How the work changed, 2020–2026

Original research edition

This is the supplied September 2026 report, including its caveats and source links. Its prices, vendor events and verification claims have not all been independently rechecked for this preview. Consult the original correction log alongside the chapters.

From API to Agent: A History of the AI Product and Productivity Market, June 2020 – September 2026

Opening historical section of the deep research brief. All dates verified against primary sources (vendor announcements, regulator and research publications) or reputable press. Contested figures are flagged as contested in the text. A list of items that could not be verified appears at the end.

I. The API Era: June 2020 – November 2022

The modern AI product market has a precise birthday. On 11 June 2020, OpenAI published a blog post announcing a general-purpose API for its text models, and — this is the part everyone forgets — explicitly stated that it was "launching today in a private beta rather than general availability." The founding condition of the industry was scarcity of access, not scarcity of ideas. For seventeen months, the only way to build on GPT-3 was to get through a waitlist, and OpenAI reviewed applications before they went live. The waitlist did not come down until 18 November 2021.

That gating shaped everything that followed. In 2020–21 the binding constraint on an AI product company was not engineering talent or capital but permission, and a small population of developers who had permission built a recognisable genre: a text box, a set of prompt templates, and a subscription. Jasper (launched as Conversion.ai in 2021), Copy.ai (which raised twice in 2021, closing an $11M round that October) and Writesonic were the canonical examples. They were marketing-copy generators with template libraries — "50 headlines for your landing page" — and for about eighteen months they printed money, because they were selling access to a capability that most buyers could not obtain any other way. Jasper's 18 October 2022 Series A — $125M at a $1.5B valuation, announced nine days before Stable Diffusion's first month was out — was the peak expression of that arbitrage.

Meanwhile a genuinely different kind of product was being built. On 29 June 2021 GitHub announced Copilot as a technical preview, built on OpenAI Codex, with a limited number of spots. It went generally available on 21 June 2022 at $10/month. Copilot mattered more than the wrapper boom did, for a reason that took the industry years to internalise: it was not a place you went to use AI, it was a behaviour that happened inside a tool you were already using, against context the tool already had. The inline completion in your editor knew your file, your imports, your naming conventions. The copy generator knew a prompt template.

2022 also brought images. Midjourney's Discord server opened on 14 March 2022 and entered open beta on 12 July 2022. OpenAI announced DALL·E 2 in April 2022, opened a beta that July, dropped the waitlist on 28 September 2022, and shipped an API in public beta on 3 November 2022. Stability AI released Stable Diffusion's weights publicly on 22 August 2022. Within five months the industry had a proprietary hosted model, a community-distributed model with no real product surface, and an openly downloadable model — three distribution strategies that are still the three distribution strategies today.

What it changed. The API era established the default architecture of an AI product: a thin application calling someone else's model, with the application's value concentrated in prompt engineering, UI and go-to-market. It also established the default business risk of an AI product, which is that your supplier can become your competitor overnight.

What people got wrong. Nearly everything about the durability of wrappers. The consensus in 2021–22 was that prompt engineering was a defensible craft and that vertical copywriting tools would compound into platforms. What actually happened is that the model vendors shipped chat interfaces that did the same job for free, the underlying models improved fast enough to erase the value of hand-tuned templates, and the incumbents whose documents the copy was destined for — Microsoft, Google, Notion — bolted equivalent features into products customers already paid for. Jasper laid off staff in July 2023, nine months after the $125M round, and cut its internal valuation. The lesson is not "wrappers are bad": Cursor is a wrapper by any strict definition and was reported in April 2026 to be raising at a $50B valuation. The lesson is that a wrapper survives only where it owns something the model vendor does not — workflow, data, distribution, or a buying centre — and in 2021–22 almost none of them did.

A second thing people got wrong: the persistent 2022 prediction that image generation would kill stock photography and commercial illustration outright. It damaged both and restructured both, but Getty and Adobe are still here, and the productivity impact landed less on "replacing artists" than on the unglamorous middle of the market — thumbnails, mockups, internal decks.

II. The ChatGPT Shock: November 2022 – December 2023

On 30 November 2022 OpenAI published "Introducing ChatGPT," a free research preview of a model fine-tuned from the GPT-3.5 series. It was not a new capability. The underlying model had finished training in early 2022 and had been available through the API. What was new was that it was free, conversational, and required no waitlist — and that turned out to matter more than any capability jump in the period.

