OpenAI Is Gaining on Anthropic With Business Users

OpenAI is closing a gap it never expected to be trailing in. According to Ramp's AI Index — which tracks real payment transactions across more than 70,000 U.S. businesses rather than relying on surveys — Anthropic held roughly 44% business adoption in July 2026 to OpenAI's 40%, but OpenAI's quarter-over-quarter growth rate of 82 now outpaces Anthropic's 76, marking the first period in months where the ChatGPT maker has had measurably stronger momentum. The reasons behind that reversal are more instructive than the headline numbers.

What's Converging

The broader story here is how fast enterprise AI allegiances shift. Ramp has been publishing its AI Index — transaction-based, not survey-based — for over a year, and the pattern has been consistent: no vendor holds a durable lead. Anthropic sat at roughly 9% business adoption in May 2025. By April 2026 it had climbed to 34.4%, at which point it passed OpenAI's 32.3% for the first time. That's nearly quadrupling in twelve months while OpenAI grew adoption by only 0.3% over the same period, according to Ramp's May 2026 report. Ramp's lead economist noted at the time that enterprise AI has no parallel for how quickly newcomers can disrupt market leaders within months — with essentially none of the vendor stickiness that has characterized prior enterprise software cycles.

That May inflection received significant attention across analysts and practitioners. A LinkedIn analysis from late May framed the shift specifically as a coding-and-developer-workflow story rather than a general-purpose productivity shift — Claude was winning where developers spend most of their time and budget, not in the chat interfaces that dominate consumer attention. Multiple reports at the time treated it as a durable reversal.

But Ramp's economist also flagged three structural headwinds for Anthropic even at the moment of peak share: an incentive structure that nudged users toward premium, higher-token models; a period of service degradation — outages and rate limits in early 2026 that Anthropic subsequently addressed through a compute deal with SpaceX and a usage-limit reset — and the risk that enterprise cost-cutting would route spend away from its more expensive tiers. All three headwinds have now materialized in the data.

The Specific Development

The August 2026 Ramp AI Index data makes two things clear. First, Anthropic's headline lead is real but narrowing. The company held 43.5% adoption among tracked businesses in July versus OpenAI's 39.7% — a margin that is smaller than many expected given Anthropic's trajectory earlier in the year. Second, and more revealing, is the token-level breakdown reported by CryptoBriefing's coverage of the same data: OpenAI's GPT-5.6 Sol accounted for 25% of tokens consumed and 23% of corporate spend in July. Anthropic's competing high-end product, Fable 5, represented just 6% of tokens and 11.4% of spend. Fable 5 costs more per token yet drives substantially less usage — exactly the margin-compressing pattern Ramp had warned about months earlier.

The Fable 5 underperformance has a specific reported cause beyond price: a mandatory 30-day data retention requirement. For enterprises managing sensitive customer data, regulatory obligations, or proprietary internal information, a provider that retains query data for a month is a compliance risk, not a footnote. Multiple sources covering the August data point to this retention policy as a primary driver of enterprise hesitation around Fable 5 — distinct from any question of model quality. The competitive gap here isn't about benchmark scores; it's about data governance terms that engineering and legal teams cannot negotiate away once set.

Our read is that this is the most underreported dimension of the story. Developers comparing GPT-5.6 Sol against Fable 5 are making a procurement decision that frequently surfaces in security review — not just in a capability sandbox. If Anthropic's data retention policy on its premium tier doesn't change, a portion of its addressable enterprise market remains structurally off-limits regardless of model performance. That's a policy constraint, not a product one, and it's a different kind of problem to fix.

The broader spending context cuts in both companies' favor simultaneously. AI adoption among Ramp's tracked businesses rose from 50% in March 2026 to 56% in July — meaning the total addressable spend is still expanding even as the allocation between vendors shifts. Both companies are adding paying customers in absolute terms. The competitive question is which one captures the next tranche of enterprise adoption as it comes online, not whether either is cannibalizing a fixed base.

What's Likely Next

The 30–90 day window is primarily a model-release story. Ramp's lead economist explicitly named GPT-5.6 Sol as "increasingly the choice for developers" — the first time a specific OpenAI model has been cited by name as a driver in the Ramp data. Anthropic's release cadence tends toward longer, more deliberate cycles, which worked in its favor when Claude 3 arrived ahead of any GPT-4 successor, but creates a vulnerability window between major launches. If Anthropic ships a capable successor to Fable 5 — and revises its data retention terms in the process — the momentum numbers could reverse again within a quarter. The Ramp data from May to August has already demonstrated that kind of speed is possible in both directions.

The more durable question is how well the Ramp dataset generalizes. Ramp's customer base skews heavily toward technology companies and the San Francisco Bay Area, and explicitly excludes large enterprises running corporate spend through American Express. That's a meaningful sample bias: the companies most likely to adopt AI early and switch providers freely are overrepresented, while the slower-moving large-enterprise segment — where Anthropic has reportedly been investing in direct sales — is underrepresented. Ramp's AI Index is a reliable leading indicator for developer-facing adoption. It is not a proxy for the full enterprise software market, and any analysis that treats it as one should be read with that caveat firmly in mind.

Sources