Issue 15 · Week 32 of 2026.

Industry- or function-specific application packages that landed this period — from frontier labs, open-weights labs, and the insurgent vertical-AI startup cohort. Each entry names the vertical it serves and the decision implication for buyers of the displaced tier.

• 2026-08-04/Salesforce/Sales/Incumbent SaaS/—

  Agentforce Coworker

  Headless-first autonomous teammate indexing 300+ enterprise sources and following users into Slack, Teams, ChatGPT and Claude — announced as GA and as beta on the same page

  Put the availability status in writing before sizing a rollout: the launch post states 'Agentforce Coworker is Generally Available' and later states it 'is in beta for Salesforce today, with web, Microsoft Teams, ChatGPT, Claude, and desktop app coming later this year.' The architecture is the more interesting commitment — building headless so one agent carries business context into rival assistants concedes that the assistant surface is not winnable and bets on context and governance instead. Verify that Platform permission inheritance actually holds when the agent is invoked from ChatGPT or Claude rather than from Salesforce, because that is where the trust boundary is novel and untested.

  Salesforce blog; Salesforce product page

• 2026-08-05/Salesforce/Security/Incumbent SaaS/—

  Agentforce 360 inside Missionforce National Security (IL5)

  IL5 authorization brings the full Agentforce 360 portfolio to national security workloads on AWS GovCloud, covering CUI and unclassified NSS data

  This is the clearest signal yet that agentic platforms are clearing the accreditation bar rather than waiting for it, and regulated commercial buyers should use it as leverage: an IL5-authorized control set is a stronger reference architecture than any vendor security whitepaper. Ask for the specific authorization boundary and which agent capabilities fall inside it, because portfolio-level announcements routinely cover fewer services than the brand implies.

  Salesforce press release

• 2026-08-03/Microsoft/Operations/Incumbent SaaS/Usage

  Copilot Studio agentic workflow designer

  No-code agentic automation reaches GA for every Microsoft 365 organization, with an agent node that reasons over unstructured inputs instead of rule-based branching

  This changes the build-versus-buy calculus for ops and IT teams that previously needed a developer for anything past simple rules, and its native grounding in SharePoint, Outlook, Teams and Planner is a real advantage over connector-based rivals reaching the same data through third parties. Two cautions: credit pricing is unpublished, so model the cost of a reasoning node running on every inbound request before enabling it broadly, and an agent node that decides the next action is a policy surface, not a workflow step — route it through the same review as any autonomous decision.

  Microsoft Message Center notice MC1442234, via secondary coverage

• 2026-08-05/Meta/Other/Frontier lab/Usage

  Muse Spark 1.2 contributor tier

  Same frontier model at $0.10/$0.20 instead of $1.25/$4.25, in exchange for permission to train on your prompts and completions

  This is the most explicit price ever put on enterprise data in a frontier API — roughly 92% off input and 95% off output for training rights — and it forces a governance decision before a cost decision. Rate limits differ too (100 RPM contributor against 3,000 RPM standard, applied per team rather than per key), so the cheap tier is not a drop-in substitute even where the data terms are acceptable. Any gateway that can route between the two tiers needs an explicit policy control, because an accidental route sends customer content into a training corpus.

  Meta developer pricing and rate limits documentation

Established SaaS vendors reacting to the agentic shift — product launches, repositioning, earnings color, partnerships, and restructuring. The system-of-record incumbents defending their turf against systems of action.

• 2026-08-04/Salesforce/Agentforce Coworker

  Built headless-first so a single agent carries business context into Slack, Microsoft Teams, ChatGPT and Claude rather than requiring users to come to Salesforce

  An incumbent conceding that it will not own the assistant surface, and repositioning onto context, permissions and governance instead, is a strategically honest move that most SaaS vendors have avoided. For buyers it means the CRM is bidding to be the policy layer underneath other vendors' assistants — which is a stronger lock-in than a UI, and worth negotiating as such while the product is still beta.

  Salesforce blog

• 2026-08-03/Microsoft/Copilot Studio agentic workflow designer

  Agentic workflow automation goes GA tenant-wide with native Microsoft 365 grounding, and unpublished credit pricing

  Distribution is the strategy: every Microsoft 365 organization now has no-code agent automation available without a procurement event, which is how Copilot Studio wins against better-designed standalone tools. Finance teams should demand the credit schedule before enabling it, because a reasoning node evaluating every inbound email has a fundamentally different cost curve from a rule.

