Anthropic’s share of tokens flowing through OpenRouter jumped after the Claude 5.5 family shipped, according to the platform’s own model analytics and its October 2026 data report. Viral posts cite a 3.6x rise in a short-window Anthropic average; the durable signal is on OpenRouter’s model compare pages, where Claude Opus 5.5 shows on the order of 29.4 trillion tokens in 30 days versus 2.17 trillion for Opus 5 — a migration, not a rounding error.
TL;DR
| Source | What it shows |
|---|---|
| OpenRouter October report | 790T tokens in October, 37% above September; Claude 5.5 launches named as drivers |
| Opus compare page | 29.4T (5.5) vs 2.17T (5) tokens, 30-day activity |
| Social 3.6x claim | 7-day Anthropic average after launch; not a full-market statistic |
| Builder takeaway | Haiku 5.5 pricing pushes cheap agent routes; revisit defaults |
What OpenRouter measures
OpenRouter sits between apps and many providers. Its public stats count prompt plus completion tokens billed through its API (with exclusions noted in footers). That makes it a useful routing thermometer, not a census of all inference.
The October 2026 newsletter says volume hit 790 trillion tokens, up 37% from September and roughly 25x from January. It names Haiku 5.5 and the wider Claude 5.5 refresh, alongside DeepSeek V4.1 Flash, as major contributors. Anthropic spend growth of 76% in one month appears in the same report’s lab growth section.
Open-weight share dipped briefly when Claude 5.5 and GPT-6.1 Sol launched, then climbed back above 70% paid tokens within weeks — competition remains real.
Anthropic Claude 5.5 token surge on OpenRouter shown as rising API traffic
Opus migration in the numbers
The Claude Opus 5 vs 5.5 compare page is the clearest public cut:
| Metric | Opus 5 | Opus 5.5 |
|---|---|---|
| 30-day tokens | ~2.17T | ~29.4T |
| Input list price | $5/M | $4/M |
| Output list price | $25/M | $20/M |
| Context | 1M | 1M |
Price cuts plus capability gains explain routing, not hype alone. Claude Code alone shows 14.3T tokens on the Opus 5.5 model page app leaderboard — agentic traffic dominates.
The 3.6x social claim versus platform data
Posts from accounts such as @peterj_walker describe a 3.6x increase in a 7-day average Anthropic token volume after Opus 5.5, Sonnet 5.5, and Haiku 5.5 arrived. That metric is short-window and lab-aggregated, useful as a pulse, easy to over-interpret as “the whole market flipped.”
Compare instead:
- Version lag newsletter shows how fast users leave older Claude versions once a 5.5 tier ships — Haiku 5.5 hit 73% of Haiku-family traffic within a week in mid-October data.
- Opus token ratio on compare pages shows order-of-magnitude model-id migration, not a gentle bump.
We label both: social claim, OpenRouter observed routing.
Haiku 5.5 and the cost ladder
Builders repeating on X that Haiku 5.5 nears Sonnet for 10–15% of the price mirror what OpenRouter’s October narrative already hinted: effective price per token keeps falling even as total volume rises.
For agent loops that fire hundreds of small calls, that tier shift matters more than leaderboard scores. Our Claude Code token guides cover how to measure spend; OpenRouter adds cross-provider routing without rewriting SDKs.
Pair this story with Vercel AI Gateway open-model share for a second gateway view. Gateways disagree on methodology; trends still rhyme.
What changed for OpenAI and open models
October’s report says launches from multiple labs moved the mix, not a single winner-take-all flip. Open-weight paid share around one quarter of spend shows builders still route cheap specialists.
When OpenAI discount rumors circulate, OpenRouter becomes the receipt book — if discounts land, they show up in effective price charts on model pages.
Practical routing checklist
| Step | Action |
|---|---|
| Audit defaults | Search code for hard-coded claude-opus-5 slugs |
| Benchmark Haiku | Run your eval set on Haiku 5.5 before Sonnet |
| Watch cache columns | OpenRouter lists cache read/write pricing per provider |
| Log task type | Agentic versus human keys differ in their classification |
| Keep a second gateway | Avoid single-router lock-in during launch waves |
Limitations readers should remember
OpenRouter excludes some BYOK traffic from rank charts. Stealth models and banned accounts are filtered in retention metrics. Direct enterprise contracts with Anthropic bypass the router entirely.
Token counts also do not equal revenue quality: subsidized promos and free tiers move volume without profit. Anthropic’s own S-1 narrative is separate from daily router stats.
What developers said on X
Early posts emphasized switching costs dropping within Anthropic’s stack — teams staying in one API but down-tiering default models. Others warned against treating router share as market share. We treat both as sentiment, not audited fact.
