Preview, Beta, GA: What Vendors Said vs What Coverage Said

Preview, beta and general availability are the three words an AI vendor uses to tell you how much to trust a feature, and they are also the three words coverage most often changes on the way to a headline. This ledger puts the two side by side for every dated vendor announcement we could find between 17 and 22 August 2026: the vendor’s status word, read verbatim from its own page, next to the word the coverage used, read verbatim from the outlet.

The question matters because a status word is often the only contractual signal a buyer gets. “Public preview” at GitHub means the feature is provided as-is and may be withdrawn without notice. “Out of beta” at Anthropic means a request header is no longer required. “GA” in a headline about Cursor means, as it turns out, nothing Cursor said. Thirty-six rows, one claim per cell, with every source linked so the comparison can be re-run.

The dataset holds labs and developer platforms, coding tools, distribution layers, martech and ads platforms, and three control rows (a regulator and two securities filings) where the vendor’s wording is definitive by construction. Rows where vendor and coverage agree stay in; so do the ten rows where no independent coverage could be located at all.

The unit of this ledger is a dated announcement, not a product. An announcement qualifies if a vendor published it on its own surface (blog, changelog, release notes, GitHub Releases, an SEC filing or a regulator’s press release) with a date inside 17–22 August 2026, and if it carries, or conspicuously lacks, an availability word. The population is the set of such items named in our own batch anchor document for the week, which was itself built by primary source sweeps; this table did not go looking for announcements beyond that list.

The five match categories were fixed before any row was scored. They are deliberately about words, not about whether the feature works: a row scores “same” if the outlet repeated or fairly paraphrased the vendor, whatever either party’s word was.

Three published definitions anchor this section, and none of them belongs to a lab. Only GitHub’s populates a cell in the table: Google Cloud’s is the reference ladder the rest are read against, and Microsoft Learn’s convention is background for a single row. Google Cloud’s launch-stage page states that at Preview “products or features are ready for testing by customers”, that “no SLAs or technical support commitments are provided” and that “The average Preview stage lasts about six months”, while General Availability products “are open to all customers, ready for production use, and covered by a Google Cloud SLA, where applicable” (cloud.google.com/products, launch-stages section). GitHub’s site-wide Pre-release License Terms define the umbrella: “‘Pre-release’ means software, online services, and additional products and features that are not yet generally available, such as private preview, public preview, early access, technical preview, or similar versions,” provided “‘AS-IS’, ‘WITH ALL FAULTS’ and ‘AS AVAILABLE.’ GitHub may change or discontinue Previews at any time without notice.” Microsoft Learn’s general convention, used as background for the LinkedIn row rather than as a single-URL quote, is that preview features come without a service-level agreement and are not recommended for production workloads.

Everything else in the table is undefined by its author. That is not an accusation; it is the reason the column exists. When Anthropic writes “is out of beta on the Claude API”, the release notes make clear that the operational meaning is that the anthropic-beta header is no longer required and that requests still sending it keep working. It is a technical statement, not a service commitment, and the table records it as such.

Every row is below, including the 21 where coverage and vendor agree and the ten where no coverage was located. Words in curly quotes are verbatim from the linked page. “Match” uses the five categories above. A handful of commercial SaaS and martech destinations carry nofollow; everything else links plainly. Outlets are named with their URLs because each row is a citation, not a criticism: the point is that the two words can be checked, not that a particular desk got one wrong.

Counting the Match column gives 21 rows scored same, 3 upgraded, 2 not in vendor copy, 0 downgraded and 10 with no independent coverage word located. Restricting to the 26 rows where coverage was actually found, 21 of 26 (81%, n=26) agree and 5 of 26 (19%, n=26) do not. Every one of the five moved toward the more definite word: toward “GA”, “unveils”, “rolls out”, “beta” or “warns”. No sampled outlet hedged below its vendor. The chart is drawn directly from that column; no derived arithmetic sits between the table and the bars.

Row 9 (Cursor) is the row that sets the denominator. Its vendor page carries no maturity word, and the firmer characterisation circulates in aggregator summaries and secondary paraphrase — but no single outlet headline carrying it was fetched and confirmed, so the row is recorded as coverage not located rather than scored as an upgrade. Scoring it as an upgrade instead would read as six mismatches in 27; the direction of the other five is unchanged either way.

Two interpretive points follow. First, the mismatches cluster where the vendor left a gap: no maturity word at all (Anthropic’s brand-new browser tool, the HubSpot feature page), a progressive verb (“starting to introduce”), or a deliberately non-final posture (“previewing”, “proposed”). Where a vendor used a crisp status word (“public preview”, “developer preview”, “Rolling out in September”, “Generally available”), sampled coverage repeated it. The pattern in this sample is not that outlets overwrite clear labels; it is that absent or soft labels get resolved to something firmer on the way to a headline.

Second, the direction is uniform in a 26-row sample, which is suggestive rather than conclusive. A larger census, or a second window, could surface downgrades (an outlet calling a GA feature “beta” out of caution, as the Grok Bot row nearly does). Until then the defensible claim is narrow: in this window, when the word changed between vendor page and coverage, it changed upward.

This table began with three suspected mismatches from the week’s anchor work. Each was re-fetched independently for the ledger rather than carried forward, because a mismatch that rests on one fetch is a claim, and one that rests on two is a finding.

