Short answer (OpenRouter's weekly rankings for September 28 – October 4, 2026): the most-used models on OpenRouter are cheap, high-volume models, and the single largest is anonymous. Space Bunny Alpha, an unnamed stealth model, processed 38.7 trillion tokens, 28.5% of the top 20. The leading named model was DeepSeek V4.1 Flash (25.6T), followed by Z.ai GLM 5.3 Flash, Xiaomi MiMo-V2.6-Flash, Tencent Hy4 preview, and OpenAI GPT-6 Luna. Frontier models such as Claude Opus 5.5 and GPT-6 Astra each had 1–2% of top-20 tokens.
High usage share does not identify the best model for your task. It reflects price, speed, and which apps send the most traffic. Use the rankings to find candidates, then use the price table and the job table below to choose.
Already have an OpenRouter key? Paste it into Teamday and run Kimi K3, GLM-5.3, GLM-5.3 Flash, or DeepSeek V4 Pro as AI employees with files, tools, and recurring missions. Connecting your own AI account lets you use Teamday free for 14 days, then $99/month with unlimited work on that account — or choose the plan and add AI credits. 14-day money-back guarantee.
Connect your OpenRouter key to Teamday →Top OpenRouter Models by Usage: October 2026
Source: OpenRouter's public rankings, fetched October 5, 2026. The weekly view covers the trailing seven days that end with the most recent complete day, so September 28 through October 4, 2026. "Change" compares those seven days with the seven before them. "Share" is each model's share of the top 20 models' tokens (135.4T in total), not of all OpenRouter traffic. Tokens are input plus output.
Prices are per 1M input / output tokens from OpenRouter's per-model endpoint API (/api/v1/models/<id>/endpoints) on October 5, 2026, and each one names its provider: the model maker's own endpoint where it serves the model on OpenRouter, otherwise the provider shown. This guide no longer uses the single price on each OpenRouter model page. That catalog-level price is one provider's quote, and it moves as providers join: on October 5 it showed GLM 5.3 at $0.05 input / $7 output from a third-party route, while Z.ai's own endpoint charged $1.40 / $4.40.
| Rank | Model | OpenRouter ID | Tokens (week) | Share | Change | Input / output per 1M (provider) | Context |
|---|---|---|---|---|---|---|---|
| 1 | Space Bunny Alpha (stealth) | stealth/space-bunny-alpha | 38.7T | 28.5% | +179% | No public price | 1,000,000 |
| 2 | DeepSeek V4.1 Flash | deepseek/deepseek-v4.1-flash | 25.6T | 18.9% | +31% | $0.15 / $0.60 off-peak; $0.30 / $1.20 peak (DeepSeek) | 1,048,576 |
| 3 | GLM 5.3 Flash | z-ai/glm-5.3-flash | 9.73T | 7.2% | −40% | $0.15 / $0.50 (Z.ai) | 1,048,576 |
| 4 | MiMo-V2.6-Flash | xiaomi/mimo-v2.6-flash | 9.63T | 7.1% | +75% | $0.14 / $0.28 (Xiaomi) | 1,048,576 |
| 5 | Tencent Hy4 preview | tencent/hy4-preview | 6.51T | 4.8% | −32% | $0.834 / $2.501; 10% less 16:00–24:00 UTC (Tencent) | 1,048,576 |
| 6 | GPT-6 Luna | openai/gpt-6-luna | 6.18T | 4.6% | +116% | $0.10 / $0.50 (OpenAI) | 1,050,000 |
| 7 | DeepSeek V4 Flash 0731 | deepseek/deepseek-v4-flash-0731 | 5.99T | 4.4% | −23% | $0.06 / $0.18 (DeepInfra) | 1,048,576 |