The adoption numbers from this period are the single most-repeated and least-verified statistics in the field. The figure "100 million users in two months, the fastest-growing consumer application in history" originates in a 1 February 2023 UBS analyst note built on Similarweb estimates, not on OpenAI disclosure. It was directionally right and precisely unfounded. Treat it as an estimate, always.

The institutional reaction was faster than the product reaction. Google reportedly declared an internal "code red" around 21 December 2022. Microsoft announced a "multiyear, multibillion dollar" extension of its OpenAI partnership on 23 January 2023 and shipped an AI-powered Bing and Edge on 7 February 2023. Google announced Bard on 6 February 2023, built on a lightweight LaMDA and opened only to trusted testers; general access in the US and UK did not arrive until 21 March 2023. That six-week gap between announcement and availability became a recurring signature of the era: everyone announced on their competitor's schedule and shipped on their own.

14 March 2023 was the densest day of the period. OpenAI released GPT-4. Anthropic launched Claude, with Notion, Quora and DuckDuckGo as named early partners. Two days later, on 16 March 2023, Microsoft announced Microsoft 365 Copilot. Notion AI had already gone generally available on 22 February 2023 as part of Notion 2.21, making it one of the first mainstream productivity products to ship a generative feature to its whole base rather than a waitlist.

Then came the gap that defines the era's real lesson. Microsoft 365 Copilot was announced in March 2023. Microsoft announced its GA date on 21 September 2023, and it became generally available to enterprise customers on 1 November 2023 at $30 per user per month — nearly eight months of demos before anyone could buy it, and with a 300-seat minimum at launch. In between, essentially every SaaS company in the world shipped a "✨ AI" button. Most of these were the same thing: a summarise/rewrite/draft menu wired to a general-purpose model, priced as an add-on, with no access to the customer's actual workflow state.

Funding followed. 2023 was the year "AI" became a fundable prefix, and a large share of the capital went to companies whose product was a thin layer over an API in a vertical — which is to say, the 2021 wrapper thesis with a bigger TAM slide.

What it changed. Two structural things. First, buying moved. Before ChatGPT, AI was bought by data science teams; after it, AI was bought by the CIO, the CFO and increasingly by end users on personal credit cards, and the shadow-IT problem became the dominant enterprise governance story for two years. Second, the free consumer chatbot became the benchmark against which every paid feature was measured. If your $20/month AI writing assistant was not visibly better than a free chat window, you had no product.

What people got wrong. The loudest failed prediction of 2023 was that search was finished. Bing's share barely moved; Google's search revenue kept growing. The second was that GPT-4 represented a rate of improvement that would continue monthly — GPT-4 shipped in March 2023 and its true successor generation did not arrive for over a year. The third, and most expensive for buyers, was the assumption that an announced enterprise Copilot was a shippable enterprise Copilot. Organisations that budgeted in March 2023 for a 2023 rollout spent the year waiting.

Conversely, the category that everyone expected to die and didn't: the document and note-taking category. Notion, Confluence, Google Docs and Microsoft Word were all supposed to be obsoleted by "just ask the AI." Instead they became the place the AI was used, because that is where the context lives.

III. The Copilot Year: 2024

2024 was the year the industry discovered that shipping an assistant and getting it used are different problems.

Microsoft 365 Copilot's first full year was hard. The product was capable in demos and inconsistent in the wild, and its weakest point was the one Microsoft had assumed was its strongest: enterprise content. Copilot surfaced whatever a user technically had permission to see, which in most large tenancies was a great deal more than anyone realised, and "oversharing" remediation became a prerequisite project rather than a footnote. Pilots stalled at renewal because usage concentrated in a small fraction of licensed seats. Microsoft did not disclose paid seat counts in 2024 at all — the first hard numbers came much later.

Model competition normalised. Google rebranded Bard to Gemini on 8 February 2024 and launched a paid Gemini Advanced tier, ending the awkward two-brand period. Anthropic released Claude 3 on 4 March 2024 and Claude 3.5 Sonnet on 20 June 2024. OpenAI released GPT-4o on 13 May 2024, notable less for raw intelligence than for being twice as fast at half the price — the first clear signal that the competitive axis was shifting from capability to cost-per-useful-token.

