GA inside the M365/Dynamics licensing most enterprises already hold, spanning Dynamics 365 Customer Service and M365 apps. The architecture signal is that the agent's value is the governed MCP tool catalog, not the chat surface — evaluate the tool catalog and its identity model, not the demo.
Sources Microsoft Dynamics 365 blog
2026-06-25 · Salesforce · Support · Incumbent Saas · Outcome Based
Agentforce Help Agent
Function heads should expect resolution-rate-based procurement to become the support-vertical norm and start scrutinizing how 'resolved' is measured — reopened cases and coerced deflection are the failure modes to contract against. Salesforce is charged only on autonomous end-to-end resolution, absorbing token-cost risk on failures, while its pending ~$3.6B Fin acquisition brings a purpose-built support model claiming ~76% average end-to-end resolution.
Sources SiliconANGLE; CX Today
2026-06-29 · Cursor (Anysphere) · Engineering · Startup
Cursor for iOS
Engineering leaders should treat mobile agent supervision as a standard procurement expectation for agentic dev tools — this follows Codex-in-ChatGPT-mobile and Claude mobile-to-desktop control, and the agent fleet is becoming a service you supervise from anywhere. Vendor-risk note: Cursor's independence ends when the announced ~$60B SpaceX acquisition closes (expected Q3 2026).
Sources Sources (Alex Heath)
2026-07-01 · Google · Engineering · Frontier Lab
Remote MCP server for Gemini Enterprise Agent Platform
CIOs should standardize on MCP as the integration contract: Google is running the reverse of the ServiceNow Build Agent pattern — instead of forcing output into a governed runtime, it exposes its governed runtime (with native IT controls) to any external agent, including rival tools like Claude Code. Either direction, the durable boundary is the governed tool catalog.
Sources Google Cloud release notes
2026-07 · Harvey / Legora · Legal · Startup
Firm-wide legal AI deployments (Harvey, Legora)
Legal function heads should note the reference deployment model outside the US/UK is now firm-wide and single-vendor, not pilot-broad — expansion is geographic (Korea, France) and workflow-deep. EvenUp's customer results (2,500+ cases managed, medical-record review from weeks to seconds at Michael Kelly Injury Lawyers) show the same depth shift in personal injury.
Sources Aju Press; Legora newsroom
Incumbent responses
2026-06-30 · Microsoft
Service Agent + MCP tools in Microsoft 365 Copilot
The bundling threat to standalone support-agent vendors (Sierra, Decagon, Fin) is now GA, not roadmap. CIOs holding M365 licenses should force every standalone support-agent proposal to price against a capability they may already own.
2026-06-25 · Salesforce
Agentforce Help Agent + pay-per-resolution
Seat-license defenders now compete against their own outcome-priced SKUs. Investors should watch the FY27 Q2 print in late August for the Agentforce ARR trajectory (Q1: $1.2B ARR, +205% YoY per Salesforce); buyers should demand the resolution telemetry that makes this pricing auditable.
Sources CX Today; Salesforce Ben
2026-06-29 · ServiceNow + Accenture
AI-powered migration services off legacy risk platforms
Incumbents are weaponizing agents for share-shift — using AI to collapse the switching cost that protects competitors' installed bases. CIOs should expect 'agent-assisted migration' pitches across every legacy category next, and recognize the same lever can be pointed at their own incumbent vendors.
Sources Business Wire via Morningstar
2026-06-30 · Snowflake
Claude Sonnet 5 on Cortex AI (same-day private preview)
Data clouds are becoming day-zero distribution for frontier models. Procurement teams should treat 'run the newest model where the governed data already lives' as a real alternative to direct API contracts — it changes who holds the security review and the spend relationship.
Sources Snowflake blog
2026-06-30 · HubSpot
Warmly acquisition (AI sales agents + intent data)
Mid-market incumbents are buying intent-data-plus-agent combos rather than building, which shifts sales-agent differentiation to proprietary signal data. Founders selling standalone sales agents without owned data should expect their category to consolidate; note this is single-sourced trade coverage with terms undisclosed.
Sources The Enterprise News
2026-06-22 · Adobe
Cannes agentic partnerships + CX skills/MCP servers GA (pre-window context)
Marketing heads should note Adobe is distributing its capabilities inside rival AI surfaces via MCP rather than defending a walled garden, and agency networks are now operational deployment channels for agentic marketing. Pre-window context (Jun 22), carried because it firms Adobe's marketing-scorecard position this week.
Sources Adobe newsroom
Startup signals
2026-07-01 · Other · $800M at $8.3B post-money
Together AI
The open-weight serving substrate behind several application-layer leaders (customers include Cursor, Cognition, Decagon, ElevenLabs, Suno; annual bookings crossed $1.15B last quarter per the company) just got capitalized at neocloud scale. Agent vendors squeezed by closed-API margins now have a funded escape valve — factor that into build-vs-buy on model serving.
