Requesty (@RequestyAI) header
Requesty (@RequestyAI) profile picture

Requesty

@RequestyAI

The control plane for production AI. 600+ models behind one API, with smart routing, spend management, governance & observability.

requesty.ai
Joined September 2023
  • Tweets266
  • Following98
  • Followers962
151Media
Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)Photo by Requesty (@RequestyAI)
Not every Claude Code call needs Opus. Auto Routing, now in beta: describe the kinds of work your team does, point each one at a model, and Requesty makes the call on the first message.
11
10
6
76
3,744,495
“We control where data goes, which models run on which tasks, and what we spend, from one place.” Michael Friedmann of @anwaltde, on 77B tokens of legal AI through Requesty (66.9B in, 7.0B out). requesty.ai/customers/anwalt
Image from Requesty (@RequestyAI)
1
56
Gemini 4 Carbon is now on Requesty. Carbon: $1 in, $5 out per million tokens. Carbon is built for multi-step coding, debugging and agent runs. Call gemini-4-carbon
Google (@Google) profile pictureGoogle
@GoogleNov 10
Gemini 4 Carbon is now available. It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning. If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
1
1
124
Claude Fable 5.2 is live on Requesty. Anthropic’s new top model: up to 30% faster than Fable 5.1, at $8/$40 per M tokens. anthropic/claude-fable-5.2
Claude (@claudeai) profile pictureClaude
@claudeaiNov 9
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
90
Caching doesn't only cut the bill. On 100k+ token requests on Requesty, median time to first token fell from 8.8s to 2.4s when the prompt was served from cache. Cost is still top of mind but latency is what your users feel.
Image from Requesty (@RequestyAI)
1
65
Claude Haiku tokens on Requesty grew 381% from September to October. Across the 13 model families we track, 9 grew and 4 shrank. Kimi fell 42%. Overall, Claude's share of successful chat tokens on Requesty went from 17% in September to 23% in October.
Image from Requesty (@RequestyAI)
1
2
137
Which branch burned the budget this week? Tag each request with X-Requesty-Branch and the Advanced tab groups spend by branch. Repo and user work the same way. docs.requesty.ai/features/an…
2
76
Your prompt doesn’t have to live in your code. Prompt Library keeps the system message, the user template, the model and the response format in one place. Change the latest version without a deploy, or pin a version and leave it alone.
3
147
Share of tokens on Requesty going to open-weights models: Week of 5 Jan: 4% Week of 26 Oct: 52% Share of spend over the same weeks: 0.5% to 17%. Half the tokens. A sixth of the bill.
2
175
Don’t pick models based on someone else’s benchmark. FrontierCode 2.0 says open models struggle. On our own set of 129 Go backend tasks, Kimi K3.1 Pro solved 81%, sitting right alongside the frontier labs. Run new models on YOUR workload. Congrats @Kimi_Moonshot team!
Image from Requesty (@RequestyAI)
1
1
3
329
Auto Routing in three steps: 1. Write the question to answer 2. List the answers it can choose from 3. Pick a model or policy for each Unsure? It falls back to your strongest model. Pinned per session, so you keep cache hits. docs.requesty.ai/features/au…
3
124
Particle-1 is DeepSeek V4.1 Flash, post-trained to finish tasks in fewer tokens. So we ran both through our internal coding benchmark: 88 real Go backend tasks, hidden tests, same harness. 39% fewer tokens per task, 20% lower cost, half the time per turn. Congrats @Suhail 🧵
Requesty logo. Kicker "Coding benchmark · Particle-1 vs its base". Headline "Same base model. 39% fewer tokens per task." Four panels compare deepseek-v4.1-flash (grey dots) with particle-1 (blue dots). Tokens per task: 2,399k vs 1,463k. Cost per task: $0.15 vs $0.12. Latency per turn: 5.4s vs 2.7s. Pass rate: 67% vs 61%, one dot per task out of 88.