The AI-native challengers established themselves in this year. Perplexity built a real audience for answer-style search. Cursor turned an AI-first fork of a code editor into the fastest-growing developer tool of the period. Glean sold enterprise search-plus-assistant into exactly the gap Microsoft's Copilot had left open — it raised $150M at a $7.2B valuation in June 2025, on the back of 2024 traction. The common thread is that all three replaced a workflow rather than decorating one.

2024 also produced the first serious evidence. The Brynjolfsson–Li–Raymond study "Generative AI at Work" (NBER working paper w31161, issued April 2023, revised November 2023, published in the Quarterly Journal of Economics in 2025) studied 5,179 customer support agents and found a 14% average increase in issues resolved per hour — but 34% for novice workers and close to nothing for the most experienced. This is the most important single result of the period and the most consistently misquoted. The finding is not "AI makes workers 14% more productive." It is "AI compresses the gap between novices and experts in a task where expertise is largely tacit knowledge that can be written down." Whether that generalises to tasks where expertise is not codifiable was, and remains, open.

Hardware was the year's comedy. Humane announced the AI Pin on 9 November 2023, shipped in April 2024, and was reviewed on 11 April 2024 with a brutality rarely seen in consumer tech; it cut the price to $499 on 23 October 2024; HP bought the assets for $116M in February 2025 and the service went dark on 28 February 2025. Rabbit's r1 followed a similar arc from CES 2024 hype to widespread disappointment. Consolidation began quietly and in an unusual form: on 19 March 2024 Microsoft hired Inflection's founders and most of its staff, with Mustafa Suleyman taking over Microsoft AI; Amazon did a comparable deal with Adept in June 2024; Google did one with Character.AI's founders, reported on 2 August 2024. These "acquihire-plus-licence" structures — buying the team and the technology while leaving a hollow corporate shell and its investors behind — became the template for AI consolidation and drew antitrust attention without, in most cases, antitrust consequences.

What it changed. Buyers started asking for evidence instead of demos, and vendors started needing to answer questions about permissions, data residency and audit before questions about quality. The unit of sale began drifting from "a licence" to "a deployment project."

What people got wrong. The dominant 2024 prediction was the dedicated AI device — that a new hardware form factor would disintermediate the smartphone. It did not come close. The inverse error was underrating coding: through most of 2024 AI coding assistance was widely described as autocomplete for juniors, and the serious objection was that it produced plausible code that did not compile. Within eighteen months, agentic coding would be the single largest revenue line in applied AI.

IV. The Agent Turn: 2025

The pivotal artefact of this period was published quietly at the end of the previous one. On 25 November 2024 Anthropic open-sourced the Model Context Protocol, a standard for connecting models to tools and data. It solved an unglamorous N×M problem: every assistant needed a bespoke integration with every system of record. MCP made the integration a property of the system, not of the assistant.

Its adoption was unusually fast for an open standard. OpenAI adopted it on 25 March 2025; Google DeepMind committed on 9 April 2025; Microsoft built it into Windows and its agent tooling over the course of 2025. On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare and Bloomberg, at which point there were more than 10,000 active public MCP servers. In a market where nothing else was standardised, the plumbing was.

Agents themselves arrived in three waves. Computer use — Anthropic shipped a public beta on 22 October 2024; OpenAI launched Operator on 23 January 2025 as a research preview for Pro subscribers. Both were impressive demos and poor products; Operator was shut down on 31 August 2025 and folded into ChatGPT agent. Deep research — long-running, multi-step, citation-producing report generation — arrived from Google in December 2024 and OpenAI in February 2025, and was the first agentic pattern that ordinary knowledge workers used voluntarily and repeatedly, because the failure mode (a mediocre report) was cheap.

Coding agents were the breakout. Cognition introduced Devin on 12 March 2024 with a claim to be "the first AI software engineer"; early hands-on testing in 2024–25 found it well short of the demo. Anthropic shipped Claude Code as a limited research preview alongside Claude 3.7 Sonnet on 24 February 2025 and made it generally available with Claude 4 on 22 May 2025. OpenAI shipped Codex, a cloud software-engineering agent, on 16 May 2025. Cursor added agent mode. By the second half of 2025 the terminal-and-repository agent, not the chat window, was the highest-value AI product form factor in existence.

