OpenAI platform bundling: Agents API, verticals, GSA deal

status: open, opened 2026-09-10, last moved 2026-10-22. AI openai agents us pricing healthcare

The thread: OpenAI is assembling every layer of the agent stack into one managed bundle — harness (Agents API on Codex), voice runtime (GPT-Live-1), verticals (Financial Services; Data agent in Work), government distribution (GSA: $0 licenses, 50% off usage) — with the Lehane policy op-ed as the narrative arm.
2026-09-10: opened from the Sep 10 ai digest. Five pieces shipped in one day, each on a different layer of the stack. The aggressive edge is the GSA pricing ($0 license fees, half-price usage, "expanded cyber defense support") — a land-grab in election season; "eligible" is the door worth watching. Lehane's "policy window" post (Sep 9, late-discovered) is the frame: capabilities require safety evidence and durable policy — while the DseWiki disclosure framework sits at day 5 of "coming weeks."
Watch for: Agents API pricing/limits and what "powered by the Codex harness" leaves for the open harness ecosystem (OpenClaw, Grok Build, pi), GSA "eligible" vs procurement reality, the next vertical (legal — Harvey just raised $550M more), GPT-Live-1 vs incumbent voice stacks.
2026-09-11: the enterprise counter-move lands — Salesforce ships Enterprise AI Harness + AI Control Plane (agent data accuracy, governance) and Slackforce Surfaces (Slackbot vibe-codes charts/dashboards/microsites in-chat from connected apps). The harness layer OpenAI bundled is now a competitive category; the enterprise buyer wants the control plane, and Salesforce sells it where enterprises already live. The Astra case-study push (Cognition, Perplexity, MIT quantum) is the same bundling week's customer-proof layer; the practitioner counter-eval (Browne, 3× fix rate) rides in the same newsfeed cluster.
2026-09-12: the bundling goes symmetrical — Grok models rolling into Microsoft 365 Copilot (Nadella: "More model choice coming to Copilot. Welcome Grok!"; Microsoft 365 account, ~11:33 AM ET; Word, Excel). The third frontier lab inside Microsoft's assistant, the same day Amodei calls for pacing. Government-side counterpoint: GitHub Copilot's FedRAMP-Moderate policy still gates to OpenAI + Anthropic models only — see new thread gov-model-gating. Note M365 Copilot ≠ GitHub Copilot; no government-compliance claim attaches to the Grok rollout. ai-social capture (Sep 12) confirms PowerPoint in the rollout scope ("rolling out in Copilot in Word, Excel, and PowerPoint").
2026-09-15: two competitive deltas. (1) Gemini 3.8 Live + 3.8 Live Extended Thinking (Google DeepMind blog, 1:05 PM ET): voice-agent line — reasoning while speaking, early verbal cues + live progress narration, ~real-time visual grounding, 97-language mid-conference switching, background tool execution; Google-selected benchmarks (AA Speech-to-Speech #1 82.6, τ-Voice 68.6%, Big Bench Audio 97.7%); in Gemini API/app/Workspace/Search today; SynthID on all audio. First big ship of the pacing week from the company with no company statement — the voice-runtime layer OpenAI's GPT-Live-1 occupies now has a Google answer with Workspace distribution. (2) Anthropic's Claude for Financial Advisors draws analyst debate (WealthManagement.com, ~11:01 AM ET) — potential praised, workflow fit questioned; the verticals race (OpenAI Financial Services, Sep 10) is now two-sided.
