ChatGPT for PowerPoint Is GA: Automating Client Decks

ChatGPT for PowerPoint is now generally available for ChatGPT Business workspaces, per the July 6, 2026 entry on OpenAI’s Business release notes — and for marketing teams, the interesting part is not the demo, it’s the meter. The add-in lives inside real Microsoft PowerPoint, drafts and revises actual editable slides, and after a one-month free window starts billing in token-based credits.

This lands in a busy week for OpenAI’s workplace push — our coverage of today’s broader ChatGPT Work launch covers that bigger story. This post stays deliberately narrow: the PowerPoint agent specifically, what it costs once billing starts on August 6, and a concrete standard operating procedure for the highest-frequency agency use case — the monthly client reporting deck.

Everything below is sourced from OpenAI’s own help-center documentation — the release notes, the product help page, and the ChatGPT rate card — plus independent trade coverage where it adds context. The credit math in Section 03 is our own modeling, computed from OpenAI’s published per-token rates.

The product is a PowerPoint-native add-in, not a separate slide generator. It installs from the Microsoft Marketplace — or via an admin-deployed manifest XML for organizations that can’t reach the Microsoft Store — and sits under Home → Add-ins inside Microsoft PowerPoint itself. That placement matters more than it sounds: because the agent operates on the actual file, output stays as editable slide structure your team can keep working on, rather than the flat exported images that plagued the first generation of AI deck tools.

Availability is broader than the Business-workspace headline suggests. Per the official help page, ChatGPT for PowerPoint is available globally to Free, Go, Plus, Pro, Business, Enterprise, ChatGPT Edu, and K-12 users — Free and Go get limited usage, while Plus, Pro, and Business usage is subject to the plan’s agentic usage limit. This is also not a surprise drop: the add-in entered global beta around May 26, 2026, so GA on July 6 closes a roughly six-week beta-to-GA arc.

As of July 6, 2026, ChatGPT Workspace Agent runs — and ChatGPT for Excel/Sheets and PowerPoint — moved to token-based credit pricing. There is no flat per-deck fee: credit use is computed from three meters — input tokens, cached input tokens, and output tokens. On GPT-5.5, the rate card lists 125 credits per million input tokens, 12.50 per million cached input tokens, and 750 per million output tokens. Two ratios do all the strategic work: output costs 6× fresh input per token, and cached input costs 10× less than fresh input.

One important note on framing: OpenAI publishes credits per task and credits per million tokens, but no credits-to-dollar exchange rate for Business and Enterprise plans — so all cost comparisons in this post stay in credits.

OpenAI’s own worked example makes the meter concrete: a run with 20,000 fresh input tokens, 80,000 cached input tokens, and 5,000 output tokens on GPT-5.5 totals roughly 7.25 credits — 2.5 for input, 1 for cached input, 3.75 for output. Note where the money goes: the 5,000 output tokens cost more than the 100,000 combined input tokens. Every optimization that follows falls out of that asymmetry.

Three more mechanics matter for budgeting. First, GPT-5.4 is a live cost lever: it runs the same three-tier meter at exactly half the GPT-5.5 rate — 62.50 / 6.25 / 375 credits per million — which makes it worth testing for lower-stakes, high-frequency runs. Second, the credit pool is shared, not siloed: ChatGPT for Excel, ChatGPT for PowerPoint, and Workspace Agents all draw from the same agentic usage pool, so a deck-heavy month compresses headroom for everything else. Third, the cutover itself caught teams off guard — the original free-until date for Workspace Agent billing was May 6, 2026, and per Tech Times’s reporting OpenAI extended it to July 6 via a quiet release-notes update on May 22 rather than a headline post.

Zoom out and this cutover is part of a measurable industry shift, not an isolated OpenAI decision. Credit-based and metered pricing models for AI features grew 126% year-over-year in 2025 — a figure from pricing analyst Kyle Poyar’s Growth Unhinged, citing PricingSaaS data, as reported by Tech Times — and the July 6 Office add-in cutover is arguably the largest single instance of the seat-fee-to-metered-credit migration so far. Poyar’s framing of why vendors like the model: credits give customers “a pool of usage that feels relatively straightforward.” For buyers, the practical translation is that workflow discipline now shows up directly on the bill — which is exactly what the next section quantifies.

The trade coverage of this launch reports the GA date and the free window, but nobody we found models the recurring cost of the one deck every agency ships on repeat: the monthly client report. So we did. The table below applies OpenAI’s published GPT-5.5 rates — 125 / 12.50 / 750 credits per million input / cached / output tokens — to four workflow disciplines for the same ~15–20 slide monthly deck. The token counts are our own illustrative estimates, not OpenAI-published task sizes; the arithmetic on top of them is exact.

Formula: credits = fresh input × 125/1M + cached input × 12.50/1M + output × 750/1M, per OpenAI’s published GPT-5.5 rates. Worked through: the from-scratch row is 18.75 + 0 + 30 ≈ 48.75 credits; the Skill-plus-fresh-data row is 5 + 1.5 + 26.25 ≈ 32.75; the data-only edits row is 2 + 1.5 + 9 ≈ 12.50; the Q&A-only row is 1.25 + 1.25 + 3.75 ≈ 6.25.