| 8 | Nemotron 3 Ultra (free) | nvidia/nemotron-3-ultra-550b-a55b:free | 5.65T | 4.2% | 0% | $0 / $0 | 1,000,000 |
| 9 | GPT-5.6 Luna | openai/gpt-5.6-luna | 4.37T | 3.2% | −49% | $0.20 / $1.20 (OpenAI) | 1,050,000 |
| 10 | DeepSeek V4 Flash 0423 | deepseek/deepseek-v4-flash | 3.23T | 2.4% | −4% | $0.09 / $0.18 (DeepInfra) | 1,048,576 |
| 11 | Jev 1.13 | typesafe/jev-1.13 | 3.10T | 2.3% | +20% | $0.042 input, no output charge (TypeSafe) | 32,000 |
| 12 | GLM 5.3 | z-ai/glm-5.3 | 2.94T | 2.2% | +5% | $1.40 / $4.40 (Z.ai) | 1,048,576 |
| 13 | Claude Opus 5.5 | anthropic/claude-opus-5.5 | 2.46T | 1.8% | +128% | $4 / $20 (Anthropic) | 1,000,000 |
| 14 | Gemini 3.8 Flash | google/gemini-3.8-flash | 2.18T | 1.6% | +1% | $0.75 / $3.75, 50% promotional discount (Google AI Studio) | 1,048,576 |
| 15 | Tencent Hy3 | tencent/hy3 | 2.07T | 1.5% | −19% | $0.132 / $0.528; $0.0825 / $0.33 16:00–24:00 UTC (Tencent) | 262,144 |
| 16 | Kimi K3 | moonshotai/kimi-k3 | 1.62T | 1.2% | +16% | $3 / $15 (Moonshot AI) | 1,048,576 |
| 17 | Muse Spark 1.3 Contributor | meta/muse-spark-1.3-contributor | 1.43T | 1.1% | −14% | $0.10 / $0.20 (Meta) | 1,048,576 |
| 18 | GPT-6 Astra | openai/gpt-6-astra | 1.41T | 1.0% | +36% | $10 / $50 (OpenAI) | 1,050,000 |
| 19 | GLM 5.2 | z-ai/glm-5.2 | 1.33T | 1.0% | −12% | $1.40 / $4.40 (Z.ai) | 1,048,576 |
| 20 | GPT-5.6 Sol | openai/gpt-5.6-sol | 1.32T | 1.0% | −21% | $2 / $10, 50% promotional discount (OpenAI) | 1,050,000 |
How to read the price column:
- Time-of-day prices. DeepSeek's endpoints charge the higher rate on weekdays from 01:00 to 04:00 and from 06:00 to 10:00 UTC. Tencent's endpoints charge less from 16:00 to 24:00 UTC.
- Rows 7 and 10 quote DeepInfra's standard, non-promotional price because DeepSeek does not serve these two V4 Flash snapshots on OpenRouter.
- Row 8 is the free
:freeendpoint. Paid endpoints for Nemotron 3 Ultra started at $0.50 / $2.20 (DeepInfra). - Jev 1.13 is not a chat model. TypeSafe describes it as a structured decision model that returns a typed choice rather than free text. It is not in OpenRouter's public model list, although its single endpoint is.
Space Bunny Alpha is an unnamed stealth model. OpenRouter added it on September 23, 2026 and says a third-party provider operates it and has chosen to stay anonymous during the preview. That provider may retain prompts and completions but, according to OpenRouter, does not use them for training. The model accepts text, image, and video input with a 1M-token context. OpenRouter's public model API listed no price or endpoint for it on October 5, so its maker and its long-term price are unknown. Stealth models do not stay as they are: an earlier one, Union Alpha, was later revealed as Unbiased's Pareto, and its free stealth period ended. Space Bunny Alpha is in the table because the rankings count it, but it is not a pick anywhere in this guide. Do not build recurring work on a model whose provider and price you cannot name.
What changed this month
- An anonymous model took the lead. Space Bunny Alpha processed 38.7T tokens in the week, up 179% on the week before. Measured by requests rather than tokens, stealth models had 10.5% of OpenRouter's text requests in the week beginning September 28, up 296%.