2026-09-15 (dig): pricing + integration detail on the voice line. Gemini 3.8 Live and Extended Thinking both: $0.005/min audio in + $0.018/min audio out (~$1.38/hr of conversation; token basis $3/$12 per 1M) vs GPT-Live-1 $0.05/min voice session billed per second plus backend model at its normal rates (~$3/hr voice-only) — Google ~2.2× cheaper with reasoning/tools included. Architecture delta: Gemini thinks inside the voice model (Extended Thinking: interleaved speak+reason with fillers and progress narration, async-only NON_BLOCKING function calling, client turn_complete=true interrupts generation; base 3.8 Live adds tool scheduling SILENT/WHEN_IDLE/INTERRUPTED) vs OpenAI's full-duplex speech layer that delegates reasoning/tools to a backend agent. Live API wrapper still tagged Preview; video frames stream by default (cost lever). Partners: LiveKit, Pipecat, Vercel, Agora, Fishjam, Vision Agents. Cloud voice wired to local tooling — async function calling makes that first-class on Gemini. AA Speech-to-Speech #1 (82.6) is Google-selected but the price gap is verifiable from both vendors' docs. Entitlements: AI Pro/Ultra cover app surfaces + AI Studio prototyping only ("AI Studio UI only" — direct API use is billed separately), but AI Pro/Ultra bundle $10/$100 monthly Cloud credits usable on the Gemini API (prepay ≥$5 active; credits consumed first; $300 Cloud welcome credits excluded post-Mar 2). OpenAI has no equivalent — ChatGPT and API billing are separate systems.
2026-09-16: the bundle gets its business-model layer — OpenAI's ads platform is official ("Reimagining advertising with AI," openai.com, ~9 AM ET): Sponsored Agents (click an ad in ChatGPT → clearly labeled conversation with a business-sponsored agent, "distinct from ChatGPT's independent answers"; testing with select US advertisers), natural-language campaign creation/analysis in ChatGPT Work via an Ads Manager plugin, AI creative tools in Ads Manager, integrations with HubSpot (first CRM partner) + Shopify (first ecommerce partner). Digiday had reported the Sponsored-Agent test Sep 14; the openai.com post is the announcement layer. Disclosed safeguard = labeling + context separation; no placement policy published. Same day: Codex for OSS renewed — $100/month Pro plans for maintainers, grants doubled 5,000→10,000 (Marsh, via steipete RT). Unbundling counterpoint in the same window: Thomson Reuters spurned provider offerings and built its own in-house LLM (ComputerWeekly) — verticals are also being defended in-house.
2026-09-16 (evening): the assistant surface reaches glasses — Snap "Specs Intelligence" (Verge, 7:40 PM ET): anticipatory AI assistant for Snap's Specs platform, iOS + Mac, "helps you manage what needs attention today, so you can make progress toward your longer-term goals"; connects other accounts (work tasks, travel). Verge pitches it in the Muse/Gemini Spark genre. Consumer-agent surface count grows again; the differentiated layer is the glasses form factor + cross-account anticipation, not the assistant itself.
2026-09-17 (backfill, tweet-grade): ChatGPT for Clinicians — Karan Singhal (OpenAI health lead), Sep 10 1:49 PM ET, surfaced late 1:30 AM ET Sep 17: verified US clinicians get free GPT-6 Astra (Pro) access; chatgpt.com/plans/clinicians. The verticals layer's health datapoint, missed in the Sep 10 window. Product page not verified this sweep; treat as Sep 10 fact.
2026-09-17/18: the legal vertical watch item lands — Astra for Law (openai.com, ~8 PM ET Sep 17): frontier intelligence for law, custom firm workflows, connected legal data sources, legal-grade controls for confidential client work. Customer-proof layer same window: Cooley "GO Public" on ChatGPT Work (IPO work, openai.com case study). Verticals now two-sided and two-deep: OpenAI (financial services, legal, health) vs Anthropic (financial advisors, LSVP). Next watch: Anthropic's legal answer, pricing/controls detail, whether the Harvey relationship deepens or competes.
2026-09-21: the verticals race gets its analysis layer — Business Standard reads Astra for Law against Thomson Reuters's in-house model and Anthropic's verticals; watch items unchanged.
2026-09-23: three moves in one window. (1) ChatGPT Voice gets connectors (email/calendar/Slack) powered by Astra/Sol/Luna, plus voice-driven docs/decks in Work — "rolling out globally today" (~1:12 PM ET); Viticci: OpenAI has "the only bidirectional voice model that supports third-party connectors." Voice becomes an agentic surface — on the same bundle where Sponsored Agents ads will run. (2) Daybreak cyber access extended to Ukraine's government for civilian-infrastructure cyber defense (~9 AM ET) — the defense-vertical datapoint. (3) Creator Product: Patreon co-founder Sam Yam (+ Rowny, Ma) hired to lead it, "ahead of upcoming announcement" (~7 AM ET) — the creator-economy layer gets a leader; watch the announcement. Also: Astra case studies (Airbnb engineering, Harvey legal drafting, invideo, Parallel) and MentalHealthBench — the verticals/eval layers both advancing the bundle.