The Windsurf saga compressed the entire market's dynamics into 72 hours. OpenAI had reportedly agreed in early May 2025 to acquire Windsurf for ~$3B. On 11 July 2025 the deal collapsed — reportedly in part because Microsoft's OpenAI agreement would have given Microsoft access to the IP — and the same day Google paid roughly $2.4B to licence Windsurf's technology and hire CEO Varun Mohan, co-founder Douglas Chen and key engineers. Three days later, on 14 July 2025, Cognition acquired what remained. The episode is the clearest illustration of the acquihire-plus-licence structure: three outcomes, one company, and remaining employees and investors left to negotiate over the residue.

Two counterweights landed in the same summer. On 10 July 2025 METR published a randomised controlled trial in which experienced open-source developers working on their own repositories were 19% slower with early-2025 AI tools — while believing they had been 20% faster. On 18 August 2025 MIT's NANDA initiative published "The GenAI Divide: State of AI in Business 2025," whose headline that 95% of enterprise generative AI pilots produce no measurable P&L return became the most-cited statistic of the year and, almost immediately, the most-contested; the sample and methodology drew sustained criticism, and the number was widely repeated without its qualifiers. Gartner had already predicted on 25 June 2025 that over 40% of agentic AI projects would be cancelled by end-2027. None of these sources agreed with each other in detail. All of them pointed the same direction: deployment, not capability, was the bottleneck.

Consolidation in productivity software became explicit. Grammarly, having acquired Coda in early 2025 and Superhuman in July 2025, rebranded the whole company as Superhuman on 29 October 2025, launching a bundled suite and a proactive assistant. Atlassian agreed to buy The Browser Company for $610M on 4 September 2025 and developer-productivity platform DX for $1B on 18 September 2025. Microsoft used Ignite on 18 November 2025 to announce Agent 365 and Work IQ — an attempt to make managing a fleet of agents an enterprise SKU.

And on 27 January 2025, DeepSeek triggered the largest single-day market-cap loss in Nvidia's history and a broad AI selloff, on the thesis that a competitive frontier-class model had been trained far more cheaply than assumed. The immediate market reaction over-corrected and reversed within weeks. The durable consequence was different and slower: it broke the assumption that frontier capability implied a permanent cost moat, and it started the price war that made 2026's economics possible.

The year closed with the competitive positions of 2022 exactly inverted. Google shipped Gemini 3 Pro on 18 November 2025 to strong reception, and on 2 December 2025 Sam Altman reportedly declared an internal "code red" at OpenAI, delaying advertising plans to focus on ChatGPT. Three years after Google's code red, OpenAI had one.

What it changed. The product primitive stopped being a response and became a task. That broke per-seat pricing, because an agent that resolves a support ticket does not map to a seat, and it broke evaluation, because you cannot A/B test a four-hour autonomous run the way you test a completion.

What people got wrong. The 2025 consensus that "2025 is the year of agents in production" was wrong for enterprises and right for developers — agents went into production overwhelmingly in software engineering, where the environment is instrumented, the output is testable, and rollback is free. The prediction that MCP would fragment into competing vendor standards was wrong. The prediction that RAG vendors and vector databases were a durable category was largely wrong; long context and native connectors absorbed most of it. And the widely repeated claim that the METR result proved AI coding tools don't work was a misreading — METR measured experienced maintainers on mature repositories with early-2025 tooling, which is close to the worst case for AI assistance, and said so.

V. 2026 to Date: The Reckoning That Hasn't Happened Yet

Through 15 September 2026, the year has been defined by three things running at once: extraordinary revenue growth, an unresolved argument about whether that growth justifies the capital behind it, and a quiet purge of everything that isn't working.

