Skip to content
PricingSaaS
Sign in
Back
Pricing briefing
CCohere
Cohere
PACKAGING CHANGE

Cohere moved Model Vault to per-GPU pricing

By , Co-founder, PricingSaaS

Published

Cohere replaced Model Vault's per-instance pricing with per-GPU hourly rates.

See before and after

Cohere’s pricing page · Change detected

MODEL VAULTfirst seen Mar 30, 2026

What changed

Cohere retired Model Vault's per-instance Performance Tier pricing — Embed, Rerank and Parse plans billed at $3 to $7 an hour per instance under its prior pricing page, before the Nov 4, 2026 update — for a per-GPU on-demand rate card. The new table prices by GPU type instead of model tier, from L4 up to GB200, at $1 to $12.50 an hour, per Cohere's pricing page as of Nov 4, 2026, with promotional pricing through March 2027.

It's Cohere's fourth pricing change in the last 90 days, per Pulse, and its first full pricing-page restructure since Dec 15, 2025, when it recast the page around enterprise custom pricing instead of per-token costs, according to Pulse's tracked history.

Why it matters

For a product leader watching AI infrastructure pricing, this is a shift from coarse instance-tier billing to usage-shaped, GPU-metered billing. Model Vault customers who outgrew the old Performance Tier buckets — fixed Small or Medium sizes per plan — now pay per GPU-hour instead, which rewards workloads that can run on a cheaper chip like the L4 rather than defaulting to the biggest GPU available.

The promotional pricing through March 2027 is doing real work here: it drops the effective entry price well below the old per-instance floor, which reads as a push to get workloads onto the new metered system before the promo window closes and rates normalize. Any team benchmarking inference costs against Model Vault should model both the promotional rate and the eventual on-demand rate, not just the one quoted today.

Watch whether Cohere lets the promotional pricing lapse quietly in March 2027 or extends it — a quiet extension would suggest the per-GPU model isn't pulling the volume Cohere wants yet.

In context

Cohere has restructured its pricing page twice before: on Dec 15, 2025 it moved from per-token pricing to an enterprise custom-pricing model [[cohere/2025Q3]], and in mid-2024 it regrouped a Free/Production split into product categories [[cohere/2024Q2]]. This is Cohere's fourth pricing move in the last 90 days, just nine days after its last price cut [[cohere/2026W44]], well above its 66-day average gap between changes, per Pulse. Full restructures aren't rare across the market Pulse tracks: 325 changes at 275 companies touched one in the last 180 days, 12.1% of 2,686 tracked changes, with Vercel [[vercel]] and Lyzr [[lyzr]] doing it most often. What the table swap understates is the metering shift: Model Vault moves from a handful of fixed per-instance tiers to continuous per-GPU billing, the same usage-based direction Microsoft [[microsoft]] and Anthropic [[anthropic]] have taken recently. The promo through March 2027 is the real lever: a time-limited discount meant to pull usage onto the new card before the old per-instance pricing is retired for good.

Common questions

Does the new pricing apply to customers already on the old Performance Tier plans?

The change is a full table swap, not a partial rollout, so the safest assumption is the per-GPU rate card now applies account-wide — confirm directly with Cohere whether any existing per-instance commitments are grandfathered.

How do we estimate our monthly spend under the new per-GPU rates?

Multiply the hourly rate for each GPU tier you use by your expected monthly GPU-hours, separately for the on-demand and promotional columns — Cohere's pricing page breaks out both for every GPU type as of Nov 4, 2026, so the sizing math runs per GPU tier rather than per seat.

Should we expect more Cohere pricing changes soon?

Cohere has made four pricing changes in the last 90 days, per Pulse, well above its 66-day average gap between changes — a cadence that suggests Model Vault pricing could keep moving, so it's worth checking back before locking in multi-month compute budgets.

Sources