  Microsoft Message Center MC1442234, via secondary coverage

• 2026-08-06/OpenAI/ChatGPT chat, ChatGPT Work and Codex

  ChatGPT chat moved to August GPT-5.6 Sol and Luna variants while Codex and ChatGPT Work stayed on the July checkpoints

  One vendor's model name now maps to different checkpoints by surface, which breaks the assumption behind most enterprise model-approval records. Governance teams should record the surface and checkpoint alongside the model name in every approval, and re-run acceptance tests per surface rather than per model.

  OpenAI Deployment Safety Hub

• 2026-08-05/Meta/Meta Model API tiering

  A two-tier rate card separating standard pricing with no training rights from a ~92% discount that grants them

  Data-for-discount is becoming an explicit, priced product rather than a buried terms-of-service clause, which is better for buyers who read carefully and worse for those who route on price alone. Expect other vendors to copy the structure, and expect the first incident to involve a gateway routing regulated content to a contributor tier by default.

  Meta developer pricing and rate limits documentation

Vertical-AI startups raising capital, winning named customers, or shipping general-availability product. The cohort that sits between frontier labs moving down the stack and SaaS incumbents defending their record-of-truth.

• 2026-08-06/Irregular/Security

  A misconfigured evaluation environment at one external cyber testing vendor produced disclosed incidents at Anthropic, OpenAI and Meta

  Third-party AI evaluators are now a named concentration risk, and almost no enterprise vendor questionnaire asks about them. Add the question — which third parties produce your safety evidence, and are they shared across our suppliers — because this week the answer was the same firm for three of the four largest labs.

  OpenAI and Meta disclosures, via Simon Willison

• 2026-08-06/Taalas/Engineering/$219M raised since 2023; purchase price undisclosed

  AMD agreed to acquire the Toronto startup that etches model weights into ROM, after $219M of venture funding since 2023

  The second specialized-inference exit in eight months after Nvidia's $20B Groq purchase, which tells application builders that per-token decode costs on high-volume agentic workloads are expected to fall through hardware rather than through model efficiency. Do not re-architect for it yet — the deal closes in Q4 and no product has been named — but treat multi-year inference cost assumptions as soft.

  AMD Newsroom; CNBC

• 2026-08-04/Liquid AI/Operations

  LFM2.5-2.6B ships with on-device throughput of roughly 30 tok/s on a phone under 2.5 GB, under a license withholding commercial use above $10M revenue

  The on-device capability is genuinely useful for offline and privacy-constrained workflows, and the license disqualifies most of the enterprises that would want it. This is the third revenue-gated open release in a month, so add license classification to the intake step of any open-weight evaluation rather than treating it as a late legal review.

  Liquid AI blog; LFM Open License v1.0

Public pricing-model shifts inside the window. The 'data owns the application' thesis predicts a structural move from seat-based to outcome-based pricing across SaaS; tracking the rate of change is itself a market signal.

• 2026-08-05/Meta

  Usage-based→Hybrid

  Muse Spark 1.2 gains a second rate card: standard at $1.25 input / $4.25 output per 1M with no training rights, and a contributor tier at $0.10 / $0.20 granted 'in exchange for permission to use your prompts and completions to train future Meta models.' Cached input falls from $0.15 to $0.002, and rate limits differ at 100 RPM contributor against 3,000 RPM standard, applied per team rather than per key. Pricing is now partly denominated in data, which makes tier selection a governance control rather than a cost setting.

  Meta developer pricing and rate limits documentation

• 2026-08-05/Meta

  Usage-based→Usage-based

  Effective cost per task rose from $0.29 on Muse Spark 1.1 to $0.40 on Muse Spark 1.2, measured on the Artificial Analysis Intelligence Index. List pricing did not change. The ~38% increase comes entirely from token consumption — input tokens up roughly 53% and output up roughly 36% per task, concentrated in long-horizon agentic evaluations. This is the clearest available demonstration that a rate card no longer predicts a bill, and any budget built on blended per-token assumptions should be re-derived after each model upgrade.