Cursor (row 9): the word that is not on the page

The 19 August changelog for always-on cloud agents contains neither “GA” nor “general availability”, confirmed on two independent fetches. The closest sentence is a continuing-work statement, quoted in full in the table. The “GA” framing is real in the sense that it circulates, but no single outlet headline carrying the literal word was fetched and confirmed, so the row is recorded as coverage not located rather than attributed to a named desk. The feature itself is covered in our always-on agents guide and in the Cursor and Codex cloud-agents piece.

HubSpot (row 25): a label with no confirmed author

The knowledge-base article for Company Surge, dated 11 August, carries no “beta” in the raw HTML body; the only occurrence is an unrelated i18n dictionary string. A separate rendered fetch showed a “Beta” token adjacent to the page title that could not be resolved as a genuine badge or as an artefact of the HTML-to-text conversion, and an in-product badge sits behind login. The row therefore attributes “now in beta” to the third-party trackers that used it and makes no claim about what HubSpot said. Feature detail is in our Company Surge note.

OpenAI (row 7): previewing versus unveiling

OpenAI’s own page says “Today, we’re previewing Private Safety Processing” and that it “is currently being tested with early customers”, with broader rollout and a technical white paper planned for September 2026. Two outlets led with “Unveils”. This is the cleanest upgrade in the table: a limited early-customer test described by the vendor as a preview, reported as a completed launch. Background on the mechanism is in our Private Safety Processing preview note.

Fourteen of the 36 vendor pages (rows 2, 6, 9, 12–14, 16–18, 21–23, 27 and 30) carry no maturity word of any kind. These are the rows most exposed to a later upgrade, because a reader, a tracker or a summariser has to supply one to write a sentence about them. Four patterns appear in the table.

The contrast between rows 10–12 is the sharpest illustration in the set. GitHub published three Copilot changelog entries three days apart. Two carry an explicit public-preview disclaimer and a pointer to terms that define what that means; one carries nothing. The sandbox-billing sentence appears in the Teams entry only, which is the kind of detail our Copilot in Slack and Teams note covers. Same vendor, same week, three different disclosure postures.

Three rows form the control group, rows 34–36: wording that is definitive by construction, so that the hedged rows have something to be measured against. The NVIDIA 8-K (row 35) and the Marvell 8-K (row 36) are legal instruments with a single verb each (“guarantees”, “issued”). Coverage of both filings matched the filings: CNBC’s “backing” and “giving” are fair paraphrases, and both outlets avoided “investment”, a word the NVIDIA filing does not support since the $105B figure is a financing ceiling.

The regulator row is the one that failed. The FTC press release (row 34) is as hedged as a regulator can be: “proposed”, “draft statement”, “seeking public comment”, plus an explicit disclaimer of authority to ban the practice outright. Quartz’s headline was “FTC warns retailers on personalized pricing disclosure rules”. There is no rule, and the release does not warn; it asks. This row and the HubSpot row are the table’s two “not in vendor copy” scores: here the word is simply absent from the release; for HubSpot the visible page copy does not confirm it, and an unresolved rendered token near the title keeps the question open.

Two rows cut against the “coverage inflates” reading and are kept for that reason. Grok Bot (row 21): xAI’s expansion notice says “now included” with no maturity label, while 9to5Mac’s framing retained “in beta”, the label the product launched with ten days earlier. Coverage was, if anything, more conservative than the vendor page. GLM-5.3 (row 22): VentureBeat’s “hits the API” is a plain factual headline for a plain factual milestone, and neither side reached for a tier word. The alias that sits on top of that model, z-ai/glm-latest (row 23), drew no independent coverage at all, which is its own small finding given that it is the row most likely to surprise a developer later; our floating-alias note explains why.

The five Anthropic rows are also a negative finding of a kind. Four of the five were scored same even though the vendor said “out of beta” and the headline said “GA”, because the two describe the same operational state. The fifth (browser use, row 2) is scored upgraded only because a blanket “GA” headline swept in a brand-new tool that was never in beta. The distinction is fine, and the table keeps it rather than rounding all five up to a mismatch. Detail on all five is in our Claude Platform betas-to-GA note.

The table supports four habits for anyone who has to turn a vendor announcement into a decision. None of them requires trusting this ledger; each requires only the vendor page.

• Read the page, not the headline. In five of the 26 rows where coverage was located, the coverage word was firmer than the vendor’s; in a sixth (Cursor, row 9) the firmer word circulates but no headline carrying it could be pinned at all. The vendor page is one click further, and it is the only wording the vendor itself published.
• Look for the definition, then assume its absence. Only GitHub’s rows link to published terms that say what “preview” costs you (as-is, may be withdrawn without notice). For every other vendor in the table, the status word means what the sentence around it says and nothing more.
• Treat “no word” as a question, not an answer. Fourteen vendor pages carry no maturity word. A missing preview disclaimer is evidence of nothing in particular until the vendor is asked; the JetBrains settings row shows how a GA reading can rest entirely on silence.
• Separate the status from the event. The Sol price cut is not a GA announcement, the Grok Bot expansion is not a launch, and the FTC statement is not a rule. In each case the underlying thing is real and the status word being attached to it is not on the page.

Our procurement questions checklist carries the contractual side of this: what to ask for in writing when the status word is undefined. Teams that want the same read-the-primary discipline applied to a vendor shortlist before a commitment can find it under our AI transformation work.