---
title: "The Gateway Charts, week 46 (Nov 9 to 15, 2026)"
series: "The Gateway Charts"
week: 2026-w46
period: 2026-11-09 to 2026-11-15 (UTC)
compared_with: 2026-11-02 to 2026-11-08
status: draft
license: CC BY 4.0
canonical: https://www.requesty.ai/data/weekly/2026-w46
json: https://www.requesty.ai/data/weekly/2026-w46/data.json
---

# The Gateway Charts, week 46: Last week in AI. Who climbed. Who slipped.

> Every request through the Requesty gateway names a model, a host and the app that sent it. Each week we rank them against the week before and pull out what changed, and what it means for what you build.

Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026. Publisher: Requesty.

## What to pay attention to

1. deepseek-v4.1-flash held #1 on the Requesty gateway in week 46 (Nov 9 to 15, 2026) with 28.3% of tokens, up 2.27 points on the week before.
2. Five coding agents (OpenCode, Claude Code, Pi, DeepSeek Harness and GitHub Copilot) sent 33.4% of tokens through the Requesty gateway in week 46 (Nov 9 to 15, 2026).
3. 67.6% of active teams on the Requesty gateway used two or more models in week 46 (Nov 9 to 15, 2026); 58.6% used two or more labs and 60.2% two or more providers.
4. claude-haiku-5-5 reached its first token 1.3x sooner on the Requesty gateway in week 46 (Nov 9 to 15, 2026): a median of 1.76 s, down from 2.21 s the week before, on streaming requests.
5. In week 46 (Nov 9 to 15, 2026), 96.1% of requests on a fallback policy whose first attempt hit a provider error were still served by another host on the Requesty gateway, up from 86.4% the week before.
6. 78.1% of input tokens through the Requesty gateway in week 46 (Nov 9 to 15, 2026) were read from a prompt cache; models read 46 input tokens for every output token.

## Findings

### 1. deepseek-v4.1-flash pulled further ahead

deepseek-v4.1-flash held #1 on the Requesty gateway in week 46 (Nov 9 to 15, 2026) with 28.3% of tokens, up 2.27 points on the week before.

- Value: 28.3% of tokens, each team capped at 2%
- Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026
- Why it matters: Last week it gave up share for the first time since September. This week it took it back, and it now carries more than the next three models combined. The order below it is unchanged: claude-opus-5-5 at 10.5%, gpt-6-luna at 8.5%.
- What to do: See your own model mix by key and team (https://docs.requesty.ai/features/usage-analytics)
- Permalink: https://www.requesty.ai/data/weekly/2026-w46/top-model
- Social card: https://www.requesty.ai/data/weekly/2026-w46/top-model/square.png

### 2. Coding agents sent a third of tokens

Five coding agents (OpenCode, Claude Code, Pi, DeepSeek Harness and GitHub Copilot) sent 33.4% of tokens through the Requesty gateway in week 46 (Nov 9 to 15, 2026).

- Value: 33.4% of tokens, each team capped at 2%
- Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026
- Why it matters: That is the second week in a row at about a third. The SDKs in the top 10 send 38.3% between them, so agents are now almost level with every SDK and framework we can name. Their requests are long, cached and end in tool calls.
- What to do: Run Claude Code through Requesty (https://docs.requesty.ai/integrations/claude-code)
- Permalink: https://www.requesty.ai/data/weekly/2026-w46/coding-agents
- Social card: https://www.requesty.ai/data/weekly/2026-w46/coding-agents/square.png

### 3. Most teams used more than one model

67.6% of active teams on the Requesty gateway used two or more models in week 46 (Nov 9 to 15, 2026); 58.6% used two or more labs and 60.2% two or more providers.

- Value: 67.6% of active teams
- Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026
- Why it matters: Single-model stacks are still the minority. Almost one team in five ran six models or more, so switching and mixing models is routine work, not a migration project.
- What to do: Switch models without changing your integration (https://docs.requesty.ai/quickstart)
- Permalink: https://www.requesty.ai/data/weekly/2026-w46/multi-model
- Social card: https://www.requesty.ai/data/weekly/2026-w46/multi-model/square.png

### 4. claude-haiku-5-5 started 1.3× sooner

claude-haiku-5-5 reached its first token 1.3x sooner on the Requesty gateway in week 46 (Nov 9 to 15, 2026): a median of 1.76 s, down from 2.21 s the week before, on streaming requests.

