The control plane for production AI. 600+ models behind one API, with smart routing, spend management, governance & observability.
- Tweets 256
- Following 98
- Followers 957
- Likes 183
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!
1
1
162
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
105
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"
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%.
34
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).
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
57
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 🧵
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
263
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
1
4
179
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
2
123
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
570
We used it ourselves.
Our Claude Code sessions on Auto Routing: 42% cheaper and 30% faster than everything on Opus 5.5.
1
7
2,037
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.
10
9
69
2,374,439
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
190
If your AI gateway guardrails are taking seconds to respond, you're adding a ridiculous amount of unnecessary latency to every single request.
And if you're running them in production, you're paying that price constantly.
Ours, 3–18 Oct:
4 ms median
15 ms p95
Just 23 ms p95 on 64k+ token prompts.
5
487
Your agent made 300 calls. One of them went wrong.
In a flat log, good luck.
Requesty rebuilds the run as one session: every prompt, reply and tool call in order, and any request open down to the token.
Replay a whole session: requesty.ai/product/observab…
1
1
6
273
Every Amber Notes meeting summary is an LLM call.
@amber_script put them all behind one Requesty integration:
> 40% lower AI cost per meeting
> 99.9% AI feature availability
> New models in minutes, not days
requesty.ai/customers/ambers…
1
162
Requesty is now a native provider plugin in Hermes Agent.
Add a fallback policy and when a provider goes down, your agent moves to the next model instead of stopping.
Change the chain in Requesty, not in your Hermes config.
docs.requesty.ai/integration…
2
7
355
Claude Code. Cursor. Codex. curl.
All on your API keys. Do you know which ones?
New in Requesty: see every agent on every key, catch leaked keys, kill them in one click.
→ app.requesty.ai/api-keys?vie…
2
1
2
171
Keys leak. Make a leaked key useless off your network.
New on Requesty: Allowed IPs.
List the IPv4 addresses and CIDR ranges that may call Requesty for your org.
Anything else is rejected, on every endpoint, and every change lands in the audit trail.
1
3
179
New @deepseek_ai v4.1 51% discounted on Requesty!
$0.098 / 1M tokens
$0.294 / 1M tokens
$0.003 cached / 1M tokens
1
2
148
A Jev decision through Requesty, median over the last 30 days:
Model: 241 ms
Requesty: 2.3 ms
That 2.3 ms is key auth, spend limits, approved-model and compliance checks, and rate limits. Under 1% of the request.
typesafe/jev-latest
2
156
Small one, but you kept asking for it.
Top up by card in USD, GBP, EUR, JPY or CNY, or iDEAL in EUR. It lands in your USD balance at the rate shown in checkout.
Auto top-up charges your card in its own currency too.
1
104
Morning spikes pushed NotarBot past provider quotas, and 429s stalled live calls. Rotating regions split the prompt cache.
Now: one stable endpoint, higher-limit deployments, cross-provider failover. Cache hits on nearly every turn.
requesty.ai/customers/notarb…
92
The fastest provider for a model changes all day.
Requesty's latency routing sends each new session to the fastest one, ranked on time to first token and throughput, corrected for errors.
Sessions stay pinned so your cache stays warm.
docs.requesty.ai/features/la…
1
121
Meta's Muse Wonder is live on Requesty.
Built for long-running agentic work: memory across sessions, persistence through blockers, check-ins before important actions.
$4 in / $20 out per 1M tokens, $0.20 cache reads, 1M context.
model: muse-wonder-1.0
1
4
272
Haiku 5.5 vs Haiku 4.5, same customers, same window:
2.9x faster output
3.8x longer answers
3.1x more context per request
44% of calls end in a tool call (14% on 4.5)
21% of output is thinking (3% on 4.5)
And 48% cheaper per request.
59
Claude 5.5 is a speed upgrade, not just a quality one.
p50 output speed on streaming Claude traffic through Requesty, Oct 8 to 11:
Haiku 5.5: 3.3x Haiku 4.5
Sonnet 5.5: 2.3x Sonnet 5
5 days after launch, Haiku 5.5 already carries 94% of Haiku tokens.
1
88
Haiku 5.5 is the fastest Claude switch we've measured on Requesty.
It took 51% of all Haiku tokens in its first 46 hours.
Same window after launch:
> Opus 4.8: 32% of Opus tokens
> Opus 5.5: 19%
> Sonnet 5.5: 17%
> Opus 4.7: 7%
3
6
456
Open-source models on Requesty.
January → September 2026:
Tokens: 5% → 53%
Requests: 12% → 42%
Spend: 0.4% → 14%
1
4
138
"Where did our prompts go, and who kept them?"