Consolidation and shutdowns. The most telling events were subtractions. In late March 2026 OpenAI shut down Sora, its consumer AI video app, roughly six months after launch; reporting attributed it to collapsing usage (a worldwide peak near one million users falling below 500,000), roughly $1M/day in compute cost, and a decision to redirect compute toward coding and agents where Anthropic's Claude Code was taking share. Disney, which had reportedly committed $1B to a Sora partnership, learned of the shutdown under an hour before the public did. On 9 July 2026 OpenAI announced it was shutting down the Atlas browser, launched only in October 2025, with the explicit conclusion that "the browser is a feature, not the destination"; its agentic browsing was redistributed into a Chrome extension and the ChatGPT desktop app. Both were described internally as "side quests." Elsewhere the pattern was absorption rather than death: Asana acquired no-code agent builder Stack AI for $75M on 28 May 2026; Google and Salesforce deepened cross-platform agent interoperability on 22 April 2026. The middle of the AI application market — single-feature tools with no proprietary workflow — is being bought for tens of millions or quietly wound down.

Model cadence and pricing. The release tempo accelerated and, critically, prices did not rise. OpenAI shipped GPT-5.5 on 23 April 2026, GPT-5.5 Instant on 5 May 2026, and on 9 July 2026 GPT-5.6 in three tiers (Sol, Luna, Terra) alongside ChatGPT Work, an agent that produces documents, spreadsheets and presentations across connected apps. Anthropic shipped Opus 4.6 (5 February), Opus 4.7 (16 April), Opus 4.8 (28 May), the Mythos and Fable 5 models (9 June), Sonnet 5 (30 June) and Claude Opus 5 on 24 July 2026 — at $5/$25 per million input/output tokens, identical to Opus 4.8, while claiming near-Fable-5 performance at half the cost. Google shipped Gemini 3.1 Pro on 19 February 2026 and a rapid series of Flash-tier updates through the summer. The competitive claim of 2026 is no longer "smarter" but "same answer, fewer tokens" — Altman's headline stat for GPT-5.6 Sol was 54% better token efficiency on agentic coding.

The shift toward per-outcome pricing. This is real but partial, and it is being forced by arithmetic rather than chosen. When an agent does the work, seats stop tracking value in both directions: a company can cut seats and increase usage, or keep seats and blow through compute budgets. Salesforce has now cycled through three pricing models for Agentforce in two years — $2 per conversation, then consumption-based Flex Credits, then per-resolution pricing for support agents in mid-2026. Intercom's Fin has sold per-resolution since 2023. Microsoft went the other way, bundling Copilot Chat into Microsoft 365 while raising base prices from July 2026 and selling agent management (Agent 365) separately. The honest summary is that the industry has agreed per-seat is broken and has not agreed on what replaces it; most vendors now run two or three models simultaneously, which is a symptom of an unsettled market, not a strategy.

The enterprise adoption picture, measured properly. This is where 2026 finally produced trustworthy numbers. A Federal Reserve FEDS Note published 3 April 2026 reconciled the competing surveys and found the apparent contradictions were measurement artefacts: roughly 18% of firms reported using AI at end-2025 (Census BTOS), about 41% of workers reported using generative AI for work (Real-Time Population Survey, November 2025), and 78% of workers were employed at firms adopting AI (Survey of Business Uncertainty) — firm-weighted, worker-weighted and employment-weighted measures of the same phenomenon. Daily use at work was 12%. A Census Bureau working paper from April 2026 covering November 2025–January 2026 found 18% of firms firm-weighted, 32% employment-weighted, 50–60% among the very largest firms, and — the key number — 57% of adopters use AI in three or fewer business functions, with AI-driven employment decreases occurring at only 2% of firms. Adoption is broad, shallow, concentrated in cognitively intensive sectors, and so far augmenting rather than displacing. Anyone arguing either "nobody is using it" or "it has already transformed work" is arguing against the data.

Agent reliability. It has genuinely improved, and the improvement is measurable. METR's Time Horizon 1.1 update on 29 January 2026 — using an expanded 228-task suite — put Claude Opus 4.5 at a ~320-minute 50%-success time horizon and GPT-5 at ~214 minutes, with the post-2023 doubling time revised down to about 131 days, i.e. progress roughly 20% faster than the earlier estimate. Notably, the same revision cut older GPT-4 scores by 35–57%, a reminder that these benchmarks are instruments under construction. Against that, Gartner warned on 26 May 2026 that applying uniform governance across heterogeneous agents is itself a leading cause of enterprise agent failure. Capability is compounding; operational practice is not compounding as fast.

