LLM Rankings Have Forked: Anthropic Wins the Taste Test, Workhorses Win the Token Bill

There are now two LLM leaderboards, and pretending they answer the same question is how teams end up with expensive architecture and vague model-selection meetings.

The first leaderboard is the taste test: Arena-style human preference rankings, where strong frontier models win pairwise comparisons because they reason better, write cleaner, recover from ambiguity, and generally make fewer messes in public. The second is the bill: OpenRouter usage, where developers route actual token volume through models that are fast enough, cheap enough, available enough, and good enough to survive production traffic. This week’s rankings make the split unusually clean. Anthropic is still sitting on the prestige layer. The token market is rewarding workhorses.

On Arena Text, the top of the board barely moved. Claude Fable 5 leads at 1508 Elo, followed by Claude Opus 4.6 Thinking at 1504, Claude Opus 4.7 Thinking at 1502, Claude Opus 4.6 at 1499, and Claude Opus 4.7 at 1493. That stability matters. It means the frontier chat leaderboard is not generating a fake daily drama cycle out of statistical dust. The useful signal is not “new king every morning.” It is that Anthropic’s high-end models are still consistently preferred in evaluated, human-facing comparisons.

Web development tells a similar story, but with more interesting movement at the edges. Arena WebDev has Claude Fable 5 at #1 with 1654 Elo, GLM 5.2 Max at #2 with 1595, then Claude Opus 4.7 Thinking at 1566, Claude Opus 4.8 Thinking at 1561, and Claude Opus 4.7 at 1556. That is a very expensive-looking top five. The tail, though, is where substitution starts: Gemini 3.5 Flash moved to #14, GPT-5.5 XHigh in a Codex harness moved to #15, Kimi K2.7 Code rose to #19, and MiMo-V2.5-Pro re-entered the visible top 20 at #20. Muse Spark dropped out.

The leaderboard you quote is not always the leaderboard you should buy from

OpenRouter’s weekly token rankings look like a different market because they are measuring a different market. DeepSeek V4 Flash is #1 with 4.59 trillion weekly tokens, up 2%. MiniMax M3 is #2 with 4.42 trillion, up 22%. MiMo-V2.5 is #3 with 3.81 trillion, up 26%. Hy3 preview is #4 with 3.62 trillion, down 12%. Claude Opus 4.7 is #5 with 2.98 trillion, up 62%.

That top five alone represents more than 19 trillion weekly tokens in the scrape. Those are not vibes. They are routing decisions. Somewhere, large volumes of real applications are deciding that DeepSeek, MiniMax, Xiaomi, Tencent, and Anthropic all belong in the same procurement conversation, even if they do not occupy the same tier in a human preference leaderboard.

This is the part teams should internalize: Arena is closer to a structured evaluation of answer quality under controlled comparison. OpenRouter is closer to a live market meter. It bakes in the parts benchmark tables usually suppress: price, latency, rate limits, context windows, free tiers, provider reliability, default routing inside apps, and whether a model is “good enough” for a particular slice of work. A model does not need to be the best model in the world to be the correct model for 800 billion weekly tokens of summarization, extraction, rewriting, test generation, or agent scratchpad work.

GPT-5.5 moving up four spots to #13 on OpenRouter with 812 billion weekly tokens is a good example of why rank alone is too blunt. Gemini 2.5 Flash Lite fell two spots to #17 while still moving 708 billion tokens. GLM 5.2 entered the top 20 with 591 billion. These are all large enough numbers to matter, but they probably reflect different jobs. Some models are being used as premium reasoning layers. Others are being used as high-throughput components in pipelines where the user never sees the raw output.

For coding agents, the middle tier is where the fight gets ugly

The WebDev board deserves separate treatment because code is less forgiving than chat. A model can sound polished and still quietly corrupt a repository. It can explain React Server Components beautifully and then delete the test that would have caught its own regression. That is why the Anthropic-heavy top of Arena WebDev is meaningful: coding agents benefit from stronger planning, better uncertainty handling, and the ability to keep a multi-step edit sequence coherent.

But the bottom of the WebDev top 20 is the more strategically interesting layer. Kimi K2.7 Code climbing to #19 and MiMo-V2.5-Pro re-entering at #20 are not proof that everyone should migrate tomorrow. They are proof that the cheaper-code-model bench is no longer decorative. This is where teams will test substitution first: doc updates, small bug fixes, scaffold generation, test expansion, migration scripts, PR review drafts, and “try three approaches in parallel” agent branches. If those models stop embarrassing themselves on narrow tasks, they will earn more surface area.

That creates a practical engineering pattern: the frontier model becomes the reviewer, planner, and escalation path; the cheaper model becomes the worker pool. You do not ask the cheapest model to redesign auth. You do ask it to generate fixtures, summarize traces, propose candidate patches, or run speculative branches that a stronger model later judges. This is not glamorous architecture. It is just the obvious result of treating tokens like compute instead of like magic.

The right answer is routing, not religion

If your internal model policy still says “we use Model X,” it is probably already stale. The better policy is a routing table backed by measurement. Use frontier models for ambiguous planning, final user-visible answers, security-sensitive reasoning, code review, incident analysis, and tasks where one subtle mistake costs more than the token bill. Use cheaper workhorses for repeatable transforms, retrieval compression, classification, metadata extraction, draft generation, batch summarization, and speculative execution.

Then measure the boring things. Track success rate by task class. Track latency at p50 and p95, not just the best demo response. Track retry rate, tool-call failure rate, JSON validity, diff acceptance rate, hallucinated file paths, and human override frequency. If you are using models for coding, measure whether generated tests actually fail before the fix and pass after it. If you are using models for support, measure whether the answer reduced follow-up tickets. “Leaderboard rank” is not an SLO.

The OpenRouter movement also argues for dependency hygiene. When a model jumps quickly in usage, do not treat popularity as proof of durability. It may be quality. It may be price. It may be a free tier. It may be one large app shifting traffic. Pin versions where stability matters, canary migrations, keep rollback paths dull, and record which provider and model produced important outputs. AI infra is now normal infra, which means the fun part ends and the runbook begins.

The editorial read is simple: the LLM market has forked into prestige and throughput. Anthropic is winning the former in the visible Arena rankings, especially at the high end of text and web development. OpenRouter is showing the latter, where DeepSeek V4 Flash, MiniMax M3, MiMo-V2.5, Hy3 preview, and GLM-family models are competing for the giant pile of work that does not require the fanciest answer in the room.

That is good news for builders. It means the winning architecture is not a bet on one vendor. It is a system that knows when to spend money and when not to. The teams that separate “best answer” from “best unit economics” will ship cheaper, faster, and with fewer theological arguments in architecture review. Everyone else will keep screenshotting leaderboards into Slack and calling it strategy.

Sources: LMArena leaderboard, OpenRouter rankings, OpenRouter model rankings API, OpenRouter model catalog API