Sources TechCrunch; Together AI blog
2026-07-01 · Other · $100M Series B
TwelveLabs
Vertical-by-data-type is a fundable thesis: video archives as the system of record, expanding from video-understanding models to a full-stack agentic system. Buyers and founders should also note the distribution mechanics — a hyperscaler using a minority check to lock preferred-cloud status. Cumulative-raise totals differ across outlets; the $100M round itself is consistent.
Sources GlobeNewswire PR; SiliconANGLE
2026-07-01 · Legal
Harvey
Customer-win signal, not funding: legal AI's expansion is now geographic and firm-wide rather than pilot-broad. Legal-tech founders and investors should treat single-vendor firm-wide commitments as the reference deployment model forming outside the US/UK.
Sources Aju Press
2026-07-01 · Finance · €6.8M Series A
Wultra
A minor but directional signal for finance-vertical buyers: as AI-driven identity attacks scale, bank-grade trust infrastructure becomes part of the application-layer bill of materials. Expansion targets are the Middle East and US.
Sources Tech Startups
Pricing shifts
2026-06-25 · Hybrid → Outcome Based
Salesforce
Agentforce Help Agent moves from seat-plus-consumption (Flex credits) to pay-per-resolution: charged only when the agent resolves an issue end-to-end autonomously, with no charge on human escalation or negative feedback and Data 360/Agentforce conversation usage unmetered. Salesforce Ben notes the model shifts token-cost risk onto Salesforce and creates adversarial-use exposure (users claiming dissatisfaction to burn vendor tokens).
2026-06-30 · Usage Based → Usage Based
Anthropic
Sonnet 5's headline rate card holds flat ($2/$10 per million tokens introductory through Aug 31, then $3/$15, identical to Sonnet 4.6) but the effective cost per completed task can rise: a new tokenizer adds up to ~35% token counts and max effort burns ~40% more output tokens per task (single-sourced via Artificial Analysis). Buyer implication: benchmark cost-per-completed-task, not per-token rates.
Sources Anthropic; TNW; Medium cost analysis (single-sourced percentages)
2026-06-10 · Unknown → Outcome Based
Adobe
Pre-window context: CX Enterprise Coworker is priced as 'low-barrier entry that scales based on value realized' — an outcomes-flavored model on the marketing side, carried here because it extends the support-first outcome-pricing pattern into a second vertical.
Sources CMSWire; Adobe PR
Vertical scorecard
As of 2026-07-04
| Vertical | Leader | Challenger | Read |
|---|---|---|---|
| Support | ServiceNow | Microsoft Service Agent | Escalating three-way fight: Microsoft's GA inside M365 licensing replaces the Fin lineage (being absorbed into Salesforce, pending close) as the clearest challenger; Salesforce counters with outcome pricing. |
| Research | Anthropic (Claude Science) | Google DeepMind / OpenAI | New vertical to track; leader unclear, but Anthropic has the distribution move this week — broad-access workbench vs Google's owned models (Gemini for Science) vs OpenAI's gated GPT-Rosalind. |
| Engineering | OpenAI Codex | Cursor | No leader change, but the supervision surface moved to mobile (Cursor iOS, following Codex-in-ChatGPT and Claude mobile); Cursor's independence ends at the SpaceX close expected Q3. |
| Legal | Harvey | Legora | Harvey and Legora consolidate international expansion (Korea, France) with firm-wide deployments; the 'Big Five' framing (Thomson Reuters, Lexis, Harvey, Legora, Clio) holds. |
| Marketing | Adobe | Salesforce | Adobe's claim firmed up from W26's co-leader listing via agency-network deployment at Cannes and MCP distribution into Claude Enterprise and M365 Copilot. |
| Sales | Salesforce Agentforce | HubSpot + Warmly | HubSpot's Warmly acquisition signals intent-data-plus-agent bundling in mid-market CRM; differentiation is shifting to proprietary buyer-signal data. |
| Operations | ServiceNow | Cursor / OpenAI automation stacks | Carried from W26: operations value keeps shifting to trigger governance, approval queues, and auditability for background agents. |
| Finance | Microsoft Dynamics | Agentic procurement startups | Carried from W26: finance and procurement agents need closed-loop controls and cost metering before broad rollout. |
| Security | Formal verification / assurance tools | Agent logs and policy wrappers | Position held, but the ServiceNow-Accenture agent-run risk-management services show agents entering the GRC stack itself — as migration weapon and managed service. |
| Other | Vertical workflow owners | Generic copilots | Carried from W26 and reinforced by Claude Science: the application-layer winner owns workflow state, artifacts, permissions, and measurement. |
Architecture watch
MCP becomes the procurement boundary
Vendors now compete on the quality and governance of their MCP tool catalogs, and capabilities travel to whatever agent surface the customer already uses — Microsoft ships its support agent on a 70+ tool MCP server, Google exposes its governed runtime to external agents over MCP, Adobe distributes into rival surfaces the same way. The CIO question shifts from 'which app' to 'whose tools, under whose identity and audit model, inside which surface.'