- Value: 1.3× median time to first token, this week vs last
- Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026
- Why it matters: It started sooner but streamed slower: its median output speed fell from 462 to 291 tokens per second. First token and stream speed can move in opposite directions, so measure the one your users feel.
- What to do: Let the gateway pick the fastest host (https://docs.requesty.ai/features/latency-routing)
- Caveat: Latency is measured at the gateway, so it includes the host and the network, and reflects the prompts teams sent that week.
- Caveat: Measured at the gateway on streaming requests, so it blends the model, the host and the prompts teams sent that week.
- Permalink: https://www.requesty.ai/data/weekly/2026-w46/speed-up
- Social card: https://www.requesty.ai/data/weekly/2026-w46/speed-up/square.png

### 5. 96% of failed requests still got served

In week 46 (Nov 9 to 15, 2026), 96.1% of requests on a fallback policy whose first attempt hit a provider error were still served by another host on the Requesty gateway, up from 86.4% the week before.

- Value: 96.1% of requests on a fallback policy whose first host returned a provider error
- Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026
- Why it matters: A provider error only reaches your users when there is nowhere else to send the request. This week most rescues ran from Bedrock to Anthropic's own API, the same model on another host.
- What to do: Set up a fallback policy (https://docs.requesty.ai/features/fallback-policies)
- Permalink: https://www.requesty.ai/data/weekly/2026-w46/rescues
- Social card: https://www.requesty.ai/data/weekly/2026-w46/rescues/square.png

### 6. 78.1% of input came from cache

78.1% of input tokens through the Requesty gateway in week 46 (Nov 9 to 15, 2026) were read from a prompt cache; models read 46 input tokens for every output token.

- Value: 78.1% of input tokens
- Period: Nov 9 to 15, 2026 (UTC), compared with Nov 2 to 8, 2026
- Why it matters: Input is most of the bill for agent workloads, and the cache decides what that input costs. Compare models on what you pay after caching, not on the list price for output.
- What to do: Turn on auto caching (https://docs.requesty.ai/features/auto-caching)
- Permalink: https://www.requesty.ai/data/weekly/2026-w46/prompt-cache
- Social card: https://www.requesty.ai/data/weekly/2026-w46/prompt-cache/square.png

## Models by share of tokens

Unit: % of tokens, each team capped at 2%.

| Rank | Move | Name | Share | Change | Peak (12 wk) | Weeks on chart |
|---|---|---|---|---|---|---|
| 1 | = | deepseek-v4.1-flash (DeepSeek) | 28.3% | +2.27 pts | #1 | 10 |
| 2 | = | claude-opus-5-5 (Anthropic) | 10.5% | +0.28 pts | #1 | 8 |
| 3 | = | gpt-6-luna (OpenAI) | 8.46% | +0.13 pts | #3 | 8 |
| 4 | = | glm-5.3-flash (Z.ai) | 5.42% | −0.62 pts | #2 | 12 |
| 5 | ▲ 1 | claude-sonnet-5-5 (Anthropic) | 5.35% | +0.08 pts | #5 | 7 |
| 6 | ▼ 1 | gpt-6.1-sol (OpenAI) | 5.04% | −0.36 pts | #4 | 7 |
| 7 | = | deepseek-v4-flash (DeepSeek) | 4.57% | +0.31 pts | #1 | 12 |
| 8 | = | claude-haiku-5-5 (Anthropic) | 3.75% | +0.25 pts | #8 | 6 |
| 9 | ▲ 1 | nemotron-3-ultra-550b-a55b (NVIDIA) | 2.40% | +0.03 pts | #9 | 12 |
| 10 | ▲ 1 | claude-sonnet-5 (Anthropic) | 2.18% | +0.17 pts | #4 | 12 |
| 11 | ▲ 2 | gpt-6.1-terra (OpenAI) | 2.16% | +0.82 pts | #11 | 3 |
| 12 | ▼ 3 | gpt-5.6-luna (OpenAI) | 2.02% | −0.37 pts | #3 | 12 |
| 13 | ▲ 1 | gemini-3.8-flash (Google) | 1.45% | +0.16 pts | #6 | 11 |
| 14 | ▼ 2 | nemotron-3.5-lightning-30b-a3b (NVIDIA) | 1.11% | −0.41 pts | #12 | 12 |
| 15 | = | glm-5.3 (Z.ai) | 0.88% | −0.35 pts | #7 | 12 |
| 16 | New | gemini-4-carbon (Google) | 0.73% | +0.73 pts | #16 | 1 |
| 17 | = | claude-opus-5 (Anthropic) | 0.59% | −0.14 pts | #5 | 12 |
| 18 | ▲ 1 | claude-sonnet-4-6 (Anthropic) | 0.58% | +0.01 pts | #13 | 12 |
| 19 | ▲ 4 | minimax-m3.1-flash (MiniMax) | 0.56% | +0.05 pts | #12 | 3 |
| 20 | ▲ 4 | leanstral-1-5 (Mistral) | 0.53% | +0.03 pts | #9 | 12 |
| 21 | ▼ 5 | ling-3.1-flash (Other) | 0.52% | −0.40 pts | #16 | 6 |
| 22 | = | gemini-4-argon (Google) | 0.47% | −0.04 pts | #22 | 5 |
| 23 | ▼ 5 | deepseek-v4-pro (DeepSeek) | 0.43% | −0.29 pts | #5 | 12 |
| 24 | ▼ 3 | gpt-6-sol (OpenAI) | 0.38% | −0.17 pts | #9 | 8 |
| 25 | ▼ 5 | gemini-3.5-flash-lite (Google) | 0.31% | −0.25 pts | #18 | 11 |
| 26 | = | gemini-3.1-flash-lite (Google) | 0.24% | −0.14 pts | #17 | 12 |
| 27 | ▲ 1 | gemini-2.5-flash-lite (Google) | 0.24% | −0.01 pts | #25 | 12 |
| 28 | ▼ 3 | kimi-k3.1-pro (Moonshot) | 0.22% | −0.22 pts | #25 | 4 |
| 29 | New | gemini-4-hydrogen (Google) | 0.20% | +0.20 pts | #29 | 1 |
| 30 | ▼ 1 | gemini-3.7-flash (Google) | 0.19% | −0.06 pts | #8 | 12 |
| 31 | ▲ 1 | claude-haiku-4-5 (Anthropic) | 0.14% | −0.07 pts | #19 | 12 |
| 32 | ▲ 3 | gemma-4-31b-it (Google) | 0.14% | −0.03 pts | #32 | 9 |
| 33 | ▲ 1 | gemini-3.1-pro-preview (Google) | 0.11% | −0.08 pts | #33 | 3 |
| 34 | ▲ 3 | gemini-2.5-flash (Google) | 0.10% | −0.04 pts | #32 | 7 |
| 35 | ▲ 3 | glm-5.2 (Z.ai) | 0.10% | −0.01 pts | #6 | 5 |
| 36 | = | gpt-5.4 (OpenAI) | 0.09% | −0.07 pts | #18 | 12 |
| 37 | ▼ 6 | gpt-5.6-sol (OpenAI) | 0.09% | −0.13 pts | #2 | 12 |
| 38 | ▼ 5 | mistral-large-4 (Mistral) | 0.09% | −0.11 pts | #31 | 6 |
| 39 | ▲ 1 | gpt-5.5 (OpenAI) | 0.09% | 0.00 pts | #28 | 11 |
| 40 | ▼ 1 | gemini-3.5-flash (Google) | 0.08% | −0.02 pts | #27 | 12 |