The bubble debate. Unresolved, and noisier than in 2025. Revenue growth is not in dispute: Anthropic reportedly reached a $65B annualised run rate at end-July 2026, up from $47B in May and $9B at end-2025, with OpenAI at roughly $40B in early August, double its end-2025 figure. Both companies filed confidentially for IPOs — Anthropic on 1 June 2026, OpenAI on 8 June 2026. Anthropic raised a $30B Series G at a $380B valuation on 12 February 2026. SpaceX acquired xAI in early February 2026 in what CNBC called the largest merger ever at roughly $1.25 trillion. And yet markets sold off twice: on 9 June 2026, and harder on 28 July 2026, when Korea's Kospi fell nearly 11%, the SOX index dropped as much as 6%, and the Nasdaq-100 fell 9.7% from its record high — triggered by Chinese memory maker CXMT's blockbuster debut, reports of Chinese domestic DUV lithography, and generalised fear about hyperscaler capex returns. Analysts on both sides were unusually blunt: one described the panic as "indiscriminate," another noted memory contract prices were rising 20–30%. As of mid-September 2026 the argument is not settled and cannot be settled from revenue data alone, because the question is not whether AI produces revenue but whether it produces enough to service several hundred billion dollars a year of infrastructure.

What people are getting wrong right now. Three things look like live errors. First, reading Sora's and Atlas's shutdowns as evidence of AI retrenchment; they are evidence of focus — both were killed to redirect compute toward agents and coding, which is where the money is. Second, treating the Fed and Census adoption figures as a ceiling; "18% of firms, three functions or fewer" is what early diffusion of a general-purpose technology looks like, and the employment-weighted figures are far higher. Third — and this is the one most likely to age badly — the assumption that frontier model prices will keep falling indefinitely. Opus 5 held price flat rather than cutting it, and the industry's entire 2026 product strategy, including per-outcome pricing, is underwritten by an expectation of continued deflation in cost-per-task that has not been guaranteed by anyone.

VI. What the Whole Arc Should Teach a Reader

Distribution beat capability, every single time. ChatGPT won on being free and frictionless, not on being a new model. Copilot won on being inside the editor. Notion AI won on being inside Notion. The six-year record contains no example of a superior model defeating an inferior model that was already where the work happened.

Announcement dates and availability dates are different products. M365 Copilot: announced March 2023, GA November 2023. Bard: announced February 2023, available March 2023. Devin: introduced March 2024, credible mid-2025. Any timeline built from announcement dates overstates the pace of the field by six to twelve months.

The evidence has always been more equivocal than the discourse. Brynjolfsson's 14% was concentrated in novices. METR's −19% was experts on mature codebases. MIT's 95% was contested on arrival. The Fed's reconciliation showed the surveys were measuring different things. Anyone citing one number without its denominator is selling something.

Categories predicted to die mostly didn't; categories predicted to be safe mostly weren't. Search survived. Documents and wikis became more valuable, not less, because they hold the context. Stock imagery survived in damaged form. Meanwhile the IDE — the safest category in software, unchanged for a decade — was restructured in under three years, and the entry-level knowledge-work task, the thing nobody wrote think-pieces about, is where the measured productivity effects actually landed.

The winning structure changed twice. 2020–22: get API access, add a template. 2023–24: attach a feature to an existing system of record. 2025–26: own a task end-to-end, price it by outcome, and be judged on reliability. Each transition destroyed most of the companies optimised for the previous one.

