Cheaper tokens, vaguer allowances: AI pricing moves in two directions

Cartesia cut its voice agent rate from $0.06 to $0.03 per minute on Free, Pro, Startup and Scale, and in the same change stopped giving away the model calls those agents make: LLM usage during calls, previously free for agents created in the UI, is now billed separately ^[24]. The page presents the lower figure as the standard rate and says the per-minute numbers "do not include telephony or LLM usage" ^[13][14]. Coverage reads the cut as promotional rather than structural: $0.03 through December 31, 2026, against a regular published rate of $0.06 ^[15]. The headline halves. The invoice does not.

Baseten did the opposite of hiding. elephant-1 and mammoth-1 input prices, previously blank on the Model APIs table, now read $0.035 and $0.10 per 1M tokens, and a new joey-1 arrived at $0.01 ^[25]. Third-party tables already carry the same figures, listing "elephant-1 at $0.035 and mammoth-1 at $0.1 per million tokens" ^[19]. The pitch for joey-1 is cost per decision rather than cost per token: coverage reports it "costs $0.01 per million tokens and answers in about 20ms" and prices sorting 10M support tickets at roughly $50 against $500 on mammoth-1 ^[17][18]. A free launch window closed and a rate card replaced it.

OpenAI moved consumption inside the subscription. Pro now includes monthly API credits of $100, $200 or $500, matching the Pro 100, 200 and 500 tiers, while Free, Go and Plus show none ^[22]. The company's own announcement puts it as arithmetic: "Pro plans get $100, $200 or $500 per month by tier" ^[6]. This is not a cheaper seat. It is API spend attached to a seat, claimed in billing settings and reset every cycle, since credits "expire at the end of each billing cycle, with no rollover" ^[7][9]. Plus, Enterprise and Edu are excluded ^[8].

GitHub kept unbundling. The add-ons grid now lists GitHub Code Quality, Code Security and Secret Protection as separate line items, and the sandbox entry dropped its public preview label ^[21]. The live page sells each by the defect it removes rather than by the plan it belongs to: Code Security to "find vulnerabilities with CodeQL and AI Scan, then assign alerts to Copilot to fix", Code Quality to "fix maintainability and reliability issues in pull requests" ^[1][2]. The split itself dates to April 2025 ^[4]. What changed this period is the Enterprise feature list, which now leads with Copilot access and agent governance ^[21].

Canva went the other way on disclosure. The Free plan's AI line, which read "Up to 20 Standard or Premium AI uses", now reads "Limited access to Standard and Premium AI", and the footnote that carried usage estimates is gone ^[23]. The live page refused our fetch, returning a forbidden response, so the capture is the evidence here ^[26]. No price moved. What moved is that a prospect can no longer tell what the free tier includes before creating an account, which makes the free tier harder to compare and easier to change later.

Read only the official pages and the period has one story: vendors are getting precise about what they charge for compute and vague about what they include. Cartesia's page sets $0.03 as the standard rate and carves the model out of it ^[13][14]. GitHub's add-on copy reads like a catalogue of separately purchasable products ^[1][2], consistent with the framing it used for the original split, that "development teams can now adopt Secret Protection or Code Security independently" ^[3]. OpenAI's announcement is one line of arithmetic tying credit to tier price ^[6].

None of these are presented as price increases, and on the published per-unit numbers none of them are. The money moves to the metered layer instead. Cartesia keeps the minute cheap and bills the model ^[14]. Baseten replaces a free window with a rate card ^[25]. OpenAI hands over API spend that expires monthly rather than discounting the seat ^[9]. In each case the subscription line stays still and the variable line does the work.

Coverage caught the inference moves quickly and the packaging moves slowly. Baseten's figures reached third-party tables within days ^[19]. Cartesia's cut was immediately read as a promotion, quoted as "$0.03 per agent minute through December 31, 2026, compared with a regular published rate of $0.06" ^[15], and reviewers are consistent that "LLM usage: billed separately on top of the per-minute rate" ^[16]. Practitioners care about shape more than price: the patch adding joey-1 to an open-source agent warns that "probabilities are estimates, not safety or action-authorization guarantees" ^[20], which is the real constraint on cheap routing models.

Where the web lags our capture is the free tier. Canva's free AI line lost its number on the page, but a directory stating its Canva pricing "was verified on November 4, 2026 from 1 independent source with medium confidence" still publishes a numeric Standard and Premium AI allowance for the free plan ^[10][11], and a guide updated days before the change still describes "up to 20 uses of Standard AI tools, or Premium AI tools" ^[12]. GitHub shows the same lag a layer up: a pricing explainer still has the "Enterprise tier at $1,050/month (50 × $21/user/month) includes all advanced security features by default" ^[5], while the page now sells Code Quality, Code Security and Secret Protection as separately priced add-ons ^[1][2][21]. Directories inherit a number once and keep quoting it after the page stops saying it.

Credits are now the default AI meter. Across the monitor window, credit language appears in 423 diffs against 271 in the prior window, from 217 companies, 176 of them for the first time ^[27]. Agent pricing grew on the same slope, 426 mentions against 286 ^[29], but it is still mostly undifferentiated: only 4 captures meter an agent per outcome, while 21 meter it per seat ^[29]. This period's moves sit on both sides of that line. Baseten and Cartesia price the unit; OpenAI prices the allowance.

Free tiers are where the pressure shows. AI mentions in free-plan copy rose to 159 from 94 ^[28], and inside that we logged 19 tightenings against 41 expansions ^[28]. Canva's is the kind of tightening that does not register as one, because it removes the number rather than the allowance. The monetization mix still leans on inclusion: 53% of AI changes are bundled into existing tiers, 30% are usage, credits or tokens, 9% create or restructure a tier, and 8% are a separate paid add-on ^[30]. GitHub's unbundling runs against that 53% and with the 8%.

Volume keeps climbing. AI-relevant changes reached 222 in W45 after 203 in W44 ^[31], and this period produced 106 AI-relevant diffs out of 324 pricing changes ^[32]. The AI price index gives no reading this week: app and infra both sit at 100, the spread is 0, and there are 0 movers on either side ^[33], which is a statement about an empty basket rather than about prices. The signal is in the mix and the direction of disclosure, not in the index.