"OpenRouter State Of Models: Jev, Open-Weights, and Tokenomics" — IndyDevDan
Full transcript: [[2026-10-05-indy-dev-dan-openrouter-state-of-models-jev-open-weights-tokenomics-transcript]] (ElevenLabs speech-to-text; YouTube subtitles were rate-limited, so proper nouns may be misspelled).
Why this is in the vault
This is a Q4-2026 snapshot of the model market, read off OpenRouter's public rankings. It records three things the vault does not yet have in one place:
- A dated usage datapoint. OpenRouter processed about 146T tokens in the week of 2026-09-21, against about 4.5T a year earlier.
- The first public reading of Jev on OpenRouter. Jev was the 12th-ranked model that week, it was growing faster than almost any other model, and OpenRouter has added a new "decisions" model category for it.
- A second Pi-plus-Jev cost comparison. About 17-20% cheaper on a single bug-fix prompt, which extends the benchmark from [[2026-09-28-indy-dev-dan-10-levels-of-jev]].
It also repeats Dan's "own your harness" argument ([[2026-05-10-harness-moat-two-layers-portability]]), now backed by OpenRouter app-share data showing Pi gaining on Claude Code. That argument matters to RDCO's agent-capability bet and to how RDCO picks models.
Episode summary
Commentary built around five trends, presented over a screen share of OpenRouter's rankings, market-share, and top-apps pages, plus his own Pi vs Pi+Jev comparison UI:
- Tokens go up. Usage is close to exponential, so in his view there is no AI bubble. The open question is whether usage keeps climbing fast enough to support Anthropic's rumored ~$2T IPO.
- Who's winning. On OpenRouter, DeepSeek, Google, and OpenAI hold the top three spots almost every week. Workhorse "A-tier" models (Gemini 3.8 Flash, DeepSeek V4.1 Flash, GPT-6 Luna — chart label actually reads GPT-5.6 Luna, see Notable claims) take most of the volume. His back-of-envelope math says Anthropic's direct traffic is larger than all of OpenRouter, so in his telling every lab is winning.
- A new model class: Jev and classifiers. Zero-shot classifiers are back. Jev's uptime and reliability are what is hard to clone. He has built Jev into his Pi coding agent's default setup.
- "Free" tokens. Stealth models like Space Bunny Alpha (rumored to be a new Minimax model) climb the rankings because they cost nothing. If it's free, your prompts are the product. Low price drives adoption.
- Top apps. Pi is gaining on Claude Code in OpenRouter app usage. He predicts Pi will pass Claude Code there, which he reads as engineers moving to own their harness.
The video ends with a plug for his own upcoming "Phase 3 agentic engineering" product (launch expected late November or early December). That is self-promotion, not a third-party sponsor.

Key arguments / segments
Timestamps are approximate. They are estimated from word position, with frame timestamps used where available.
- [~00:35] Trend 1, tokens go up. OpenRouter weekly volume went from 4.5T to 146T in about a year. More tokens is not the same as more value: generating code is the easy part, and burning tokens to feel productive is not progress. He aims for a steady rise in his own compute use rather than spikes.
- [~01:49] The Anthropic IPO framing. OpenRouter mostly excludes Anthropic's direct traffic. The bull case for the ~$2T IPO (Anthropic's projection of $200B annual recurring revenue by end of 2028) needs usage to keep compounding. If it stalls, he expects Anthropic, then OpenAI, to fall. He leans optimistic.
- [~04:49] Trend 2, who's winning. DeepSeek, OpenAI, and Google are the top three almost every week. He rejects the claim that Google is out of the race: Gemini 3.8 Flash is his default in Pi because it has the best balance of performance, speed, and cost. GPT-6 Luna is credited with keeping OpenAI in the rankings.
- [~07:01] Napkin math on Anthropic. He assumes a ~$60B run rate made up entirely of Opus 5.5 tokens. That gives ~65T tokens/day, or ~450T/week, about 3x all of OpenRouter. Even after halving it, the figure is still ~1.5x OpenRouter. His conclusion: OpenRouter mostly shows individuals and small and mid-size businesses, while enterprise spend goes direct to the labs.
- [~08:30-10:10] The distillation loop. Frontier labs lead, Chinese open-weights models make that capability cheap by distilling from frontier models, and then frontier labs get the credit again. For engineers this means falling costs and rising intelligence. His summary: every lab wins.