Milestone Timeline

11 Jun 2020 — OpenAI launches the GPT-3 API in private beta, gated by waitlist
29 Jun 2021 — GitHub announces Copilot technical preview, built on OpenAI Codex
18 Nov 2021 — OpenAI removes the GPT-3 API waitlist
Apr 2022 — OpenAI announces DALL·E 2 (CLIP-latents paper published 13 Apr 2022)
21 Jun 2022 — GitHub Copilot reaches general availability at $10/month
12 Jul 2022 — Midjourney enters open beta
22 Aug 2022 — Stability AI publicly releases Stable Diffusion weights
18 Oct 2022 — Jasper raises $125M Series A at a $1.5B valuation
30 Nov 2022 — OpenAI launches ChatGPT as a free research preview (GPT-3.5)
21 Dec 2022 — Google reportedly declares an internal "code red" over ChatGPT
23 Jan 2023 — Microsoft announces a multiyear, multibillion-dollar OpenAI investment
1 Feb 2023 — UBS note using Similarweb data estimates ChatGPT at ~100M MAU (estimate, not disclosure)
6 Feb 2023 — Google announces Bard for trusted testers
7 Feb 2023 — Microsoft launches AI-powered Bing and Edge
22 Feb 2023 — Notion AI ships generally available in Notion 2.21
14 Mar 2023 — OpenAI releases GPT-4; Anthropic launches Claude
16 Mar 2023 — Microsoft announces Microsoft 365 Copilot
21 Mar 2023 — Bard opens to users in the US and UK
Apr 2023 — NBER w31161 "Generative AI at Work": +14% average, +34% for novices (revised Nov 2023; QJE 2025)
17 Jul 2023 — Jasper lays off staff nine months after its $125M round
1 Nov 2023 — Microsoft 365 Copilot reaches GA at $30/user/month
9 Nov 2023 — Humane announces the AI Pin
6 Dec 2023 — Google announces Gemini
8 Feb 2024 — Google rebrands Bard to Gemini and launches Gemini Advanced
4 Mar 2024 — Anthropic releases Claude 3
12 Mar 2024 — Cognition introduces Devin, "the first AI software engineer"
19 Mar 2024 — Microsoft hires Inflection's founders; Mustafa Suleyman leads Microsoft AI
11 Apr 2024 — Humane AI Pin reviews published; widely panned
13 May 2024 — OpenAI releases GPT-4o: 2x faster, half the price of GPT-4 Turbo
2 Aug 2024 — Google hires Character.AI's founders, completing the year's acquihire pattern
22 Oct 2024 — Anthropic ships computer use in public beta with upgraded Claude 3.5 models
25 Nov 2024 — Anthropic open-sources the Model Context Protocol
23 Jan 2025 — OpenAI launches Operator as a research preview
27 Jan 2025 — DeepSeek-driven selloff; Nvidia posts a record single-day market-cap loss
18 / 28 Feb 2025 — HP acquires Humane's assets for $116M; AI Pin services shut down
24 Feb 2025 — Anthropic ships Claude 3.7 Sonnet and Claude Code as a limited research preview
25 Mar 2025 — OpenAI adopts MCP; Google DeepMind commits on 9 Apr 2025
16 May 2025 — OpenAI launches Codex, a cloud software-engineering agent
22 May 2025 — Claude Opus 4 and Sonnet 4 launch; Claude Code reaches general availability
10 Jun 2025 — Glean raises $150M Series F at a $7.2B valuation
25 Jun 2025 — Gartner predicts over 40% of agentic AI projects will be cancelled by end-2027
10 Jul 2025 — METR RCT finds experienced developers 19% slower with early-2025 AI tools
11 Jul 2025 — OpenAI's Windsurf deal collapses; Google licenses the technology for ~$2.4B and hires its leaders
14 Jul 2025 — Cognition acquires the remainder of Windsurf
18 Aug 2025 — MIT NANDA "GenAI Divide" report: 95% of enterprise pilots show no P&L return (contested)
31 Aug 2025 — OpenAI shuts down Operator, folding it into ChatGPT agent
4 / 18 Sep 2025 — Atlassian agrees to buy The Browser Company ($610M) and DX ($1B)
29 Oct 2025 — Grammarly rebrands as Superhuman and launches Superhuman Go
18 Nov 2025 — Google ships Gemini 3 Pro; Microsoft announces Agent 365 and Work IQ at Ignite
24 Nov 2025 — Anthropic releases Claude Opus 4.5
2 Dec 2025 — Sam Altman reportedly declares an internal "code red" at OpenAI
9 Dec 2025 — Anthropic donates MCP to the Agentic AI Foundation under the Linux Foundation
29 Jan 2026 — METR Time Horizon 1.1: post-2023 doubling time revised to ~131 days
2–3 Feb 2026 — SpaceX acquires xAI; CNBC values the combination at ~$1.25 trillion
12 Feb 2026 — Anthropic raises a $30B Series G at a $380B valuation
19 Feb 2026 — Google releases Gemini 3.1 Pro
Late Mar 2026 — OpenAI shuts down the Sora consumer video app
3 Apr 2026 — Federal Reserve FEDS Note reconciles AI adoption surveys (18% of firms; 41% of workers)
17 Apr 2026 — Cursor reported in talks to raise $2B+ at a $50B valuation on ~$2B ARR
22–24 Apr 2026 — Google Cloud Next: Gemini Enterprise agent platform, Agent Marketplace, Workspace Intelligence
23 Apr 2026 — OpenAI releases GPT-5.5
29 Apr 2026 — Microsoft reports 20 million paid Microsoft 365 Copilot seats (FY26 Q3)
26 May 2026 — Gartner warns uniform agent governance is itself a cause of enterprise agent failure
28 May 2026 — Asana acquires Stack AI for $75M; Anthropic releases Claude Opus 4.8
1 / 8 Jun 2026 — Anthropic, then OpenAI, file confidentially for IPOs
9 Jun 2026 — Anthropic previews Claude Mythos 5 and releases Fable 5; AI stocks sell off
9 Jul 2026 — OpenAI releases GPT-5.6 (Sol/Luna/Terra) and ChatGPT Work; announces the Atlas browser shutdown
24 Jul 2026 — Anthropic releases Claude Opus 5 at unchanged Opus pricing
28 Jul 2026 — Chip and AI selloff: Kospi −11%, SOX −6%, Nasdaq-100 9.7% off its high
17 Aug 2026 — Anthropic reported at a $65B annualised run rate; OpenAI at ~$40B

