Is Claude Opus 5.5 Taking Traffic From OpenAI? What Vercel's Spend Data Actually Shows — and Whether You Should Switch

*Short answer: the numbers are real, but they show Anthropic recovering share with Opus 5.5, not taking it from OpenAI. On Vercel's AI Gateway — one developer gateway, measured in dollars not usage — Anthropic's share of spend fell from about 70% in late July to 34% on September 21, the day before Opus 5.5 shipped, then climbed back to 44% by September 24 as Opus 5.5 reached roughly 10% of all gateway spend in three days. OpenAI's rise from 10% to 24% happened over those two months, mostly before the launch. And on OpenRouter, the other big public gateway, Anthropic is about 3% of requests and Opus 5.5 doesn't crack the top ten. Spend dashboards are a better signal than arena screenshots — but only if you read the axis.*

What the data is — and isn't

Vercel AI Gateway is a single API key in front of hundreds of models, and Vercel publishes its usage as a public leaderboard under a CC BY 4.0 license, updated daily, with a JSON export. That's what Rauch's chart animates, and it's what we pulled. Three things to hold in mind before reading any of it: it's one gateway's customers (skewed toward web developers building on Vercel), it measures spend at list price, so a model that gets cheaper shrinks in share even if usage is flat, and Vercel publishes 'percent share, never absolute volumes,' so a lab's share can fall while its actual usage grows. Rauch's own caveat from a week earlier applies: spend is 'for inference of the model across providers… not revenue going directly to' the labs.

The daily numbers, from Vercel's export

Read the middle of that table, not the ends. Anthropic's two-month decline was finished before Opus 5.5 existed. Where did the share go? Rauch says it himself: 'Kimi K3 + DeepSeek took ~half of Anthropic's loss' — the export agrees, with Moonshot and DeepSeek up 14 points against Anthropic's 26-point fall, and OpenAI up 9. In tokens the picture is more lopsided still: on September 24 DeepSeek carried 59% of all gateway tokens, OpenAI 9.9%, Anthropic 8.6%. 'OpenAI now leads in tokens' is true only relative to Anthropic. What Opus 5.5 did was arrest the slide — from 0.9% of spend on launch day to about 10% three days later, the fastest ramp for a new model in the dataset, pulling Anthropic's total back up ten points in two days.

The price trap in 'spend share'

This is the part that makes every one of these charts hard to read in September 2026. Opus 5.5 is 20% cheaper per token than Opus 5 and, by Anthropic's account, about 40% cheaper per task. GPT-6 Sol and Luna halved their prices the same afternoon. When both labs cut prices, identical usage shows up as less spend — and Vercel's own September index documented exactly this effect when developers moved from Fable 5 to the cheaper Opus 5 in August. So a lab can win users and lose spend share in the same week. Anthropic's token share on the gateway (33% → 8.6%) and request share (15% → 6.5%) also fell over the two months, but those denominators were flooded by DeepSeek's tokens and TypeSafe's Jev model — which alone is 27% of requests — so 'share fell' doesn't tell you 'usage fell.'

The second gateway tells a different story

OpenRouter's rankings, with data through September 24, are a useful check because they measure tokens and requests rather than dollars. There, the week of September 14 had DeepSeek at 25% of text requests, Google 19%, OpenAI 17%, and Anthropic 2.7% — and Opus 5.5 appears in none of the top-ten token tables for the day, month, or trending views. It does sit #1 on OpenRouter's Artificial Analysis Intelligence Index tab at 57.6, but that's a benchmark, not usage. The gap between the two gateways is the real finding: Vercel's customers are the developers who pay for frontier quality; OpenRouter's are the ones optimizing cost, and Anthropic is a premium brand in a market where the volume is moving to cheap open-weight models. Opus 5.5 at $4/$20 is Anthropic's attempt to be in both rooms.

What the developers posting are actually saying

Sentiment is running well ahead of the data. The most-shared take — 'the best model I've ever used… After switching completely to Codex, I now have to go back to Anthropic' (@kimmonismus) — is an opinion, not a dashboard. Several of the viral head-to-heads people cite as Opus wins aren't even Opus tests: the liquid-glass UI comparison pitted Fable 5.1 against GPT-6 Astra, and the 500-hour Zelda recreation was built with Astra, Fable 5.1, and Opus 5.5 together. The independent benchmarks do favor Opus 5.5 — #1 on the Artificial Analysis Intelligence and Coding Agent indexes and the Vals Index at launch — so the enthusiasm has a basis. The claim that it's already draining OpenAI doesn't.

Should you switch from ChatGPT to Claude?

• If you pay $20 for an assistant: this is the strongest month Claude Pro has had. Opus 5.5 — the top-scoring model in independent testing — is the everyday model on the $20 plan, while ChatGPT Plus still answers in Chat with GPT-5.6 and puts GPT-6 Sol and Luna only in Work and Codex. If the model is what you're paying for, switch or add Claude Pro now. If you're paying for breadth — images, voice, the most polished consumer agent — ChatGPT is still the all-rounder. Our updated ChatGPT vs Claude comparison makes that call dimension by dimension.
• If you build on the API: don't switch on a share chart. Run your own evals on Opus 5.5 at $4/$20 against GPT-6 Sol at $2/$10 — the Artificial Analysis coding index puts them at 66 and 57, at $13.04 and $2.99 per task. For agentic coding the gap is worth the price; for high-volume extraction it isn't.
• If you're waiting for DevDay: it's September 29 at Fort Mason, keynote 10 a.m. PT. OpenAI already answered Opus 5.5 once, ninety minutes after it launched, with the Sol and Luna price cuts and a banked usage reset; Astra stays at $10/$50. Four more days of waiting costs you nothing on a monthly plan.
• Capacity: we checked status.claude.com — one major incident on September 22 was sixteen hours before the launch, and there have been zero since. Anthropic raised five-hour limits on Pro and Max with the release. No documented Opus 5.5 outages or limit complaints as of today.

Our read

Rauch's chart is good data telling a more interesting story than the one it went viral for. Anthropic spent August losing developer spend to cheaper models — Kimi, DeepSeek, and OpenAI's discounted GPT-5.6 tier — and Opus 5.5 is its answer: a Fable-class model at a Sol-adjacent price that, in three days, became the single biggest line item on a major gateway. That's a comeback, not a conquest, and OpenAI's share was rising before and after. The switch decision for a subscriber is easier than the one for a developer: Claude Pro is the best $20 in AI this month on model quality alone, and the DevDay counterpunch is four days out. Watch the same export next week — if Anthropic holds above 40% while prices stay where they are, the recovery is real; if it drifts back toward 34%, Opus 5.5 was a launch spike. We'll update this page either way.