  Artificial Analysis

• 2026-08-03/Alibaba / Qwen

  Unknown→Usage-based

  Qwen3.8-Max reaches API GA at $2.00 input / $6.00 output per 1M with cache hits at $0.25. The 88% cache-hit discount is unusually large and makes prompt-caching discipline the dominant cost lever on this model. Offsetting it, Artificial Analysis measured 150M output tokens to run its Intelligence Index against a 66M class median and labelled the model 'very verbose' — so the effective bill depends far more on the workload's caching structure than on the headline rate.

  Artificial Analysis model page

• 2026-08-03/Microsoft

  Seat-based→Usage-based

  The Copilot Studio agentic workflow designer moves ops automation from developer-built workflows inside a seat-priced suite to credit-metered agent execution. The capability is now available tenant-wide without a procurement event, while the credit schedule is not published in the rollout notice. A reasoning agent node that evaluates every inbound email, document or request has a materially different cost profile from the rule-based branching it replaces, so enable it against a named budget and a metered pilot rather than tenant-wide by default.

  Microsoft Message Center MC1442234, via secondary coverage

A snapshot of leader-vs-challenger by vertical. Useful for procurement shortlists when matching workload to vendor cohort. Rows refresh weekly as leadership shifts.

• Legal

  Leader: Harvey

  Challenger: Legora

  No qualifying in-window movement; the category still has no independent matter-level ROI study, which remains the gap that would settle it.

• Security

  Leader: Salesforce (Agentforce 360 at IL5)

  Challenger: Microsoft (MDASH + MAI-Cyber)

  IL5 authorization for a full agentic portfolio is the strongest accreditation event in the category to date and moves the leader position on regulated deployability rather than on model quality.

• Engineering

  Leader: Anthropic (Claude platform)

  Challenger: OpenAI Codex / GPT-5.6 + GitHub Copilot

  Epoch's MirrorCode leaderboard shows Claude Fable 5 at 64% against GPT-5.6 Sol at 20% on long-horizon coding, which widens rather than narrows the runtime lead recorded since W30.

• Finance

  Leader: Microsoft Dynamics

  Challenger: Oracle Fusion Agent Studio + Gemini

  Unchanged; no in-window product date cleared the bar, and Alphabet's DeepMind reorganization adds execution uncertainty to Oracle's announced Gemini path.

• Support

  Leader: Salesforce (Agentforce)

  Challenger: Sierra (Horizon)

  Agentforce Coworker extends the position into cross-surface work, but its availability status is stated inconsistently in the launch post so the extension is not yet contractible.

• Commerce

  Leader: Salesforce (Agentforce Commerce)

  Challenger: OpenAI (ChatGPT checkout)

  Unchanged; transaction ownership remains decisive and no in-window commerce product event moved the position.

• Operations

  Leader: Microsoft (Copilot Studio)

  Challenger: ServiceNow

  GA of the agentic workflow designer puts no-code agent automation in front of every Microsoft 365 tenant without a procurement event, which is a distribution advantage ServiceNow cannot match on workflow quality alone.

• Research

  Leader: OpenAI (Academic Researchers)

  Challenger: Google (Gemini Enterprise Agent Platform)

  Unchanged on product, but Google's research organization was restructured in-window and four senior leaders left, which is execution risk against the challenger position.

• Marketing

  Leader: Adobe

  Challenger: Salesforce

  No in-window shift; both incumbents continue to benefit from falling inference costs without surrendering data or distribution.

• Other

  Leader: Snowflake (Cortex AI Gateway)

  Challenger: Provider-native MCP surfaces

  The Endpoint Accuracy Index strengthens the gateway case by making per-endpoint serving quality a governable variable, provided the gateway is willing to disclose which endpoint it routed to.

Patterns that crossed multiple vendors this period. One pattern, several exemplars, what it changes for procurement, cost, or vendor-risk posture.

• Approved-model catalogs govern labels, not behavior

  GPT-5.6 Sol now maps to different checkpoints in ChatGPT chat versus Work and CodexIdentical open weights lose up to ~40% of reference accuracy depending on the serving endpointAgentforce Coworker announced as both generally available and in beta on the same page

  Most enterprise AI governance is built on a list of approved model names, and this week produced three independent demonstrations that a model name does not determine what runs. A vendor can change the checkpoint behind a name by surface, a gateway can serve the same weights at materially lower accuracy, and a launch page can describe the same product at two different maturity levels. The remedy is not more approvals — it is recording the resolvable facts alongside the name: surface, checkpoint, endpoint, serving configuration, license class, and availability status. Any approval record that cannot answer 'what exactly was tested' is a compliance artifact rather than a control.