Sources

Not Verified

The following were used with hedging, or deliberately excluded, because I could not confirm them against a primary or high-authority source:

DALL·E 2's exact announcement day. OpenAI's page confirms April 2022 and the associated paper is dated 13 April 2022; the commonly cited 6 April 2022 blog date was not independently confirmed. Timeline entry left as "Apr 2022."
Founding dates for Jasper (as Conversion.ai), Copy.ai and Writesonic. Only secondary/company-profile sources; no precise founding dates asserted in the text.
Jasper's post-layoff internal valuation figure (reported as a cut from $1.5B to ~$1.2B). Reported only by secondary outlets; described qualitatively.
Rabbit r1's CES 2024 unveiling date and shipping dates. Not confirmed from a primary source; mentioned without dates.
DeepSeek-R1's release date (widely reported as 20 January 2025). Only the 27 January 2025 market selloff is dated in the timeline.
OpenAI Deep Research's exact launch day (commonly given as 2 February 2025) and Google Gemini Deep Research's (11 December 2024). Described by month only.
GPT-5's launch date (widely reported as 7 August 2025) and the subsequent GPT-4o-removal backlash. Referenced only indirectly; not placed in the timeline.
Microsoft's "30 million paid Copilot seats" claim. Only an aggregator/syndicated source; the timeline uses the confirmed 20M figure from the 29 April 2026 FY26 Q3 earnings call.
Gemini 3.5 / 3.6 / 3.7 / 3.8 Flash release dates (May–September 2026). Sourced only from Wikipedia's release table; referenced in the text as "a rapid series of Flash-tier updates" without specific dates. Gemini 3 Pro (18 Nov 2025) and 3.1 Pro (19 Feb 2026) are corroborated by press.
The exact ChatGPT Atlas shutdown date (aggregators say 9 August 2026). Only OpenAI's 9 July 2026 announcement is dated here.
Anthropic's Claude Mythos 5 / Fable 5 / Sonnet 5 release dates. From Wikipedia's model table, corroborated indirectly by the Opus 5 announcement's references to Fable 5 and Mythos 5; treat the exact days as provisional.
Perplexity's 2026 valuation, revenue and user figures. Available only from low-authority aggregators; excluded entirely.
Microsoft 365 July 2026 list-price increases and the E7 SKU. Reported only by partner/reseller blogs; described directionally, not with figures.
Salesforce Agentforce's precise pricing-model transition dates. The three successive models are well attested in aggregate; individual effective dates are not.
The claim that Disney had committed $1B to a Sora partnership and learned of the shutdown under an hour before the public. Reported by TechCrunch citing a Wall Street Journal investigation; the WSJ original was not directly accessible.
The 9 June 2026 selloff's magnitude. CNN's report was inaccessible (robots.txt); the event is described qualitatively, with the better-documented 28 July 2026 selloff carrying the numbers.