OpenRouter Coding Leaderboard — Monthly Analysis

September 2026 snapshot — the first first-party capture since June

Snapshot date: 2026-09-06 · Previous measured board: 2026-06-23 (third-party) · Live leaderboard ↗

The ranking

Top-20 total: 3.79T tokens. OpenRouter publishes no share percentages, so the share column is each app's share of this top-20 total, computed by us — it is not market share and will overstate every row. Fell out of the top 20 since the June board: Descript, GitLawb, Pieces.app, Mira, Roo Code, HammerAI, SillyTavern, GDevelop.

What's interesting (our take)

• The fork chain has inverted, and that is the finding of the month. We have tracked the Cline → Roo → Kilo chain as a consolidation story since April: each fork out-ranking its parent, Kilo at #2 in June with Cline down at #10. It now runs the other way. Cline is #3 at 347B, Kilo Code is #4 at 291B, and Roo Code has fallen out of the top 20 entirely. The original is back above the fork and the middle link is gone. We do not know why from this data alone — but "forks accumulate and the parent decays" was the working model, and it no longer describes the board.
• Claude Code is #2 with 444B, up from #4. It is the only app besides Hermes above 400B, and it now carries more of the top-20 total than Cline and Kilo Code separately. Given that Claude Code has a first-party subscription path most users take, its OpenRouter volume is a floor on real usage rather than a measure of it — which makes #2 on a board that only sees the routed fraction the more striking number.
• A fork nobody was discussing entered at #5. Zazen (Freebuff fork) appears at 201B from outside the top 20, ahead of pi, Codex and OpenClaw. We have no coverage of it and no read on where the volume comes from. A fork arriving at #5 on first appearance is exactly the pattern that produced Kilo, and it is worth watching before it is obvious.
• DeepSeek Harness at #10 is a model-migration signal. A harness named for an open-weights model reaching the top ten says the cheap-tier routing story is now big enough to show up in app rankings, not just in pricing pages. It lines up with the head-to-head we covered in August, and with DeepSeek V4 Flash’s arrival in our cost calculator at roughly a twentieth of the new flagships’ input price.
• OpenClaw fell from #3 to #9. Read this one carefully: the June board was all-category and cumulative, so part of the drop is measurement, not market. But OpenClaw also went seven weeks without shipping before 2.0, and the 2.0 upgrade path broke for a meaningful number of existing users. If the next first-party capture still has it at #9, that stops being a source artifact.
• Cursor is #18 with 30.5B, and that is not the indictment it looks like. Cursor is among the loudest names in the space and ranks eighteenth here — but it routes overwhelmingly first-party, so OpenRouter sees only its overflow. We flag it because the gap between rank and discourse is usually the interesting signal on this board; this is the case where the boring explanation is the right one.
• Our two sources disagree on third and fourth place, and we are not resolving it. CodeSOTA’s 30-day window to 2026-08-24 puts Kilo Code #3 above Cline #4; the first-party board for the window ending 2026-09-06 puts Cline #3 above Kilo #4. Both can be true of their own windows — the crossover would have happened in late August. It is also possible one of them is simply wrong. Either way, treat the Cline/Kilo ordering as contested rather than settled.

What we checked

After three months of source failure, both primary sources came back in the same run. We are not assuming that holds.

What we'll watch next month

• Whether Cline stays above Kilo Code. One capture is not a trend, and our two sources already disagree about it. If the October first-party board still has Cline ahead, the fork-consolidation model we have been using since April needs replacing rather than amending.
• Zazen. A fork at #5 on first appearance either keeps climbing or evaporates, and both outcomes are informative. We will also try to establish what it actually is before next month.
• Whether DeepSeek Harness holds the top ten now that GPT-6 Astra and Fable 5.1 have both launched at $10/$50. If cheap-tier routing keeps growing while two flagships land at the same premium price, that is the clearest evidence yet that the market is splitting by cost tier rather than by capability.
• Whether OpenClaw recovers from #9 after 2.0, and whether the upgrade breakage shows up as a sustained dip or a one-month notch.