## Apps by share of tokens

Unit: % of tokens, each team capped at 2%.

| Rank | Move | Name | Share | Change | Peak (12 wk) | Weeks on chart |
|---|---|---|---|---|---|---|
| 1 | = | OpenAI Python SDK (SDK or framework) | 21.3% | +2.16 pts | #1 | 12 |
| 2 | = | OpenCode (Coding agent) | 13.1% | −1.00 pts | #2 | 12 |
| 3 | ▲ 1 | Vercel AI SDK (SDK or framework) | 10.1% | +0.77 pts | #3 | 12 |
| 4 | ▼ 1 | Claude Code (Coding agent) | 9.83% | −1.65 pts | #3 | 12 |
| 5 | = | Pi (Coding agent) | 6.64% | +3.11 pts | #4 | 12 |
| 6 | ▲ 1 | Anthropic SDK (SDK or framework) | 2.51% | −0.16 pts | #6 | 9 |
| 7 | ▲ 1 | OpenAI SDK (other) (SDK or framework) | 2.36% | −0.13 pts | #3 | 12 |
| 8 | ▼ 2 | DeepSeek Harness (Coding agent) | 2.15% | −0.54 pts | #6 | 11 |
| 9 | ▲ 1 | LiteLLM (SDK or framework) | 2.11% | +0.21 pts | #9 | 11 |
| 10 | ▼ 1 | GitHub Copilot (Coding agent) | 1.60% | −0.33 pts | #6 | 12 |
| 11 | = | Kilo Code (Coding agent) | 1.39% | +0.12 pts | #8 | 12 |
| 12 | ▲ 2 | Zoo Code (Coding agent) | 0.27% | +0.09 pts | #11 | 12 |
| 13 | = | Open WebUI (App) | 0.25% | −0.01 pts | #13 | 12 |
| 14 | ▼ 2 | LangChain (SDK or framework) | 0.21% | −0.09 pts | #10 | 12 |
| 15 | ▲ 2 | LoreBary (App) | 0.01% | 0.00 pts | #15 | 10 |

## Labs by share of tokens

Unit: % of tokens, each team capped at 2%.