- [~10:10] Trend 3, Jev. Jev processed about 2.6T tokens, up 257% week over week (transcript figure — the "2.68T / ~200%" variant does not appear verbatim in the transcript or in any captured frame, so treat 2.6T/257% as the sourced number). It ranks 12th even though each call uses far fewer tokens than an LLM call inside an agent. The hard part to clone is a general zero-shot classifier with steady latency and uptime. A small model trained on your own data can beat Jev on one task, but not across many tasks.
- [~11:53] Reliability as differentiator. He cuts to the Claude status page during an outage (Sep 29 2026) to contrast with Jev's flat uptime. He says the leading labs treat reliability as a second or third priority.
- [~12:11] Pi vs Pi+Jev benchmark. Same prompt (described as a relatively simple ~400-500K-input-token task) on Sonnet 5.5, comparing Pi alone vs. Pi+Jev. He states he saved "about five cents" and "about twenty percent" with Jev added, and that he still needs to add a speed comparison to the UI — cost is the only axis shown. No dollar totals, input-token counts, per-call Jev pricing, or specific file names are stated on camera or in the transcript; an earlier version of this note presented invented specifics (exact dollar figures, a 434k/286k token split, "5 Jev calls at $0.002", a changed file named
stock.ts) that do not appear anywhere in the source and have been removed. Treat "~20% savings, n=1" as the actual claim. "Think in ands, not ors": combine models rather than pick one. - [~13:29] OpenRouter "decisions" category. OpenRouter has added a model category for Jev-style decision models. He expects many clones and more "new species" of models (classifiers, decision-makers).
- [~15:06] Trend 4, free tokens. Space Bunny Alpha jumped to #2 by being free. He assumes free providers keep and train on your prompts and would rather pay. The wider lesson: cheap-but-capable models win usage, and OpenRouter's user base is weighted toward price-sensitive users.
- [~18:06] Haiku pricing. Based on the Sonnet launch pricing, he expects Anthropic to keep Haiku's price where it is. He wants it under $1 so it can compete with Gemini 3.8 Flash, but doesn't expect that.
- [~19:03] Trend 5, top apps. Pi is slowly gaining on Claude Code in OpenRouter app usage, with Cline and Kilocode trailing. He predicts (in his December predictions video) that Pi passes Claude Code on OpenRouter. He notes the caveat himself: Claude Code's true usage is far higher once subscriptions are counted. His thesis: own the harness instead of renting it.
- [~20:20] Local hardware. He is buying an M5 Ultra Mac Studio (512GB unified memory) in October to benchmark local open-weights models.
- [~22:01] Wrap-up and method. Start with state-of-the-art models on a new problem. If the problem repeats, move it down to cheaper, faster models. He will publish an agent-updated model-stack tier list. Use Jev-style models to route work to the right model, agent, or workflow.

Note: frame-000829.jpg was extracted near the Pi vs Pi+Jev benchmark segment but turned out to be a black "DON'T" title card from an unrelated moment, not the comparison UI — removed from this note rather than left mis-captioned. The source video was not re-downloaded to re-extract the correct frame; the benchmark numbers in Key arguments above are sourced from the transcript's spoken audio only.
Notable claims
These are claims as stated in the video. None were verified against primary sources unless noted.
- OpenRouter volume: ~146T tokens/week (week of 2026-09-21; the on-screen chart shows a 146T total and data through Sep 28), against ~4.5T about a year earlier, roughly 32x. Internal inconsistency: his closing slide says "~80x, 4.5T to 445T". 445T is his estimate of Anthropic's weekly volume, not OpenRouter's. Do not cite 80x as OpenRouter growth.
- Top models that week (from the on-screen chart): DeepSeek V4.1 Flash 19.6T, GLM 5.3 Flash 16.3T, Space Bunny Alpha 13.9T, Hy4 preview 9.64T, GPT-5.6 Luna 8.53T, Nemotron 3 Ultra (free) 5.67T. He calls the model "GPT-6 Luna" out loud, but the chart label reads GPT-5.6 Luna.
- Jev on OpenRouter: 2.6T tokens, +257% week over week (transcript figure), ranked #12. He says it is the fastest grower apart from MiMo Flash and Space Bunny Alpha.
- Anthropic: rumored ~$2T IPO target, losses of "almost $50B", run rate of ~$60-65B, and a company projection of $200B annual recurring revenue by end of 2028. All come from browser-tab headlines, not filings.
