Credits as the universal currency

What's happening — and why

How it works

Evidence over time

29 supporting · 5 counter — hover or tap a point for detail, click to jump to the row.

Evidence

Trivia

• Credits became the most common billing unit in the corpus (roughly 35% of companies with taxonomy data at 97 companies), yet no two credit systems define a credit the same way: Cursor's $1 credit = $1 of underlying API cost is the most transparent, while most creative-tool vendors deliberately decouple the credit from any published dollar-to-compute ratio so they can change underlying costs without renegotiating. The same word describes a spectrum of disclosure from fully transparent to fully opaque.

• The 71% correlation between credit-metered vendors and freemium tiers (observed at 61 companies) is the strongest cross-axis correlation in the corpus's credit analysis — stronger than the credit-enterprise or credit-hybrid correlations. The credit pool is architecturally suited to freemium because the free monthly allotment and the paid pool share one currency, making the upgrade trigger (running out of credits) unambiguous to the user.

• The credit-currency pattern spread into enterprise software via GitHub Copilot's June 2026 AI Credits launch — the first time a product with tens of millions of users adopted the credit abstraction for its primary metering unit. Prior to that, the pattern had been concentrated in creative-app and developer-tool SMB segments; GitHub's adoption confirmed that the credit abstraction scales to enterprise procurement volumes.

• At 158 companies the credit lost its crown: seats (78 arrays) overtook credits (70) for the first time across four syntheses, after credits had led at 122 (59) and every prior count. The reversal wasn't a flight from credits — it was the corpus absorbing 36 new companies skewed toward seat-anchored vertical SaaS, exposing that "credits are the most common unit" was always a sampling artifact of an AI-native-heavy corpus.

• Credits and seats turned out not to compete but to stack: in the seat+credit hybrids that dominate AI-coding and outbound sales-tech, the same bill carries both units, so a company can push both counts up at once. That's why seats (78) and credits (70) can each be "near-majority" simultaneously in a 158-company corpus — they measure orthogonal things (who logs in vs. what they consume), not rival pricing philosophies.

• Qodo's 2026-06-30 switch breaks the "credits and seats stack, never substitute" finding: it dropped per-seat pricing entirely for a pure credit pool, the corpus's first AI-coding vendor where the credit didn't layer on top of the seat but replaced it. The orthogonality holds as a corpus-wide average, but at the vendor level the credit can now eat the seat — when the team-wide pool is the only meter, "who logs in" stops being billable at all.

• The seat-versus-credit horse race that kept this trend at "weakens" for seven weeks was never a race: at 353 companies 88 of the 137 credit vendors (64%) also bill seats, so nearly two-thirds of the "credit" column is sitting inside the "seats" column on the same invoice. The 2026-07-30 recount also killed the direction of the old claim — credits per seat-vendor rose from 123/175 (0.70) at 338 companies to 137/188 (0.73) at 353, so the seat lead narrowed while the label said it was widening.

• Poe (2026-07-28) is the corpus's cleanest measurement of the credit as a silent price lever: the $19.99/mo Premium sticker did not move, but the bundled allowance fell from 1,000,000 to 660,000 compute points — a 34% cut in usage that works out to a ~51% increase in price per point. Nothing on the pricing page changed except a number inside the pool, which is exactly the disclosure gap the credit abstraction creates.

• Ideogram (2026-07-22) repriced by changing the denominator, not the numerator: Ideogram 4.0 entered the credit matrix at the same 2/4/6 credits as 3.0 but quoted per SINGLE image where 3.0 is quoted per FOUR, so a Plus subscriber's 1,000 monthly credits stretch to ~2,000 images on 3.0 Turbo and only ~500 on 4.0 Turbo. Inside one pool the per-image credit spread now runs about 48x, from Ideogram 2a Turbo up to 4.0 Quality — all invisible in the $15 plan price.

• A quarter of the inference cohort has abstracted the token out of existence: `model-inference` → `tokens` holds for only 54 of 115 companies (47%), and 29 of those 115 (25%) bill credits with no token unit published anywhere — abacus-ai, anyscale, cartesia, descript, elevenlabs, freepik, gladia, hedra, heygen, ideogram, livekit, midjourney, murf-ai, playground-ai, poe, recraft, resemble-ai, runway, stability-ai, suno, synthesia and vellum among them. Buyers of a quarter of the inference market cannot see a token rate at all, which is the credit winning as an abstraction even in the one category where the natural unit is universally known.

• Marketing is the corpus's most credit-saturated use case AND its clearest layering cohort: 37 of 51 `marketing` companies (73%) bill credits against a 39% corpus base, and 30 of the 51 (59%) bill seats AND credits on the same invoice against a 25% base. Where the output is heterogeneous — a post, an image, an email sequence, a video — the credit displaces the natural unit rather than replacing the seat.

• LangChain (2026-07-21) shows the abstraction can hide a 100% price increase in plain sight: it collapsed seven LangSmith meters into two normalized units (1 LCU = $1.50, 1 LSU = $1.00) and stopped publishing dollar-per-1,000-trace rates entirely. Because the sandbox rates convert back to exactly their old dollar values, the conversion factors are verifiable — and by the same arithmetic a base trace went from $2.50 to $5.00 per 1,000. The unit change was the price change.

For buyers

A credit hides the dollar-to-compute conversion, and that rate can change unilaterally — Cursor's 2025 credit-pool switch was opaque enough to force a public refund. Before committing, ask for the dollar value of a credit per feature and whether that ratio is contractually fixed. Note where the credit sits: in most plans it rides on top of a seat (you pay for both), but a handful of vendors (Qodo's 2026 switch to a pure team-wide pool) now let the credit replace the seat entirely.

For vendors

A credit economy needs a metering layer that converts every feature's real cost into credits, a balance/top-up system, and clear in-product disclosure of burn rate — plus the governance to change conversion rates without a trust rupture. Credits remain the natural meter for the usage layer; for the base, seats still dominate the corpus (175 vs 123), so most vendors stack credits on a seat rather than replace it.

Outlook — what to watch

Credits will keep spreading on the usage layer as vendors ship more heterogeneous features (video, voice, agents) under one plan — where vendors restructured usage in 2026 (Qodo, Dust, Zapier), the credit or task pool was the destination. But the corpus-wide lead stays with seats, and that gap has widened, not closed: the thesis weakened from 'most common unit' to 'most common consumption unit.' The status flips back toward 'holds/sharpens' only if seat-anchored vendors begin dissolving the seat into the pool the way Qodo did; it weakens further if more vendors reprice without credits at all (Shortwave, Creatify, lemlist all did in 2026).

FAQ

Related trends

• Unit proliferation at the infra layer — with the corpus vocabulary now saturating and consolidation a live countermove
• From subscriptions to usage — and back to hybrid, sometimes within six weeks
• Token prices fall as access tiers rise