October volume in context
OpenRouter’s October report is not only an Anthropic story. Total tokens rose 37% month over month to 790 trillion, with open-weight share dipping near 52% the week all major 5.5-tier closed models launched, then recovering above 70% within two weeks as DeepSeek V4.1 Flash traffic returned.
That pattern matters for capacity planning: a single lab launch can temporarily swing router share without changing long-run open-model economics. Builders who hard-code “always cheapest open model” rules should expect whipsaw weeks around flagship releases.
The same report introduces decision models as a fast-growing request class — separate from chat completions. If your product routes through OpenRouter’s decisions API, token charts understate your bill shape. Check both dashboards.
Agentic versus human traffic classification
OpenRouter classifies API keys monthly using signals such as tool-call rate, cache hit rate, and inter-request timing. Keys scoring 0.50 or higher count as agentic. That split explains why Claude Code dominates Opus 5.5 app leaderboards while many human chat apps never appear.
When you interpret token surges, ask whether the spike is new agents or existing agents retargeted to cheaper tiers. Haiku 5.5 adoption can lower spend per task while raising token counts if agents loop more freely.
Retention and version-lag lessons
The version lag report shows Anthropic customers migrate quickly: 80% of Opus traffic hit 5.5 within weeks of launch, while other families linger on older versions longer. Stale model strings in production therefore waste money fast — you pay list price for deprecated SKUs fewer providers optimize.
For multi-model apps, OpenRouter’s ~latest aliases help, but pinned slugs in agent configs override them. Schedule a quarterly slug audit the way you rotate dependencies.
How this connects to Claude Code and IDE defaults
Claude Code’s 14.3T token line on the Opus 5.5 page is a proxy for how much autonomous coding moved to the new stack without developers typing model names manually. Cursor, JetBrains, and CLI harnesses feeding the same router will show similar concentration if their defaults track Anthropic’s latest Opus.
If your org budgets per team, attribute spikes to harness defaults, not mysterious user preference shifts.
Pricing tables worth re-reading
Compare pages list weighted average paid prices below headline list rates because of caching and provider selection. Opus 5.5’s weighted averages sat near $0.76/M input effective in public charts when we checked — far below nominal $4/M input.
That gap is the real builder takeaway: prompts with cache-friendly system instructions change effective price more than negotiating list rates.
Claimed versus verified summary
| Claim | Verified on OpenRouter pages | Unverified |
|---|---|---|
| 3.6x Anthropic 7-day average | Mentioned on X, not on static page | Exact window definition |
| Opus 5.5 dominates Opus traffic | 29.4T vs 2.17T 30-day tokens | Off-router Anthropic direct API |
| October record month | 790T in newsletter | N/A |
What this means if you do not use OpenRouter
Even teams on direct Anthropic contracts should read router stats as early telemetry. When a flagship launch moves millions of developer-hours overnight, provider capacity, rate limits, and queue latency shift for everyone. If Haiku 5.5 absorbs tasks you routed to Sonnet last month, your internal cost model needs new default assumptions.
For multi-cloud shops, mirror the exercise on whatever gateway you use. The exact trillion-token counts will differ; the direction after a coordinated 5.5 release should not surprise you.
Competitor labs and open-weight bounce-back
DeepSeek’s V4.1 Flash refresh appears in the same October report as a volume driver after the closed-model launch week. That is the competitive check on Anthropic’s surge: builders chase price-performance back to open weights quickly unless a closed tier wins on a must-have skill.
Z.ai and MiniMax growth lines in the report show the same release-calendar sensitivity — timing of a drop matters more than brand loyalty in router data.
Governance and procurement notes
Procurement teams sometimes treat gateway dashboards as spend truth. OpenRouter separates billed usage from BYOK traffic in footnotes. When you export CSVs for finance, confirm which column you are using and whether test keys are segregated.
Also document model alias policy: who may pin Opus versus who must default to Haiku for batch jobs. Token surges often trace to one rogue cron job left on Opus after a demo.
FAQ for finance partners
Finance readers often ask whether a token spike equals revenue for the lab. It does not automatically. Router growth can reflect promotional credits, internal dogfood, and subsidy wars. Pair OpenRouter charts with your actual invoice from Anthropic or your reseller.
Related reading
- Vercel AI Gateway open-model volume
- Claude Opus 5.5 SQLite parse research
- Token efficiency with Claude Code
- OpenAI discount claims fact-check
- Neocloud bubble fact-check
- What is MCP
OpenRouter numbers reflect pages accessed November 9, 2026; model activity changes daily.