Editorial note. The honest summary of this month is that we got our data back and immediately found the thing we had been asserting for four months is no longer true. The fork chain was the cleanest structural story this series had — Cline begets Roo begets Kilo, each fork out-ranking its parent, consolidation running one direction. This board has Cline back above Kilo and Roo gone from the top 20 altogether. We are stating that rather than explaining it away, because a source gap is exactly the interval in which a trend you have stopped measuring can quietly reverse. The comparison is imperfect and we have said where; that is a reason to hold the conclusion loosely, not a reason to keep repeating the old one. Next month is the first clean first-party-to-first-party diff since April, and it is the one that will settle this.

August 2026 — no snapshot this month (source gap)

Run date: 2026-08-01 · Last measured board: 2026-06-23 · Live leaderboard ↗

What we checked

What we learned anyway

• The app-rankings API is real — it is gated, not missing. We had recorded that no public endpoint existed. Probing it directly returns 401 “No cookie auth credentials found”, not a 404. That is a meaningful correction: the data is served, it just needs a logged-in browser session rather than a bearer key. An unsolvable scraping problem turns out to be a solvable authentication one.
• Three consecutive months without a first-party capture. The leaderboard renders from streamed Next.js RSC chunks, so a static fetch gets a shell. This is now a structural limitation of the series, not a bad month.
• Our fallback has gone quiet. CodeSOTA carried the July entry; it has published nothing new in the ~six weeks since. If it is still frozen at 2026-06-23 next month, it has stopped being a usable backstop.
• Model boards are not app boards. The trackers that are updating daily rank models, not the applications routing to them. They cannot answer the questions this series asks — which agent platforms are gaining share — no matter how fresh they are.

What we'll watch next month

• Whether CodeSOTA resumes updating, or needs replacing as the fallback source.
• Whether an authenticated one-off capture is worth wiring up now that the endpoint is known to exist behind a session cookie.
• The two July questions still unanswered and carried forward: whether Hermes's ~45% normalises on a coding-only board, and whether Kilo Code holds #2 over OpenClaw across that ~4pp gap.

July 2026 snapshot

Snapshot date: 2026-07-01 · Previous: 2026-06-01 · Live leaderboard ↗

The ranking (top 10)

*From CodeSOTA (all-category board), data-dated 2026-06-23. Δ is versus our 2026-06-01 coding-only snapshot; some "NEW" rows reflect this source's blended, longer-window board rather than genuinely new coding entrants. Dropped from last month's top-20: LangChain, Zed Editor, Qwen Code, Agent Zero, Studs.gg, Crush, zorai, Ito, OpenSquilla — several are coding tools displaced by companion apps on the blended board, not necessarily by falling volume.

What's interesting (our take)

• Hermes's ~45% is real concentration and a source artifact. Share on the apps board roughly doubled from ~20.8% (our June 1 snapshot) to ~45.3% / 20.57T tokens. Hermes genuinely dominates — reportedly ~669B tokens/day, more than Claude Sonnet 4.6 and Opus 4.8 combined — but this is the all-category, cumulative-window board, not the coding-only weekly view, so the jump overstates the month-over-month move.
• Kilo Code edged past OpenClaw for #2 (12.9% / 5.84T vs 8.9% / 4.04T) — a clean fork-chain win. The Cline→Roo→Kilo lineage keeps pulling volume upward while its ancestors fade: Cline slid to #10 (1.7%) and Roo Code to #15 (0.6%), with Aider/Goose/OpenCode still off-chart at tens of billions per month.
• Companion apps are crowding the coding cohort on the blended board. Five top-20 slots are roleplay/companion apps, not dev tools — ISEKAI ZERO (#8), Janitor AI (#9), Mira/Telegram (#13), HammerAI (#16), SillyTavern (#17). Filtered to coding, the true top-6 is Hermes / Kilo Code / OpenClaw / Claude Code / pi / Cline — largely stable.
• First-party CLIs held mid-pack. Claude Code stayed #4 and ticked up to 6.2% (2.80T); OpenAI's Codex re-entered at #14 (353.8B). Both are steady but neither is closing on the agent-class leaders. CLI and agent tools together still exceed 70% of weekly tokens — programming automation has overtaken chat as OpenRouter's dominant workload.

