Loading today’s stories
Build1 distinct publisher3 min readUpdated
Free quotas expire on the vendor's schedule, and on AI platforms that schedule is measured in months. The exit costs more than re-plumbing, so it belongs in the integration budget.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The useful unit here is notice. Reddit's API stayed free for more than ten years before its 2023 repricing, and Heroku's free dynos ran over a decade before being switched off in 2022, with fraud and abuse cited as the reason [7][2]. Mistral opened a free API tier in September 2024 with no credit card required [8], and the dev.to tally that collects these cases puts the open phase at five months [10]. Measured against Heroku's window, the time you get to plan in has contracted by a factor of at least 24 [21].
The pattern itself is old and named: Cory Doctorow's "enshittification" dates from 2022 and was the American Dialect Society's word of the year in 2023 [20], and someone now maintains a chronological Free Tier Graveyard with a cause of death per entry [5]. Outside AI it still moves at the old pace. SendGrid's free 100 emails a day "forever" became a 60-day trial in 2025 [3], and Firebase moved its free 5GB of Cloud Storage to a paid plan in February 2026 [4].
Inside the AI window, the deprecation clock ran ahead of the pricing clock. In November 2024, about two months after the free tier opened [24], Mistral retired a batch of endpoints including Mistral 7B and Mixtral 8x7B and gave four months to migrate [11]. That landed before the February 2025 free-tier caps and the $14.99 Le Chat Pro tier [9]. By June 2026, OCR pricing had doubled and smaller models were 50 to 100 percent dearer [12]. The rewrite arrived before the first invoice did.
The mechanism previous cycles lacked sits on the input side. Google's Gemini free tier is explicit about the trade: free tokens in exchange for using prompts and generated responses to improve Google's products, with human reviewers permitted to read and annotate them, and the paid tier exempt, which is why the terms say not to submit sensitive, confidential or personal information to the Unpaid Services [16]. Meta's unpaid Llama API terms grant the same right over inputs and outputs [17]. GitHub Copilot has trained on free and Pro interactions since April 2026 unless the user opts out [18]. Twitter never trained on your API calls and Heroku never trained on your app's behaviour [19], which is why leaving those platforms was a re-plumbing job and nothing more.
Perplexity presented its May 2026 reset as a way to invest in faster reasoning models [14]. According to the dev.to account, it arrived weeks after Google tightened free API access, OpenAI placed ads in ChatGPT's free tier, and Anthropic tightened peak-hour limits [15].
The replacement price is sometimes legible in advance. Le Chat Pro at $14.99 a month is $179.88 per seat per year [22]; Perplexity's five Pro searches a day works out to roughly 150 a month before you pay [23]. Those are the easy numbers. With open phases measured in months, the exit has to sit in the same budget line as the integration, or it is not budgeted at all.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
In April 2024 PlanetScale killed its free tier for all users, not just new ones, removing 5GB of MySQL storage and a billion row reads per month, with no grandfathering or gradual phase-out; production databases had to migrate on a deadline.
Heroku killed its free dynos in 2022 after more than a decade, citing "fraud and abuse".
SendGrid's "free 100 emails/day forever" became a 60-day trial in 2025.
Firebase Cloud Storage moved its free 5GB to the paid plan in February 2026.
Twitter's free API was available for years before being killed in 2023.
Reddit's API was free for over a decade before the pricing change in 2023.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Dated specifics, single unverified voice
The cluster rests entirely on one self-published dev.to post. Its strength is specificity: named vendors, dated changes, quoted contract language, concrete caps and prices, all internally consistent. Its weakness is that nothing is corroborated — no vendor announcements, changelogs, terms links or second outlet appear in the supplied material, and the two load-bearing interpretive claims (concurrent lab tightening, absence of training in earlier cycles) are asserted without support.
Broad vendor practice, no usage data
Adoption here means how widely the described behaviour is actually in force, and the observation set is broad: eleven-plus distinct vendors across databases, email, hosting, social APIs, image, music and LLM platforms with dated pricing, licence and deprecation changes, most already effective at publication. What is missing is scale — no counts of affected accounts, workloads, revenue or migrations, and no usage disclosure showing how many teams depended on the withdrawn tiers.
Timeline solid, causal story overstated
The underlying inventory of changes is more concrete than the headline needs, but the framing runs ahead of it in two places: the implication that rival labs gave each other 'cover' to close free access is presented as a coordination pattern with no dates or statements behind it, and the claim that no previous platform trained on free-tier usage is an unverified absolute. With no vendor rationale or serving-cost context, the interpretation is moderately overstated relative to what the single source proves, though the dated events themselves are not inflated.
Advocacy framing, no disclosed vendor stake
The single source is a pseudonymous self-published post arguing a thesis, with the enshittification frame chosen in advance and examples selected to fit it — an engagement and persuasion incentive. Offsetting that, the supplied material shows no sponsorship, competing product, affiliate offer or paid placement, and the piece criticises incumbents rather than promoting an alternative, so no commercial conflict is evidenced.
Verifiable specifics, one uncorroborated source
Confidence is limited by single-publisher sourcing and the absence of any primary link, but is not low: the assertions are precise enough to be checked against vendor pricing pages and terms, are internally consistent across a dozen vendors, and mostly describe changes already in effect. The interpretive layer deserves materially less trust than the timeline.
product
Incogni ranks 13 AI assistants by privacy risk: bigger is worse, except ChatGPT1 distinct publisher
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
invest
The AI trade's weak link is the buyer: Anthropic's best model took 6% of its tokens1 distinct publisher
build
Developer habit, priced at $965B: what Anthropic's run actually proves1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 24, 2026