AT&T now routes 40 percent of its employees’ AI requests to open-weight models instead of proprietary offerings from OpenAI or Anthropic, according to AT&T vice president Mark Austin, with an internal target of pushing that share to 60 or 70 percent within a few years. The company processes roughly 45 billion tokens a day across its internal AI systems, and Austin says open-source routing has already cut coding-related AI costs by 56 percent, for what the company measured as only a 2 percent drop in output quality. It’s a concrete, quantified example of a shift the New York Times recently framed in stark terms: corporate America is getting hooked on open-source AI, and that shift poses a genuine “good enough” threat to the commercial position of frontier labs like Anthropic and OpenAI.
Why “good enough” is the phrase that matters
The threat frontier labs face here isn’t that open-weight models have caught up to the absolute ceiling of what Claude Opus 5 or GPT-5.6 Sol can do on the hardest, most demanding tasks. It’s that for a large and rapidly growing share of everyday enterprise workloads, routine coding tasks, internal documentation, first-pass customer support drafts, an open-weight model running at a fraction of the cost delivers output indistinguishable enough from a frontier model’s output that the price difference stops being justifiable. AT&T’s own numbers make that calculation concrete: a 56 percent cost reduction against only a 2 percent quality decline is, for the overwhelming majority of business use cases, a straightforwardly rational trade to make, and one that compounds dramatically at AT&T’s actual usage volume of 45 billion tokens daily.
This dynamic mirrors a pattern our own explainer on small versus large AI model tradeoffs covers directly: efficient, purpose-tuned models increasingly close the practical gap with larger, more expensive ones for well-defined, high-volume tasks, even when they remain behind on the hardest, most open-ended problems. What’s new here isn’t the underlying technical trend, it’s a Fortune 100 company putting a specific, verifiable percentage on how far that substitution has already gone inside a single organization’s real production workload.
Why this specifically threatens Anthropic and OpenAI’s business model
Anthropic and OpenAI have both built substantial parts of their revenue on enterprise customers paying premium API rates for access to frontier-tier models. Anthropic in particular has leaned into an enterprise-first strategy, with Menlo Ventures reporting the company holds 42 percent market share in coding applications compared to OpenAI’s 21 percent, and 32 percent of broader enterprise AI usage against OpenAI’s 25 percent. That enterprise concentration is exactly the customer base an internal AT&T-style shift toward open-weight routing puts pressure on directly: the more that large enterprise customers build internal infrastructure to intelligently route routine requests to cheaper open models while reserving frontier models only for tasks that genuinely require their extra capability, the smaller the addressable spend frontier labs can count on from their highest-value customer segment.
That pressure compounds with the broader trend our coverage of DeepSeek’s training efficiency and Meta’s Muse Glimmer release have both tracked this year: the quality gap between leading open-weight models and closed frontier models has been narrowing steadily, even as open models’ cost advantage has stayed roughly constant or widened. AT&T’s routing strategy isn’t a one-off cost-cutting experiment, it’s the kind of infrastructure investment that becomes more valuable, not less, as that quality gap keeps closing.
Why frontier labs aren’t standing still
Both Anthropic and OpenAI have visible countermeasures already in motion. Anthropic’s own revenue growth, reportedly exceeding a 30 billion dollar annualized run rate with year-over-year growth above 1,400 percent according to recent reporting, suggests that even with pressure from open-weight substitution at the margins, aggregate demand for frontier-tier capability is still climbing fast enough to support that growth, at least for now. Both major labs have also leaned harder into tiered pricing and cheaper, faster model variants, exactly the kind of response that acknowledges open-weight competition without conceding the premium end of the market. OpenAI’s recent Ultrafast tier, offering dramatically faster inference at a premium price, is one example of frontier labs trying to justify continued price premiums through genuinely differentiated capability, speed in that specific case, rather than competing purely on raw output quality where open models are closing the gap fastest.
What this means for how enterprises should actually think about AI spend
AT&T’s approach offers a genuinely useful template for other large organizations rethinking their own AI cost structure: rather than treating “which AI vendor should we use” as a single, company-wide decision, route different task categories to different models based on actual measured quality tolerance for that specific task, and revisit that routing regularly as both open and closed model quality keeps shifting. A 2 percent quality drop might be entirely acceptable for internal documentation or routine code review, and completely unacceptable for a customer-facing legal or financial output where an error carries real consequences. The value in AT&T’s disclosed numbers isn’t a blanket argument that open models are now equivalent to frontier ones, it’s a concrete demonstration that intelligent, task-specific routing between the two, rather than an all-or-nothing vendor choice, is where real, substantial cost savings are actually available right now.
Whether frontier labs successfully defend their premium positioning against this kind of substitution, or whether the AT&T pattern becomes the norm across large enterprises over the next several years, is likely to be one of the more consequential open questions shaping AI industry economics through the rest of this decade, well beyond the specific companies and models involved in this particular story.
See AI Weekly’s coverage, drawing on the original New York Times reporting, for additional detail.