The pattern is the point. A team that rebuilds its deck from scratch every month sits near the top of OpenAI’s own 10–50 typical range; a team that builds a Skill once, caches its template and style guide, and asks only for data-level updates lands around a quarter of that spend on our estimates — and a pure question-and-reformat run drops below the range floor entirely, costing about what a light Excel task does. On these same estimates, a 20-client agency roster rebuilt from scratch runs roughly 975 credits a month against roughly 250 for the disciplined workflow. Caching is not a UX nicety here; on this meter it is the difference between a rounding error and a line item.

The economics above dictate the operating procedure: build the reusable context once, cache it, and spend fresh tokens only on what changed this month. Here is the SOP we’d run for a recurring client reporting deck, with the human checkpoints mapped to OpenAI’s own disclosed limitations rather than to generic caution.

Step 1 — build the Skill once. Encode the client’s template, brand style rules, slide structure, and output expectations into a Skill, invoked from the sidebar via the “+” menu or an @-mention. This is the single highest-ROI hour in the whole workflow — it moves your template and style guide into cached context at a tenth of the fresh-input rate and stops the team rebuilding the same prompt every month.

Step 2 — duplicate, then draft. OpenAI’s help page explicitly advises duplicating the file first for important work — make that non-negotiable for client files. Then let the agent draft from this month’s source material: performance data, meeting notes, connected apps. Teams already automating client reporting will recognize the shape — the deck becomes the last rendering step of a pipeline, not a hand-built artifact.

Step 3 — three named human checkpoints. OpenAI’s help page discloses three specific failure modes, and each one maps to a named review pass before anything reaches a client. Template QA: generated or edited slides “may not always match a preferred style perfectly” — someone owns brand fidelity. Formatting QA: advanced presentation edits — charts, shapes, formatting, slide management — “may be limited or still in development,” so a human finishes the visual layer. Fact QA: OpenAI tells users directly to review claims, numbers, and citations before relying on the output — the numbers person who owns your client reporting dashboards signs off the figures.

Step 4 — ship, then tighten the Skill. Whatever the reviewers corrected this month becomes a rule in the Skill next month. Over a quarter, the correction rate should fall — and if it doesn’t, the deck is telling you the workflow belongs closer to the data source. Designing that layer — which context gets encoded, which QA stays human, which steps chain into the agent run — is exactly the kind of system we build in our content engine engagements.

Note what the SOP does not automate: the narrative. The agent is genuinely useful at asking-the-deck questions — where the story sags, what a slide fails to say, whether the structure fits the audience — but the strategic read on a client’s month is the part they pay an agency for. Automate the assembly; keep the argument human.

Agencies are picking a stack here, not a benchmark winner — the deciding variable is which ecosystem your client files already live in. OpenAI’s add-in is the Microsoft-native play with the deepest ChatGPT-plan integration. Anthropic’s competing Claude for Excel and PowerPoint — released March 11, 2026 — differentiates on shared conversation context across the two add-ins, so the model that analyzed your spreadsheet remembers that analysis while building the deck. And Google’s equivalent push into Slides is the default answer for Workspace-native shops that were never in PowerPoint to begin with.

The honest cross-vendor read: all three now agree on the architecture — the agent goes to where the file lives, not the other way around. The remaining differences are billing mechanics, context handling, and which suite your clients standardized on years before AI entered the picture. If that decision is genuinely open at your organization, it deserves an evaluation against your own decks, not a vendor comparison table — the kind of structured pilot we run inside AI transformation engagements.

The GPT-5.6 disconnect nobody is flagging. GPT-5.6 went GA in ChatGPT the same day this post publishes, and OpenAI’s GA announcement carries a quote from Canva’s Danny Wu describing the model as especially strong on presentations and roughly 1.6× more token-efficient for slide creation. Read carefully, though: that is a claim about GPT-5.6 the model, from Canva’s own usage — not about the ChatGPT for PowerPoint add-in. As of July 9, the rate card lists rates only for GPT-5.5 and GPT-5.4 on Excel/PowerPoint/Workspace Agents, with no GPT-5.6 row. The deck agent did not verifiably get smarter or cheaper this week, and any coverage implying it did is ahead of OpenAI’s own documentation.

Confidentiality has a second party in the room. Because the product runs in Microsoft PowerPoint, use is governed by Microsoft’s Marketplace Terms of Service in addition to OpenAI’s — and per those terms, Microsoft may have the ability to read the content of your PowerPoint files. For unreleased campaign work, M&A material, or anything under NDA, that is a genuine review-with-legal item before rollout, not boilerplate.

The cadence is now a pattern you can plan against. Sibling product ChatGPT for Excel entered beta around early March 2026 and reached GA around early May; PowerPoint entered beta in late May and hit GA July 6. Two data points make a cadence, not a law — but a roughly two-month beta-to-GA arc per Office surface, with billing following about a month after GA, is a reasonable planning assumption for whatever OpenAI ships into the suite next. Budget owners should also assume the GPT-5.6 rate-card row arrives eventually, and — if the GPT-5.4-to-5.5 doubling is any guide — not at a discount.