- DeepSeek V4.1 Flash is the top named model. It processed 25.6T tokens in the week (+31%) and 65.9T over the 30 days to October 4, the most of any model in that window. GLM 5.3 Flash was second over 30 days (54.9T), ahead of Space Bunny Alpha (52.5T), but its weekly volume fell 40%.
- Newer cheap models gained share. Xiaomi MiMo-V2.6-Flash, added on September 21, reached 9.63T (+75%). GPT-6 Luna more than doubled (+116%) while GPT-5.6 Luna fell 49%.
- Frontier models are small by volume but growing. Claude Opus 5.5 grew 128% to 2.46T and GPT-6 Astra 36% to 1.41T. The four frontier-priced models in the top 20 (Claude Opus 5.5, Kimi K3, GPT-6 Astra, and GPT-5.6 Sol) together had about 5% of top-20 tokens.
- The newest flagships are not in the top 20 yet. Claude Sonnet 5.5 (added September 28) and GPT-6.1 Sol (September 29) lead OpenRouter's new-and-trending list with 789B and 731B tokens in the week, below the 1.32T needed for 20th place.
- Requests and tokens tell different stories. Google had 21.5% of text requests in the week, second only to DeepSeek's 21.8%, yet Gemini 3.8 Flash was its only model in the top 20 by tokens.
- Corrections to the September version. It quoted catalog-level prices. GLM 5.3's $0.84 / $2.64 matches a discounted third-party route (Phala, 40% off on October 5), while Z.ai's own endpoint charges $1.40 / $4.40. Grok 4.7's $1.60 / $4.80 has become $2 / $6, xAI's own price below 200K prompt tokens.
Frontier Models on OpenRouter: Current Prices
These models rank lower by volume because they cost more per token, but they are the candidates for final review, hard coding, and long-running agent work. Prices are per 1M tokens from each model maker's own OpenRouter endpoint on October 5, 2026.
| Model | OpenRouter ID | Input / output (provider) | Context | Use it for |
|---|---|---|---|---|
| Claude Fable 5.1 | anthropic/claude-fable-5.1 | $10 / $50 (Anthropic) | 1,000,000 | Hardest long-horizon work |
| GPT-6 Astra | openai/gpt-6-astra | $10 / $50 (OpenAI) | 1,050,000 | Hardest reasoning, coding, research |
| Claude Opus 5.5 | anthropic/claude-opus-5.5 | $4 / $20 (Anthropic) | 1,000,000 | Complex coding and agents |
| Kimi K3 | moonshotai/kimi-k3 | $3 / $15 (Moonshot AI) | 1,048,576 | Long-context reasoning |
| Claude Sonnet 5.5 | anthropic/claude-sonnet-5.5 | $2 / $10 (Anthropic) | 1,000,000 | Daily coding and knowledge work |
| GPT-6.1 Sol | openai/gpt-6.1-sol | $2 / $10 (OpenAI) | 1,050,000 | Agentic coding and professional work |
| GPT-6 Sol | openai/gpt-6-sol | $2 / $10 (OpenAI) | 1,050,000 | Coding and agent work |
| Qwen3.8 Max | qwen/qwen3.8-max-0902 | $2 / $6 (Alibaba) | 1,000,000 | Multimodal agents |
| Grok 4.7 | x-ai/grok-4.7 | $2 / $6 below 200K prompt tokens; $4 / $12 above (xAI) | 500,000 | Long-running software engineering |
| GLM 5.3 | z-ai/glm-5.3 | $1.40 / $4.40 (Z.ai) | 1,048,576 | Lower-cost coding |
| Gemini 3.8 Flash | google/gemini-3.8-flash | $0.75 / $3.75, 50% promotional discount (Google AI Studio) | 1,048,576 | Fast multimodal work |
| DeepSeek V4 Pro 0813 | deepseek/deepseek-v4-pro-0813 | $0.66 / $1.98 off-peak; $1.32 / $3.96 peak (DeepSeek) | 1,048,576 | Low-cost reasoning and coding |
| GPT-6 Luna | openai/gpt-6-luna | $0.10 / $0.50 (OpenAI) | 1,050,000 | High-volume routing and extraction |
Cache reads, long-context tiers, reasoning tokens, and tool calls change the real bill. The same model also costs different amounts on different endpoints. OpenAI's flex endpoints on OpenRouter charged half the prices above and its fast endpoints double, and US or EU regional endpoints on Azure, Amazon Bedrock, and Google Vertex charged 10% more for several of the Claude and OpenAI models. For direct-API prices and release details, see Best AI models of 2026.