2026-09-29: DevDay keynote day — 1 PM ET / 10am PT (OpenAI: "10am PT, on the dot"; Sama's day-before tease: "We have found a new thing"). Pre-keynote framing (Verge): OpenAI has fallen behind in continuously-running consumer-facing agents — rumored answer is Aeon, positioned against Meta's Muse, SpaceX's Grok Bot, OpenClaw. If it ships, the stack becomes: Agents API + voice + verticals + ads + consumer agent. Watch what Aeon actually is (agent vs orchestrator vs feature), pricing, and whether the verticals push continues.
2026-09-29 (DevDay outcome, evening): the stack's consumer, model, harness and billing layers all moved in one day. (1) Dots — the rumored "Aeon" shipped as Dots: always-on, Astra-powered proactive assistants with customizable avatars and their own cloud computer (openai.com + Verge) — the consumer-agent layer, Meta Muse's direct competitor. Stack read: Agents API + voice + verticals + ads + consumer agent = complete. (2) GPT-6.1 Sol — "near-Astra" at one-fifth Astra API prices (gpt-6-astra thread's price ladder extends; Theo claims TB4 > Opus 5.5 at ~1/30th the price — vendor claim). (3) Sign in with ChatGPT federates subscriptions — ChatGPT Plus/Pro covers AI usage in Notion (nichochar), Pi users access OpenAI subs/credits — the subscription, not the API call, becomes the distribution instrument. (4) Open models in Codex (Kiely, ~5:36 PM ET): GLM-5.3 Flash + Kimi K3 natively in Codex, spend counted against OpenAI enterprise commitments — the harness is now a storefront for its cheapest competitors' flagship models (efficient-tier). (5) Codex Security Cloud with Daybreak Blue cyber models default. Same evening Altman made the endgame instrument public: no OpenAI IPO "until the company can make confident safety claims" (Q&A after the keynote, ~8:19 PM ET; no timeline).
2026-10-05: three deltas on the bundle. (1) The ad layer moves into the output itself — visual ad format puts images of sponsored products on screen during generated-image outputs ("initially" when you generate images; US testing "later this month," openai.com + Verge ~9 AM ET; ledgered review-by 2026-10-19, hard 10-31), plus measurement/attribution partners and brand-suitability controls. The Feb slate (name/logo/link sponsored section) is now media-embedded; disclosure moves from chat slots to image content. Same week textGrain watermarking landed (provenance thread) — the trust stack is being built in one product. (2) Codex Security Cloud in research preview inside ChatGPT — GitHub repository scanning, Daybreak Blue models only (mixed-news syndication, single outlet; the DevDay cyber bundle's first productization). (3) Competitor symmetry, from ai-social: tetsuo RT — Grok 4.7 now on Amazon Bedrock (on record Sep 29) and now Google's Gemini Enterprise Agent Platform (X-grade, single tweet, docs check pending); the model-distribution symmetry race follows Grok→M365 Sep 12. Pricing: 500k context, $2/$6 per M tokens below 200k prompt.
2026-10-06: the customer/partner layer accretes — Ironclad (openai.com: training/evaluating agents on complex contracting workflows, to "advance computer use for professional work"); Atlassian partnership expansion (frontier models connected with enterprise knowledge); Jump Trading quant-research case study (longer-running ChatGPT workflows, human review). All self-published case-study grade; the verticals shelf grows a contracting-workflow skill. (Separately today: the OpenAI math batch release — see openai-navier-stokes.)