What we'll watch next month

• Whether Hermes's ~45% share holds on the coding-specific board once the live JS page is fetchable again, or normalizes toward ~25–30% when the cumulative-window / all-category effects wash out.
• OpenClaw vs Kilo Code for #2 — only a ~4pp gap that could flip back; watch OpenClaw's IM-automation token growth against Kilo's multi-IDE / model-routing base.
• Codex's trajectory after re-entering the top-20 — does OpenAI's CLI keep climbing, or fall back off-chart the way Qwen Code did between May and June?

June 2026 snapshot

Snapshot date: 2026-06-01 · Previous: 2026-05-04 · Live leaderboard ↗

The ranking

*Estimated from third-party reports; live page JS-rendered. Ranks 11–20 include Qwen Code, GDevelop, OpenHands, Agent Zero, Studs.gg, Portkey AI, Crush, zorai, and others. Lemonade and Agent Zero dropped from top 10. Total 30-day coding volume: ~1,640B tokens (est.), up ~28% MoM.

Note — daily leaderboard: Kilo Code reclaimed the daily #1 spot in late May by routing 313B tokens in 24 hours, 222B of them through GLM-5 (Z.ai). The monthly rolling figures above reflect the full 30-day window, where Hermes still leads.

What's interesting (our take)

• Hermes takes the monthly crown for the first time. With ~341B tokens and 20.8% share, Hermes Agent moved from #2 to #1 in the 30-day rolling window — up from 217B (16.96%) in May. This is the first month Hermes leads the monthly leaderboard rather than just the daily view. The growth is consistent and broad-based, not a single-day spike.
• OpenClaw's share collapsed 5.8 percentage points — the only top-3 app to lose absolute tokens. Despite the total market growing ~28%, OpenClaw fell from 240B (18.78%) to ~214B (13.0%). That shrinkage in an expanding market means OpenClaw lost users, not just share. It held #1 for just one month (May) before being displaced. Whether this is a model-launch anomaly or structural displacement is June's most important question.
• Kilo Code's GLM-5 arbitrage: model routing as a growth hack. In late May, Kilo Code routed 71% of a 313B-token day through GLM-5 — Zhipu AI's new frontier model at ~$1/M input tokens, roughly 3–10x cheaper than comparable models. This pushed Kilo Code to daily #1 and helped recover the monthly total from a mid-May trough of 149B to ~208B. It's the first time we've seen an app visibly use model-routing economics to inflate leaderboard position. GLM-5 may be auto-selected for background agentic tasks (retries, plan loops), making the spike partially a scaffolding artifact.
• Descript entered the top 5 at ~55B tokens — an AI video and podcast editor that has no traditional code-generation use case. Its appearance in the coding category confirms that OpenRouter's attribution is broader than IDE usage: any app that calls LLM APIs for content generation may be counted here. This is the first vertical-specialist story that isn't a dev-tool: a non-IDE app in the top 5.
• LangChain appeared in the top 10 — the orchestration framework's OpenRouter-attributed usage entering the ranked apps list signals that the infrastructure/SDK layer is now generating enough attributed tokens to compete with end-user apps. This is a category-composition shift worth watching.
• Market expansion decelerating: +28% MoM vs. +56% MoM last month. Still growing, but the hyper-growth phase of OpenClaw's entry appears to have inflated May's numbers. The underlying trend is healthy expansion at 25–30% monthly — well above typical SaaS benchmarks.

What we'll watch next month

• Can Kilo Code sustain GLM-5-driven volume for a full 30-day window, or does it fade when users discover most tokens are going to a cheap background model rather than their chosen frontier model? A sustained monthly #1 would make this a real growth story; a one-week spike would confirm it's routing arbitrage.
• OpenClaw's third month: two consecutive months of absolute-token decline in an expanding market would be the strongest structural signal yet that the coding category is diversifying away from the original platform. Watch whether OpenClaw drops further or stabilizes around 10–12% share.
• Descript and LangChain durability: vertical-specialist and infrastructure entries historically spike once and drop (see: Lemonade in April, Agent Zero in May). If both hold top-10 positions in July, it signals a genuine broadening of the coding category definition on OpenRouter.