Discounts and batch pricing
Searchers often ask for "discounted" OpenRouter models. Two separate mechanisms exist:
- Promotional discounts. OpenRouter's discounted models collection listed GPT-5.6 Sol at $2 / $10 (50% off) on October 5, plus deeper temporary cuts on smaller models such as Solar Pro 4 (70% off, $0.09 / $0.36). The page states that promotions start and end without notice. The endpoint API marks the same thing per provider: DeepInfra's GLM 5.3 Flash route, for example, was 50% off at $0.075 / $0.25, below Z.ai's $0.15 / $0.50.
- Batch API. OpenRouter launched Batch on September 22, 2026: providers finish requests within a 24-hour window and "generally charge 50% (and sometimes less)" of the normal per-token price. On October 5 the model API listed 73
:batchvariants, for exampleopenai/gpt-6-astra:batchat $5 / $25 andanthropic/claude-sonnet-5.5:batchat $1 / $5. Use it for evaluations, backfills, and bulk extraction, not for interactive chat.
Free models are a third option. The current free catalog is covered in OpenRouter free models.
Best OpenRouter Model by Job
| Job | Start with | Cheaper alternative to test | Route into Teamday |
|---|---|---|---|
| Multi-file coding | Claude Opus 5.5 or GPT-6.1 Sol | GLM 5.3 or DeepSeek V4 Pro 0813 | Max, the app builder, in a coding harness |
| SEO research | Cheap extractor plus a strong reviewer | DeepSeek V4.1 Flash for extraction | Sarah, the SEO manager |
| Content production | Mid-tier drafter plus frontier editor | GLM 5.3 Flash for first drafts | Maya, the content creator |
| Competitive intelligence | Long-context model | Kimi K3 or Gemini 3.8 Flash | Kai, the CI analyst |
| Data reporting | Structured-output model plus SQL-capable reviewer | GPT-6 Luna | James, the data analyst |
| Strategy memos | Frontier reasoning model | Claude Sonnet 5.5 | Nova, the CMO |
| Image or video planning | Strong multimodal model plus media tools | Gemini 3.8 Flash | Iris and Reel |
These are starting points based on price, context, and each provider's stated strengths, not results from an independent benchmark. Run the evaluation suite below on your own work before committing.
The Model Market Map
Think of OpenRouter models in work tiers, not brand tiers.
| Tier | Examples (October 2026) | Best use | Failure mode |
|---|---|---|---|
| Frontier reasoning | GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5 | Architecture, final review, complex debugging | Expensive if used for every step |
| Coding-specialist | GPT-6 Sol, Grok 4.7, GLM 5.3 | Patches, migrations, tests | Looks right but breaks the build |
| Fast general | Gemini 3.8 Flash, GPT-5.6 Luna, Hy4 preview | Drafting, summaries, rewrites | Generic output |
| Long-context | Kimi K3, GLM 5.3 Flash (1M) | Transcripts, research packs, repository scans | Misses details in huge context |
| Cheap structured-output | DeepSeek V4.1 Flash, GPT-6 Luna, MiMo-V2.6-Flash | Classification, JSON extraction, routing | Schema drift and retries |
| Free | Nemotron 3 Ultra, Laguna S 2.1 | Experiments, low-risk drafts | Rate limits and endpoint changes |
How to Route Work Across Tiers
A practical model-routed mission looks like this:
- A cheap or free model — for example DeepSeek V4.1 Flash or one of the OpenRouter free models — extracts facts from source material.