| Rank | Move | Name | Share | Change | Peak (12 wk) | Weeks on chart |
|---|---|---|---|---|---|---|
| 1 | = | DeepSeek | 34.1% | +2.04 pts | #1 | 12 |
| 2 | = | Anthropic | 24.4% | +0.50 pts | #1 | 12 |
| 3 | = | OpenAI | 19.7% | −0.51 pts | #2 | 12 |
| 4 | = | Z.ai | 6.62% | −1.17 pts | #2 | 12 |
| 5 | = | Google | 4.47% | +0.37 pts | #4 | 12 |
| 6 | = | NVIDIA | 3.59% | −0.39 pts | #5 | 12 |
| 7 | = | MiniMax | 0.94% | −0.15 pts | #7 | 12 |
| 8 | ▲ 1 | Mistral | 0.70% | −0.03 pts | #7 | 12 |
| 9 | ▼ 1 | Moonshot | 0.69% | −0.17 pts | #6 | 12 |
| 10 | = | Alibaba | 0.33% | +0.07 pts | #8 | 12 |
| 11 | Re | Xiaomi | 0.22% | −0.73 pts | #7 | 5 |
| 12 | ▼ 1 | xAI | 0.09% | −0.05 pts | #10 | 11 |

## Hosts by share of tokens

Unit: % of tokens (hosts are not team-capped).

| Rank | Move | Name | Share | Change | Peak (12 wk) | Weeks on chart |
|---|---|---|---|---|---|---|
| 1 | = | Azure | 16.3% | +0.21 pts | #1 | 12 |
| 2 | = | Bedrock | 12.8% | +0.05 pts | #2 | 12 |
| 3 | = | Sference | 10.4% | −1.11 pts | #1 | 12 |
| 4 | ▲ 2 | Melious | 9.69% | +3.90 pts | #4 | 6 |
| 5 | = | DeepInfra | 7.59% | +0.05 pts | #2 | 12 |
| 6 | ▼ 2 | Vertex (Claude) | 7.09% | −0.68 pts | #4 | 9 |
| 7 | ▲ 1 | Anthropic | 4.33% | +0.18 pts | #4 | 12 |
| 8 | ▲ 2 | Vertex (Gemini) | 4.16% | +0.21 pts | #3 | 8 |
| 9 | ▼ 2 | OpenAI | 3.74% | −0.86 pts | #6 | 12 |
| 10 | ▲ 2 | TensorX | 3.58% | −0.03 pts | #5 | 12 |
| 11 | ▼ 2 | DeepSeek | 3.43% | −0.70 pts | #2 | 11 |
| 12 | ▼ 1 | NVIDIA | 3.06% | −0.57 pts | #8 | 12 |
| 13 | ▲ 3 | Fireworks | 1.12% | +0.30 pts | #11 | 12 |
| 14 | = | Mistral | 0.83% | −0.31 pts | #12 | 12 |
| 15 | = | Novita | 0.57% | −0.46 pts | #9 | 10 |
| 16 | ▲ 2 | Z.ai | 0.29% | −0.04 pts | #12 | 12 |
| 17 | ▲ 3 | Alibaba | 0.14% | −0.01 pts | #15 | 12 |
| 18 | Re | Xiaomi | 0.14% | −0.68 pts | #18 | 2 |
| 19 | Re | Inceptron | 0.11% | +0.02 pts | #19 | 9 |
| 20 | ▼ 1 | Moonshot | 0.10% | −0.13 pts | #17 | 10 |

## Vital signs

| Metric | This week | Last week | Note |
|---|---|---|---|
| Input read from cache | 78.1% | 77.1% | Share of input tokens served from a prompt cache. |
| Input per output token | 46 : 1 | 46 : 1 | Tokens read for each token written. |
| Output spent thinking | 31.5% | 33.9% | Reasoning tokens as a share of output. |
| Turns ending in a tool call | 18.5% | 20.1% | Responses that asked for a tool instead of answering. |
| Median first token | 1.41 s | 1.52 s | Across every streaming request. |
| Median stream speed | 171 tok/s | 156 tok/s | Once the model starts talking. |
| Routed by a policy | 35.9% | 36.8% | Fallback, latency or load-balancing policies. |
| Routed by latency | 19.0% | 21.6% | Tokens where the gateway picked the quickest host. |

## Method

- Aggregated from the Requesty production gateway for the week (Monday to Sunday, UTC), compared with the week before and the ten weeks before that.
- We publish shares, rates, ranks and latency only: no volumes, no customer data, no prompt content.
- In the model, lab and app charts each team counts for at most 2% of a week's tokens, so one heavy team can't bend the chart.
- An entry is named only when at least 20 teams used it and no single team is more than 60% of it.
- Latency is measured at the gateway on streaming requests. It includes the host and the network, and reflects the prompts teams sent, not a controlled benchmark.
- Bring-your-own-key traffic, internal test traffic, synthetic load and abusive accounts are excluded.

## Cite this

Source: Requesty, The Gateway Charts, week 46 (Nov 9 to 15, 2026). https://www.requesty.ai/data/weekly/2026-w46

License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