- Napkin math: $60B/yr spent entirely at Opus 5.5 list price gives ~65.2T tokens/day. This is an upper bound under a deliberately unrealistic assumption, which he states himself.
- Pi+Jev savings: "about five cents" and "about twenty percent" cheaper on one ~400-500K-input-token prompt, stated on camera with no dollar totals or token counts shown. That is n=1. The "~20% on every agent" figure is his extrapolation. It is not a benchmark result. (An earlier draft of this note cited specific dollar figures, a token-count split, and a changed filename for this comparison — none of that appears in the source; it has been removed as unsupported.)
- Model naming as of Oct 2026, as he states it: Claude Opus 5.5 and Sonnet 5.5 (with a new Haiku "coming soon"), GPT-6 Sol / Luna / Terra / Astro, Gemini 3.8 Flash, DeepSeek V4.1 Flash.
- Bias flags: (a) He builds and sells Pi-centric training, so the "Pi overtakes Claude Code" framing helps his business. (b) He is openly enthusiastic about Jev but has disclosed elsewhere that he takes no sponsorships ([[2026-09-28-indy-dev-dan-10-levels-of-jev]]). (c) The video ends with a pitch for his own product.
Mapping against Ray Data Co
- Jev-as-gate, a second datapoint. RDCO has the
typesafe:typesafe-aiskill installed and has already filed a security review ([[2026-09-18-security-review-typesafe-skills-plugin]]) and a usage guide ([[2026-09-23-every-jev-usage-guide]]). This video adds cost evidence (unverified specifics beyond "~20% savings, n=1" — see Key arguments) for putting a classifier call inside an agent loop to cut input tokens, rather than only using classifiers as standalone verifiers. A useful RDCO test: put a Jev "is this file relevant?" gate in front of the subagent fan-outs in/process-newsletterand/process-youtube(or the vault-write critics), then measure input tokens before and after. Treat the ~20% as a hypothesis until RDCO measures it. - Model-selection method matches RDCO practice. "State-of-the-art first, move repeated work down to cheaper models" is the same pattern as the founder's standing delegation guidance (Opus for everything by default as of 2026-10-06, with cheaper subagent types such as
fanoutfor bounded per-item work — this rule lives in session working memory, not yet a vault note). No change needed. This confirms the current approach. - OpenRouter undercounts Anthropic. This matters for any phData client conversation or Sanity Check piece that uses OpenRouter share charts as "market share". OpenRouter mostly reflects individuals and small businesses; Dan's own napkin math (see Key arguments, ~07:01) argues enterprise/direct-to-lab spend is several times OpenRouter's visible total. The claim that phData's specific enterprise customers route through Snowflake Cortex and Bedrock rather than OpenRouter is this note's own synthesis, not something Dan states — flagging it as inference rather than sourced fact. Pair this note with [[2026-08-17-stratechery-stripe-openrouter-aggregating-ai]] before citing OpenRouter share charts.
- Reliability as a selection criterion. The Claude outage cut-in is a reminder that RDCO's always-on agent depends on one provider. A fallback chain was discussed in [[2026-06-08-indy-dev-dan-ranked-cloudflare-software-factory-s-tier-tokenomics]]. Low urgency.
- Mapping strength: medium. The Jev in-loop cost evidence is directly actionable. The market-share and IPO commentary is context only.
Related
- [[2026-09-28-indy-dev-dan-10-levels-of-jev]] — previous episode; the Jev framework this video builds on
- [[2026-09-14-indy-dev-dan-agentic-engineering-benchmarks]]
- [[2026-09-21-indy-dev-dan-self-compact-pi-agent-zero-hype-devlog]] — Pi harness work
- [[2026-08-24-indy-dev-dan-intelligence-explosion-harness-engineering]]
- [[2026-06-08-indy-dev-dan-ranked-cloudflare-software-factory-s-tier-tokenomics]] — earlier tokenomics framing
- [[2026-08-17-stratechery-stripe-openrouter-aggregating-ai]] — OpenRouter as an aggregator
- [[2026-08-21-innermost-loop-anthropic-ipo-stripe-openrouter-token-economy]] — Anthropic IPO and token economy
- [[2026-09-15-every-typesafe-jev-vibe-check]]
- [[2026-09-23-every-jev-usage-guide]]
- [[2026-09-27-alphasignal-system-one-models-jev-laya-clm8b]] — Jev clones and the "System One" model class