- A mid-tier model drafts a structured artifact.
- A frontier model reviews claims, risks, and missing context.
- The agent writes the final file into the workspace.
- The result is published for review, as in Sarah's weekly SEO report or Max's ship log.
- A recurring mission reruns the workflow.
Ranking Criteria That Actually Matter
Approved output per dollar
Token price is incomplete. Measure cost per accepted artifact after retries, editing, and review time. A $0.10 model that needs three retries and a heavy edit can cost more than a $2 model that is right the first time.
Tool reliability
For agents, tool calls matter more than chat quality. The model must follow schemas, call tools correctly, and recover when a tool fails. The runtime matters too — the harness encyclopedia compares how each one handles tools, files, and tests.
Context fit
A cheaper long-context model can beat a stronger short-context model if the task needs whole-document or whole-repository awareness. Check the endpoint's context, not the model family's headline figure: on October 5 most GLM 5.3 Flash endpoints on OpenRouter reported 1,048,576 tokens, but Reka's reported 262,144.
Endpoint stability
Preview models, stealth models, and dated snapshots change. DeepSeek V4 Flash currently has two live snapshots on OpenRouter, and even one provider prices them differently: on October 5 DeepInfra charged $0.09 / $0.18 for deepseek/deepseek-v4-flash (the 0423 build) and $0.06 / $0.18 for deepseek/deepseek-v4-flash-0731. Prices also differ across providers of one ID: GLM 5.3 ran from $0.03 input with $12 output (Relace) to $2.80 / $8.80 (Alibaba's fast endpoint). Pin the exact ID you tested, and price the provider you will actually use against your own input/output mix.
Fallback behavior
Production work needs fallback routing. If a free or cheap endpoint is unavailable, the agent should escalate to a stronger model or stop with a clear error.
Practical Evaluation Suite
Run the same work through candidate models and score the output.
| Test | Pass condition |
|---|---|
| Long brief summary | Captures source facts without inventing missing details |
| JSON extraction | Valid schema on first try, no extra prose |
| Code patch | Build or tests pass, or the failure is correctly explained |
| SEO brief | Names exact pages, queries, actions, and expected metric |
| Executive memo | Makes tradeoffs explicit and avoids fake certainty |
| Tool-output synthesis | Cites source rows and separates data from opinion |
| Fallback test | Recovers or stops cleanly when a model or tool call fails |
Using OpenRouter Models in Teamday
What Teamday supports today:
- OpenRouter key: save an
OPENROUTER_API_KEYin runner settings. The Pi harness then runs Kimi K3, GLM-5.3, GLM-5.3 Flash, and DeepSeek V4 Pro through OpenRouter. Other OpenRouter model IDs are not first-class picks in the model picker. - Other frontier models run through their own harnesses with your own account: Claude on Claude Code, GPT-6 on Codex, and Gemini on the Google harness.
- Pricing: connecting your own AI account lets you use Teamday free for 14 days. Then it's $99/month with unlimited work on your own AI account, or prepaid AI credits ($50 or $200, never expire) at the AI model's price plus 25% if you don't have one.
Put these models to work with your own key. Connect OpenRouter, pick Kimi K3, GLM-5.3, GLM-5.3 Flash, or DeepSeek V4 Pro, and give Sarah, Maya, or Max a real task. Your token usage bills to your OpenRouter account.
Use my OpenRouter key in Teamday →Recommended Mission: Monthly Model Review
Create a monthly "Model Router Review" mission:
- Pull the latest OpenRouter rankings and model prices.
- Re-run the evaluation suite across the current candidates.
- Record which model was promoted, demoted, or removed, with the reason.
- Update the agents' model settings only after the results are reviewed.
Run the review as a recurring mission. Max, Teamday's app builder, can run the evaluation on a schedule with your OpenRouter key and write the results to your workspace for review. Connect your own provider and use Teamday free for 14 days.
Connect OpenRouter and start the review →