Open-weight models took the tokens. The money barely moved.

On 2026-09-28 the Financial Times published an analysis arguing that US corporate AI spending is shifting from closed frontier models to cheaper open-weight models. Its headline framing combined three separately sourced data points that the FT presentation bundled together:

1. Open-weight models ran 56% of all tokens on Vercel AI Gateway in August 2026 — up from 7% in December 2025.
2. Open-weight models now run about 40% of AT&T's AI workloads.
3. AT&T reports roughly 56% lower inference cost; Coinbase reports roughly 50% lower.

INDEPENDENTLY VERIFIED (source traced to publisher): Vercel published these numbers itself, before the FT article. Vercel's AI Gateway Production Index, September 2026 (published 2026-09-17, data through August 2026, authors Amelia Charles and Eric Dodds) states: "Open-weight models ran the majority of gateway tokens for the first time, up from 7% in December to 56% in August." Vercel also disclosed that open-weight models accounted for only 14% of estimated spend in the same month, that Anthropic took 64 cents of every dollar, and that the average price per token fell 23.2% in August, a third straight monthly decline. The New Stack independently reproduced and contextualised these figures on 2026-09-18.

COMPANY CLAIM (AT&T): AT&T's Chief Data and AI Officer Andy Markus wrote in an official AT&T blog on 2026-07-23 that AT&T averages 45 billion tokens per day, has built an internal AI Gateway with cache-aware routing that "can even change models partway during multi-turn sessions", and is "reducing AI costs as much as 90%, already saving millions". Separately, VP Mark Austin (who oversees employee-facing AI) told The Information on 2026-08-20 that LiteLLM-based model routers cut coding and advanced-task costs by up to 56% with about 2% quality degradation, and that targeted applications moved to open models cut costs by 80–90%.

COMPANY CLAIM (Coinbase): On/around 2026-06-27 CEO Brian Armstrong publicly stated Coinbase had reduced its internal AI spending by nearly half. Three levers were named: cheaper default models (Zhipu's GLM 5.2 and Moonshot's Kimi K2.7 Code), automated routing, and caching (cache hit rate on Coinbase's internal LibreChat instance raised from 5% to 60%). Armstrong added that 91% of engineers were never hitting their prior usage caps, so defaults — not rationing — were the lever.

FACT (Vercel's own published data, through August 2026):

• Open-weight models crossed the majority threshold on one gateway for the first time: 7% of tokens (Dec 2025) → 13% (April) → 36% (July) → 56% (August).
• The money did not move with the tokens. Open-weight = 56% of tokens but 14% of estimated spend. Anthropic held 64% of spend, and has not fallen below 61% of spend in any month since December 2025.
• Within-lab substitution dominated cross-lab switching. Anthropic's most capable model "Fable 5" fell from 13.2% of gateway spend in July to 4.9% in August, while "Opus 5" rose to 22.5%. Nine in ten teams that ran Fable cut usage; most of the displaced volume moved down Anthropic's own lineup at roughly half the price.
• Google's share of gateway token volume fell from 30% to 5%, with Gemini 3 Flash accounting for 22 of the 25 percentage points lost; over three-quarters of that volume went to other labs.

COMPANY CLAIM (AT&T and Coinbase): Large enterprises claim that routing — not capability upgrades alone — is now the primary cost lever on AI workloads, with savings between 40% and 90% depending on how the claim is scoped.

INTERPRETATION: The change is not "open models won". It is that the unit of model selection moved from the application to the gateway, and that made cost a per-request decision variable rather than an annual procurement decision. Token share moved from 7% to 56% in eight months at one gateway; spend share moved from a minor share to 14%. The economically decisive number is the price gap per token, not the token share.

Both architectures described here are the same idea: a proxy sits between application code and model providers and makes the per-request decision.

1. Preprocessing / classification. The gateway inspects the prompt (task type, complexity estimate, cache status) before dispatch. Coinbase's version "matches task type, cache status, and per-token pricing to select the most cost-effective model capable of completing the job". AT&T's weighs "speed and cost with the expected quality of the output".
2. Price/quality arbitration. Cheaper or open-weight models absorb the long tail; frontier models are reserved for architectural planning and novel reasoning. Coinbase's stated principle: "Humans shouldn't be choosing models — AI can automate this task."
3. Cache awareness — the underappreciated lever. AT&T's gateway is explicitly cache-aware, because cached tokens cost roughly one-tenth of uncached ones. Austin's counterintuitive finding, as relayed by Fierce Network: on long agentic coding sessions, staying on a premium model can be cheaper, because the accumulated conversation context is already cached and switching models mid-session discards it. Coinbase improved cache hit rate on LibreChat from 5% to 60%.
4. Mid-session switching. AT&T's gateway can change models partway through a multi-turn session, with the trade-off above.
5. Vendor-neutral surface. Both Vercel AI Gateway and OpenRouter expose many providers behind one interface with unified usage and cost accounting — which is what makes aggregate measurement possible at all.
6. Price mix feedback. As cheap tokens displace expensive ones, blended price per token falls, which funds more inference from the same budget, which pushes share further. Vercel: open-weight adoption "helped push the average price per token across the gateway down 23.2% in August."