  OpenAI; Artificial Analysis; Salesforce

• Vendor questionnaires do not reach the evaluation supply chain

  Anthropic, OpenAI and Meta incidents all traced to one external evaluator's misconfigurationUK AISI ran agents with internet access and cyber classifiers disabled by deliberate configurationOpenAI now plans to give third-party testing partners recommended security controls

  Enterprise AI risk assessments ask vendors about their own controls and take published safety evaluations as evidence. This week showed that the evidence is produced by a small number of third parties whose configurations are invisible to the buyer and, in at least one case, shared across competing suppliers. The practical addition to a questionnaire is three questions: which third parties produce your safety and capability evaluations, what containment do those environments guarantee, and do you know whether our other AI suppliers use the same firms. None of that requires new tooling, and until this week nobody was asking it.

  OpenAI and Meta disclosures via Simon Willison; UK AISI technical paper

• Agents are being built headless so they follow the user into rival assistants

  Agentforce Coworker designed to run in Slack, Microsoft Teams, ChatGPT and ClaudeCopilot Studio agent nodes grounded natively in SharePoint, Outlook, Teams and PlannerOpenAI replaced its App Directory with a Plugin Directory packaging skills and templates

  The application layer has stopped fighting for the chat window. Salesforce is explicitly building an agent that carries business context into competitors' assistants, Microsoft is embedding agent nodes wherever its data already sits, and OpenAI is repackaging integrations as portable plugins. The strategic implication for buyers is that the durable asset is context and permission inheritance rather than the interface, and the durable risk is that permissions asserted inside one vendor's platform have to hold when the agent is invoked from another. That boundary is new, largely untested, and the right subject for a proof of concept before a rollout.

  Salesforce; Microsoft; OpenAI

• Data rights are becoming an explicitly priced line item

  Meta contributor tier at $0.10/$0.20 against standard $1.25/$4.25 for training permissionContributor rate limits of 100 RPM against 3,000 RPM, applied per team rather than per keyChatGPT Work usage billed on a credit system with admin-set workspace and group spend limits

  Two different pricing structures arrived this week that make governance decisions inseparable from cost decisions. Meta's contributor tier puts a roughly 92% discount on granting training rights, which turns a data-protection question into a budget conversation and creates an obvious failure mode where a routing layer sends regulated content to the cheap tier. Credit-metered agent products create the mirror problem, where an autonomous agent's spend is bounded by admin configuration rather than by task scope. In both cases the control that matters is a policy at the routing or workspace layer, and it has to exist before the first production workflow rather than after the first invoice.

  Meta developer documentation; OpenAI ChatGPT Work admin documentation

Forward catalysts in the next 7–30 days that would change the read materially — earnings prints, conferences, expected product launches, regulatory decisions, and competitive responses.

• By Sept 30

  A contractible GA date for Agentforce Coworker

  The launch post describes the product as both generally available and in beta, with web, Teams, ChatGPT, Claude and desktop 'coming later this year'. Rollout sizing and any committed business case depend on which of those is written into the order form.

• By Aug 31

  Copilot Studio agentic workflow designer credit pricing

  The capability is GA tenant-wide with the credit schedule unpublished. A reasoning node evaluating every inbound request has a materially different cost curve from the rules it replaces, and finance cannot model it until the schedule exists.

• By Sept 30

  Whether any lab publishes containment requirements for third-party evaluators

  OpenAI said it will provide testing partners with recommended security controls for high-risk evaluations. Whether that becomes a published standard determines if this week's shared failure is fixed industry-wide or one private contract at a time.

• By Aug 31

  Astra's revised launch date and final capability rating

  Any roadmap that assumed a named unreleased frontier model now carries schedule risk from a safety framework rather than from engineering. Buyers with dependent commitments should re-plan on the July checkpoints in the meantime.

• Ongoing

  Whether gateways begin disclosing which endpoint served a request

  If serving configuration moves measured accuracy by tens of percent, then a gateway that will not name the endpoint cannot support an acceptance test. This is the single most useful disclosure any model-routing vendor could add.