How to read this timeline
This is a dated record of AI model and product events through 2026, consolidated from twelve separate DevPik posts. Individually those posts were news items with a short useful life. Together, with one thing added, they become a reference.
The thing added is a status label on every entry, because a substantial fraction of what gets reported as an AI "release" is not one:
- Confirmed — officially announced by the company, with a primary source.
- Reported — credible outlet reporting, not officially confirmed.
- Unverified leak — circulated claims with no official confirmation and no independent verification. Treat as rumour.
This distinction is the reason the page exists. In a fast news cycle, leaks and confirmed launches get covered in identical language and are remembered as equally real. Six months later nobody can recall which was which. Labelling them is unglamorous and turns out to be the most useful thing you can do with an archive like this.
Where a claim was later contradicted or quietly never materialised, the entry says so. Entries are not edited to look prescient in hindsight.
April 2026
| Date | Event | Status |
|---|---|---|
| 2 Apr | Claude Code source code reportedly leaked — claims of 512,000 lines, 44 feature flags, an autonomous daemon mode | Unverified leak |
| 6 Apr | GPT-5.4 Computer Use — model operating a desktop environment directly | Confirmed |
| 7 Apr | OpenClaw banned, then shipped a major release days later | Reported |
| 7 Apr | Google TurboQuant — claimed 5× LLM compression with no accuracy loss | Confirmed (vendor claim) |
| 8 Apr | OpenAI "SPUD" — superintelligence programme tied to a four-day workweek | Reported |
| 10 Apr | Claude Mythos / Project Glasswing — capability described as withheld on safety grounds | Unverified leak |
| 13 Apr | Apple N50 AI smart glasses, targeting 2027 | Reported (Bloomberg) |
| 15 Apr | OpenAI shuts down Sora, reportedly $15M/day burn | Reported |
| 16 Apr | Stanford AI Index 2026 published | Confirmed |
| 17 Apr | Mozilla Thunderbolt — $15 open-source AI client | Confirmed |
| 20 Apr | "Spud" GPT-5.5 leak circulating | Unverified leak |
What actually held up. The confirmed entries — GPT-5.4 Computer Use, TurboQuant, the Stanford AI Index, Mozilla Thunderbolt — are all still referenceable and still matter. Of the three unverified leaks, none received official confirmation. The Claude Code source leak in particular generated a great deal of specific-sounding detail (exact line counts, named internal systems) which is characteristic of leak reporting that cannot be checked. It is recorded here because it happened as an event, not because the claims are established.
The pattern worth noticing. April 2026 was dense with capability announcements about agents operating computers directly. That direction was real and continued. The individually dramatic items — the leaks, the "too dangerous to release" framing — went nowhere. Sorting one from the other in real time is close to impossible; sorting it afterwards is easy and almost nobody bothers.
May 2026
| Date | Event | Status |
|---|---|---|
| 6 May | Open Design — open-source alternative to Claude Design | Confirmed |
May was quiet by comparison, which is itself worth recording. Coverage volume in this space tracks announcement cadence rather than actual progress, and a thin month in the news is not a thin month in development. The archive this page replaces had a dozen April entries and one for May — not because April mattered twelve times more, but because April had more launches to write about.
What the 2026 record actually shows
Reading the year as a whole rather than as a sequence of headlines, four things stand out.
Agentic capability was the real throughline. Models operating desktops, running autonomously for hours, calling tools in long chains. Nearly every confirmed release in this record moves in that direction. It was the genuine story of the year, and it was consistently less exciting to read about than the leaks were.
Open weights closed the gap faster than expected. By August 2026, open-weight coding models were posting SWE-Bench Pro figures at or above the proprietary models they were compared against, at one to two orders of magnitude lower cost. In April that was a prediction. It stopped being one. See the LLM API pricing comparison for where that landed.
Shutdowns became as newsworthy as launches. Sora's closure on a reported $15M/day burn was the clearest signal that the economics of frontier AI products had started to bite publicly rather than privately. Capability was never the only constraint.
Leak coverage had a near-zero hit rate. Every entry labelled "unverified leak" above remains unconfirmed. Not one was later validated. That is the single most useful thing in this page: in a year of intense coverage, the confirmed announcements were worth reading and the leaks were worth ignoring, and they were reported with equal confidence at the time.
If you take one operational lesson from this archive: when evaluating AI news, check whether there is a primary source before you plan around it. The label matters more than the headline, and almost no coverage supplies the label.