• This is a cost-architecture signal, not an open-source-values signal. Both AT&T and Coinbase describe the switch as price-driven. Nothing in either company's stated reasoning is ideological.
• The token/spend divergence is the most important qualification in the whole story. 56% of tokens, 14% of spend. Reporting "open models took over production AI" from a token-share figure inverts the revenue picture and understates how intact the frontier labs' economics remain.
• It changes where the cost conversation sits. If a gateway is choosing the model, then cost control becomes a platform discipline (routing policy, cache policy, evaluation gates), not a procurement discipline (annual vendor negotiation).
• Consulting relevance: "Use a cheaper model" is not an insight. "Deploy a cache-aware gateway with a quality floor and a spend-downshift policy" is an implementable, budgetable programme — and both AT&T and Coinbase published enough detail for it to be copied.
• It undercuts "closed models dominate production inference" only partially. They dominate spend and premium tiers. The shift is real but it is happening in the cheap tail of the distribution, where closed-model premium pricing matters least.

FACT: Organisations can now hold token volume roughly flat-or-growing while holding AI spend flat or falling, using routing rather than access restriction.

INDEPENDENT EVIDENCE: AT&T reports ≈45B tokens/day (≈5.6x year-on-year) with roughly flat AI spend. Coinbase reports roughly halved AI spend while internal token usage continued to rise.

INTERPRETATION: Three things became newly feasible that were not, a year ago:

• Budget-linked capacity planning. Inference capacity can be planned as "spend envelope + routing policy" instead of "seat count × model".
• Graceful downgrade instead of access denial. Both companies replaced rationing with cheaper defaults. Coinbase's data point is that 91% of engineers were never near their caps — meaning most caps were pure friction.
• Portable routing policy. Because routing decisions are now made in one place, changing model vendors (or routing model traffic to a self-hosted cluster) becomes a config change rather than a codebase rewrite. Vercel's CEO made exactly this argument: enterprise adoption of model-agnostic patterns is still early because "harnesses, CLIs, IDEs, SDKs, etc need to be adapted to be model agnostic."

• What is AT&T's actual denominator for "40%" — request count, token share, or cost share? Do request share and token share diverge at AT&T as sharply as they do on Vercel?
• On what audit basis is Coinbase's ~50% AI-spend reduction stated, and would it survive excluding one-off migration costs?
• Why does open-weight share of spend (14%) lag tokens (56%) so persistently, and when does that gap close?
• Does the 7% → 56% curve continue, or has Vercel's customer base saturated its cheap tail?
• What is the actual quality floor AT&T enforces before downshifting a task, and how is it re-validated as models change?
• Are enterprises that are not already on a routing gateway following at all, or is this a gateway-market effect rather than an industry effect?
• Does AT&T's "rethinking investments in OpenAI and Anthropic" become a renegotiation, or a multi-year migration?

• Vercel publishes its next AI Gateway Production Index in October 2026 covering September data; watch whether open-weight share of spend (14%) starts converging toward its share of tokens (56%). That convergence, not the token share, is the real signal.
• AT&T's stated 60–70% open-model target inside roughly a year gives a checkable milestone. Its next earnings cycle is the first place to expect a filed number rather than a press quote.
• Coinbase faces the sharpest follow-up pressure, because it made the most specific cost claim and made it on Chinese open-weight weights. Expect a security/provenance reckoning that forces either mitigation detail or a partial reversal.
• Open-model capability convergence continues or stalls. At roughly a four-to-five-month reported lag, open weights are close enough to compete on routine workloads and far enough to fail on hard ones. That gap governs how far routing can go.
• Frontier labs respond with cheaper tiers rather than better top-end claims — Opus 5 is the template. Expect continued within-lab price-tier cannibalisation.
• Governance catches up. Open-weight share of enterprise traffic is now large enough that procurement, residency and provenance standards will be written this cycle, not the next.

The FT headline — "Corporate America embraces cheaper 'open' AI models" — is directionally right and analytically loose in three specific ways, and all three matter.

First, this was never an earnings-call story. Vercel does not have earnings calls; it published a blog index. Coinbase's number came from a CEO social post. AT&T's number comes from executive press interviews. The FT's genuine transcript-based finding — a sixfold year-over-year rise in open-model mentions on US earnings calls — is a real signal about discourse, and it is a different claim from the three numbers used to illustrate it.

Second, tokens are not money. Open-weight models carried 56% of tokens and 14% of dollars on the same gateway in the same month, with Anthropic holding 64% of spend. Both facts are true simultaneously. Any story that reports the first and skips the second is telling its audience the wrong thing about who is winning.

Third, the interesting engineering is not model choice — it is cache state. AT&T's cache-aware gateway, the finding that staying on a premium model is sometimes cheaper on long agentic sessions, Coinbase's 5% → 60% cache hit rate: these are the mechanisms that actually produced the savings. "Use a cheaper model" is the slogan. Prompt-cache economics and sticky routing policy are the substance.

The genuinely notable structural change is smaller than the headline and more durable: model selection moved out of application code and into a gateway, and once it is there, it becomes a per-request financial decision that a platform team can tune continuously. That is an infrastructure fact, not a values fact — which is exactly why it is spreading.