ALT Requesty logo. Kicker "Coding benchmark · Particle-1 vs its base". Headline "Same base model. 39% fewer tokens per task." Four panels compare deepseek-v4.1-flash (grey dots) with particle-1 (blue dots). Tokens per task: 2,399k vs 1,463k. Cost per task: $0.15 vs $0.12. Latency per turn: 5.4s vs 2.7s. Pass rate: 67% vs 61%, one dot per task out of 88.

2
5
287
Speed stood out most. At 2.7s median per turn, Particle-1 was the fastest of all 19 models we tested, ahead of glm-5.3-flash (3.4s) and gpt-6-sol (3.6s).
Requesty logo. Kicker "Coding benchmark · Latency". Headline "Fastest of 19 models. 2.7s median per turn." A ranked bar chart of median seconds per model turn for 19 models. Particle-1 is first at 2.7s, in blue, followed by glm-5.3-flash 3.4s and gpt-6-sol 3.6s. Its base, deepseek-v4.1-flash, is at 5.4s. The slowest are deepseek-v4-pro at 17.0s and hy4-preview at 22.6s.

ALT Requesty logo. Kicker "Coding benchmark · Latency". Headline "Fastest of 19 models. 2.7s median per turn." A ranked bar chart of median seconds per model turn for 19 models. Particle-1 is first at 2.7s, in blue, followed by glm-5.3-flash 3.4s and gpt-6-sol 3.6s. Its base, deepseek-v4.1-flash, is at 5.4s. The slowest are deepseek-v4-pro at 17.0s and hy4-preview at 22.6s.

1
71
The trade-off: 61% of tasks solved vs 67% for its base, level with DeepSeek V4 Pro among open-weight models. Frontier models still lead, Opus 5.5 at 90%. For high-volume coding loops where tokens and turn time add up, worth a try: model="particle/particle-1"
Requesty logo. Kicker "Coding benchmark · Pass rate". Headline "61% of tasks solved, level with DeepSeek V4 Pro." Eight open-weight models, each drawn as 88 dots, one dot per task, with solved tasks filled. hy4-preview 68%, deepseek-v4.1-flash 67%, qwen3.8-flash 64%, kimi-k3 64%, particle-1 61% (in blue), deepseek-v4-pro 61%, glm-5.3 60%, glm-5.3-flash 58%.

ALT Requesty logo. Kicker "Coding benchmark · Pass rate". Headline "61% of tasks solved, level with DeepSeek V4 Pro." Eight open-weight models, each drawn as 88 dots, one dot per task, with solved tasks filled. hy4-preview 68%, deepseek-v4.1-flash 67%, qwen3.8-flash 64%, kimi-k3 64%, particle-1 61% (in blue), deepseek-v4-pro 61%, glm-5.3 60%, glm-5.3-flash 58%.

38
Fewer tokens per task is the kind of progress we love to route. Congrats @Suhail on Particle-1: 28 to 52% fewer output tokens across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, pass rates in line with the base. On Requesty today as particle/particle-1
Image from Requesty (@RequestyAI)
1
4
191
Sensitive project data. Public-sector clients. 5 to 15 models. LOUPZ points its AI chats, apps and agents at one European endpoint, router.eu.requesty.ai. Data goes to the US only when strictly required, and no app is tied to a provider. requesty.ai/customers/loupz
Image from Requesty (@RequestyAI)
2
138
We used it ourselves. Our Claude Code sessions on Auto Routing: 42% cheaper and 30% faster than everything on Opus 5.5.
Image from Requesty (@RequestyAI)
1
7
2,487
The part we care most about: the choice is pinned for the session. The model never switches mid-task, so prompt caches keep working. If the classifier isn't sure, it falls back to the model you trust most. Set up a policy: app.requesty.ai/routing-poli…
3
630
Last week Jev was the #1 model on Requesty by requests. By tokens it was #12. A decision model answers with a typed decision, not an essay, so a call uses ~12% of the tokens of an average LLM call. Token leaderboards barely see them.
1
197
Load more