2026-10-07: the consumer face of the rollout — GPT-6 + Intelligent UI goes global in ChatGPT ("GPT-6 for everyone," openai.com, feed-anchored 8:00 PM ET; Verge 3:10 PM ET): responses compose interactive visuals — diagrams, charts, forms, tappable buttons — "rolling out to all users alongside GPT-6," with the model trained on when to favor visuals over text. The consumer default keeps diverging from the work tier (Astra), and this model switch doubles as a UI overhaul. Watch: roll-vs-opt-in detail, ads-format interaction (ledger review-by Oct 19).
2026-10-08→09: the case-study shelf keeps accreting — Asana browser agent "76× cheaper, 5× faster" (Oct 9; a source mismatch is on record: openai.com snippet says GPT-6 Astra in Codex while the headline says GPT-6.1 Sol — flagged unverified); LegalOn Codex costs −65% (Oct 8); Sophos on Daybreak: threat-investigation time −96%, 52% of MDR cases automated (Oct 9); Radisson hotel discovery in ChatGPT; Oracle days-to-minutes (Oct 8). All self-published; the DevDay cyber bundle now has its first customer-claim layer (Sophos/Daybreak).
2026-10-15 (ai-social window 2, ~2:19 PM ET — https://x.com/OpenAI/status/2110797830361436604): GPT-Live-1.1 — "New voices. Better accents. And less 'checking...' bringing these updates to Voice in ChatGPT, Codex, and dots, with new voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese. Rolling out today to all users. Coming soon to the API." First version bump on the bundle's voice-runtime layer, and the first update spanning all three voice surfaces — ChatGPT, Codex, and the Dots consumer agents. Shipped on the consumer side (a shipped artifact, not an announcement); API availability a "coming soon" claim. Watch: API arrival and its pricing against the recorded $0.05/min per-second-billed ladder.
2026-10-15: two verticals land on the bundling stack, answering the Sep 10 opening watch that named legal as the next vertical. (1) Sullivan & Cromwell's Agreement Analyzer (openai.com, published Oct 15, no time in capture — https://openai.com/index/sullivan-cromwell): developed with OpenAI to identify issues in a draft agreement — company-graded single-engagement case study. (2) The Rosalind Academic Plan (~3:00 PM ET — https://openai.com/index/introducing-the-rosalind-academic-plan): eligible academic labs get sponsored access to GPT-Rosalind and life-sciences research tools — the name's first record appearance, adjacent to the Anthropic wet-lab genre and the Gemini 4 Argon science watch. Both announced, not shipped: no customer counts, no eval artifacts, "eligible"/"sponsored" undefined. Watch: Agreement Analyzer proof beyond the single case study, Rosalind's first named labs, GPT-Rosalind science claims held to the validation standard.
2026-10-16 (openai.com, ~3:00 PM ET): the customer-proof/verticals layer adds a regulated-market bank — Itaú building generative AI for its entire customer base via the ia.i conversational interface and the IARA governance platform, applying its AI First strategy to customers and employees in one of Brazil's most regulated sectors. Companion: the Stack Overflow for Agents post gets its primary URL — signups tripled and up to 10× growth in agent creation via the ChatGPT plugin — the knowledge-share layer joining tooling and harness on the bundle. https://openai.com/index/itau-generative-ai
2026-10-22 (openai.com, no time in capture — https://openai.com/index/uber-chatgpt-work-rollout): Uber rolls ChatGPT Work out to 30,000+ employees — research, data analysis, presentations, complex workflows; the bundle's customer layer moves from small-business franchisee case studies to a named corporate rollout at company scale. Self-published, no third-party check.
2026-10-22 (~10:00 AM ET, MSN syndication — https://www.msn.com/en-us/technology/artificial-intelligence/openai-drug-discovery-swarm-decoded-7-7-million-brain-cells-in-32-minutes/ar-AA2fPjHC): the GPT-Rosalind strand (first record appearance: the Oct 15 Rosalind Academic Plan) gets its first quantitative science claim — OpenAI, PerturbAI and Scripps Research ran nearly 2,000 GPT-Rosalind agents that decoded 7.7 million brain cells in 32 minutes, in what the report calls a drug-discovery milestone. Single-outlet syndication, company-graded; the standing "GPT-Rosalind science claims held to the validation standard" watch now has a first datapoint to test.