May 2026 snapshot

Snapshot date: 2026-05-04 · Previous: 2026-04-28 · Live leaderboard ↗

The ranking

Ranks 11–20: Qwen Code (3.29B), GDevelop (3.13B), OpenHands (2.94B), Studs.gg (2.11B), Portkey AI (1.79B), Deep Agents CLI (1.75B), Crush (1.1B), zorai (0.781B), Ito (0.504B), OpenSquilla (0.388B)

What's interesting (our take)

• OpenClaw debuted at #1 with 240B tokens. The original platform that all major forks descend from reclaimed the top position in a single month — it either wasn't tracked by OpenRouter previously or had a dramatic activation event. At 18.78% share it sits above Kilo Code's previous peak (22.9% in April, but measured against a smaller total market).
• The fork chain is fragmenting. Kilo Code (a Cline-fork) fell from #1 to #3 and lost 28B tokens in absolute terms — the only app in the top 10 to shrink month-over-month. Roo Code and Cline also fell in rank. The forks are being cannibalized by their upstream as OpenClaw's own feature velocity accelerated.
• Total market volume grew ~56% month-over-month (estimated ~821B → ~1,278B total tokens), making this expansion, not zero-sum competition. Nearly every app gained absolute tokens; Kilo Code's absolute shrink is therefore notable — it lost users while the category boomed.
• pi surged from #7 to #5, more than doubling its tokens (13.9B → 37.8B). pi is primarily a conversational companion product, not a dev-focused agent. Its rise into coding territory likely reflects developers using it for pair-programming discussion and architecture review rather than direct code generation — a use pattern worth watching.
• Lemonade reversed sharply, falling from #4 to #7 despite absolute token growth. Last month's vertical-specialist story (Roblox-focused coding) appears to have been a timing coincidence rather than durable adoption. Vertical specialists may face stickier competition from general-purpose agents that can be prompted into any vertical.
• Long-tail expansion: 8 new entrants in positions 11–20 including Qwen Code, Zed Editor, Studs.gg, and Portkey AI. Top-3 concentration dropped from ~55% to ~48% of total volume. The category is broadening even as the top two positions consolidated.

What we'll watch next month

• Can Kilo Code arrest its absolute token decline, or will it continue ceding ground to OpenClaw? A second month of shrinkage would be a structural signal, not noise.
• Will pi's coding-category surge continue? If it crosses 5% share it will be the clearest signal yet that conversational agents are encroaching on dev-tool territory.
• Can Qwen Code or Zed Editor break into the top 5? Both entered the top 20 this month — Qwen Code especially benefits from its open-weight model base and lower per-token cost on OpenRouter.
• Does OpenClaw hold #1, or was this entry spike a model-launch artifact? The May snapshot will tell us if the 240B total is a new floor or a one-time activation event.

April 2026 snapshot (baseline)

Snapshot date: 2026-04-28 · Inaugural snapshot — no previous data for comparison · Live leaderboard ↗

The ranking

Context

• Kilo Code led at 22.9% share with 188B tokens — a commanding position driven by its aggressive fork-chain innovation on top of Cline and Claude Code's base. The top-3 (Kilo, Hermes, Claude Code) held ~55% of total coding volume.
• Lemonade cracked the top 5 as a vertical specialist — the Roblox-focused coding agent entered the top 4 despite a narrow use-case focus. An early signal that niche agents can compete on OpenRouter volume.
• Fork chain clearly dominant — Kilo Code (#1), Cline (#5), and Roo Code (#6) collectively held ~27% share, demonstrating the Cline→Roo→Kilo lineage's reach.
• Hermes Agent strong at #2 — 178B tokens across coding tasks, confirming Hermes's cross-functional usage beyond just scheduling and assistant work.

What we watched for May

• Whether Lemonade's vertical-specialist position would hold or be a one-month spike.
• Whether any new major entrant would challenge Kilo Code's top position.
• Whether total market volume was growing or purely zero-sum competition between existing apps.