- Examples
- Microsoft Service Agent (70+ MCP tools), Google remote MCP server for Gemini Enterprise Agent Platform, Adobe CX skills/MCP servers in Claude Enterprise and M365 Copilot, Workday Agent-Ready Tools (Jun 2, context)
Sources Microsoft Dynamics 365 blog; Google Cloud release notes; Adobe newsroom
Outcome pricing crosses from thesis to GA
W26 covered the BVP pricing thesis; this week it shipped — Salesforce's pay-per-resolution reaches GA and Adobe prices Coworker to scale with value realized. The counterweight is Sonnet 5's unit-economics tension: agentic models do more work per task, so vendors absorbing outcome risk must control token burn. Expect outcome-priced SKUs to appear selectively where vendors trust their resolution telemetry — support first.
- Examples
- Salesforce pay-per-resolution (GA July), Adobe CX Enterprise Coworker value-scaled pricing
Sources CX Today; CMSWire; Anthropic
The vertical workbench: workflow layer over an unchanged model
Claude Science is explicitly 'not a new model' (TechCrunch) — it is packaging: tool integration, auditable artifacts, compute access, and ecosystem partnerships on top of Opus 4.8. The same-week contrast with OpenAI's gated specialist and Google's owned-model bundle gives buyers three named architectures for vertical capture that will replay in law, finance, and engineering.
- Examples
- Claude Science (workflow wrapper on Opus 4.8), GPT-Rosalind (gated specialist model), Gemini for Science / AlphaFold (proprietary-model bundle)
Sources TechCrunch; Anthropic
Data clouds as day-zero model distribution
Snowflake shipped Sonnet 5 same-day inside its security perimeter as an Anthropic launch partner, while Together AI's raise (with Salesforce Ventures and SE Ventures participating) capitalizes the open-weight alternative serving Cursor, Cognition, and Decagon. Model access is becoming a feature of the governed data layer rather than a separate contract — the 'data cloud is the platform' thesis strengthens.
- Examples
- Snowflake Cortex AI (same-day Sonnet 5 private preview), Together AI $800M raise
Sources Snowflake blog; TechCrunch
Watchlist
Jul 15
Anthropic AI for Science applications close
The funded-project cohort (up to 50 projects, up to $30K credits each) signals which domains Anthropic subsidizes next in its workbench playbook.
Jul 22
ServiceNow Q2 2026 earnings
First commercial read on whether governed-agent positioning holds up; watch non-seat revenue mix and the $1M+ ACV cohort. Date confirmed by ServiceNow primary announcement.
Jul 22-23
Alphabet, Microsoft, and SAP report
Microsoft is the one to watch: Azure growth consensus ~34-36% and the first material Copilot revenue disclosure. Dates for these three are aggregator-estimated, not vendor-confirmed.
During July
Salesforce Help Agent and pay-per-resolution reach GA
First customer-reported resolution-rate data will matter more than the launch — it either validates or breaks outcome pricing's measurability premise in support.
Aug 31
Sonnet 5 introductory pricing ends
The rate card moves from $2/$10 to $3/$15 per million tokens just as agent-heavy deployments see the tokenizer and effort-level usage curve hit invoices.
Q3 2026
SpaceX/Anysphere (Cursor) close expected
Watch enterprise procurement reaction and whether rival-lab model availability on Cursor changes post-close — a live vendor-risk test for agentic dev tooling.
Late Aug 2026
Salesforce Q2 FY27 and Workday Q1 FY27
Agentforce ARR trajectory (Q1: $1.2B, +205% YoY per Salesforce) is the first scoreboard for outcome-priced support; timing directional per pre-brief coverage.
Changelog
- Scorecard refresh: support challenger changes from the Intercom/Fin lineage (being absorbed into Salesforce, pending close) to Microsoft Service Agent; marketing leader firms from Adobe/Salesforce co-leaders to Adobe; new research, legal, and sales rows added; engineering and security notes updated; operations, finance, and other rows carried from W26.
- W26's BVP outcome-pricing thesis moved from prediction to shipped SKU this week (Salesforce pay-per-resolution GA); pricingShifts now tracks the first GA outcome-priced support agent.
- New vertical under coverage: science/R&D, opened by the Claude Science vs Gemini for Science vs GPT-Rosalind three-way distribution experiment.