Requesty's new Compliance page answers that for every request: where the model was hosted, whether it retains or trains on data, and what scanning caught.
One click exports it as an evidence report.
app.requesty.ai/compliance
87
ZoomInfo evaluated over 7 AI gateways, then routed Claude Code, Copilot and its agents through Requesty.
"Requesty's enterprise readiness and configurability is what set it apart."
- Arkady Landes, Senior Cloud DevOps Manager, ZoomInfo
requesty.ai/customers/zoomin…
115
New provider on Requesty: @MeliousAI 🇪🇺
Melious serves open-weight models from European infrastructure. Now you can route to it with the same Requesty key:
• deepseek-v4.1-flash
• glm-5.3-flash
• glm-5.3
• deepseek-v4-flash-0731
requesty.ai/models?providers…
11
1,174
New in Requesty: connect your Claude Platform and OpenAI Platform orgs.
Users, API keys, workspaces and spend all show up in one admin panel, with a Source badge on every row.
One place to see usage, ownership and spend across every platform.
docs.requesty.ai/features/pl…
2
266
Decisions models return a probability for every option in your schema. Routing, triage, moderation.
Until now, none of them ran in the EU.
Clef by @Cloudflare, hosted by @sference: EU region, zero retention, no training. Live on Requesty.
requesty.ai/model/sference/c…
6
403
We ran 22,000 decisions through 3 decision models and 8 LLMs.
On human-labeled gateway tasks:
Clef (27B): 83.2% at $0.13 per 1K items
Opus 5.5: 86.6% at $3.80 per 1K items
Check it here: requesty.ai/blog/decision-mo…
1
1
6
370
On Claude Opus 5.5, a cache hit costs $0.20 per million input tokens. A miss costs $4.
That's 20×. Agents resend the whole conversation every turn, so hits vs misses decides the bill.
How caching is priced, what breaks it, and how to keep it warm ↓
3
7
428
How fast do tokens move to a new model?
Share of tokens on the new model vs. its predecessor, first 72 hours on Requesty:
- GPT-6 Sol = 49%
- Claude Fable 5.1 = 43%
- GPT-6 Luna = 35%
- Claude Opus 5.5 = 23%
- DeepSeek V4.1 Flash = 23%
- Claude Sonnet 5.5 = 5%
1
228
Same model. Very different latency.
claude-opus-5-5 p50 by endpoint today:
> vertex@us 1.32s
> bedrock 17.9s
The fastest endpoint changed 13 times in 24h. Pin one and you inherit its bad hours.
requesty.ai/rankings/latency…
1
6
237
The best model changes every couple of weeks.
Mozart AI wanted switching to be a config change, not an engineering project.
77 models and 17 providers later, it's still one endpoint on Requesty.
How they built it: requesty.ai/customers/mozart
1
145
Introducing the Requesty CLI.
One command to run Claude Code, Codex, OpenCode, Pi and Hermes through Requesty, on 600+ models, with every agent's spend in one dashboard.
Check it out here: docs.requesty.ai/integration…
1
3
305
Jev on Requesty, share of active orgs that have sent a request:
18 Sep → 1.2%
21 Sep → 2.4%
24 Sep → 3.6%
28 Sep → 4.0%
Roughly 1 in 25 orgs so far. TypeSafe's decision model, one model string away.
docs.requesty.ai/features/de…
1
4
223
New model launches, measured in teams rather than tokens.
Share of Requesty customers using each model 7 days after launch:
DeepSeek V4.1 Flash 11%
Gemini 3.8 Flash 8%
Claude Opus 5 8%
GPT-6 Astra 4%
Claude Fable 5.1 2%
Flash tier onboards about 3x more teams in week one.
229
More AI traffic shouldn't mean another integration project.
Ethira grew monthly requests 4.4x from Nov 2025 to Jun 2026, on the same Requesty integration.
See how the team scaled:
requesty.ai/customers/ethira
136
New @OpenAI GPT 6 Sol and Luna now on Requesty!
GPT-6 Sol
Input: $4 → $2
Cached input: $0.40 → ~$0.20
Output: $20 → $10
GPT-6 Luna
Input: $0.20 → $0.10
Cached input: $0.02 → ~$0.01
Output: $1.20 → $0.50
Lower costs across input, cached input, and output.
Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.
GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.
We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.
3
242
Claude Opus 5.5 is live on Requesty.
$4/M in, $20/M out, down from $5/$25.
Cache reads dropped to $0.20/M. If your agents re-read long contexts, Opus just got a lot cheaper than the sticker price suggests.
Same endpoint, new model string. anthropic/claude-opus-5-5 →
1
3
1,223
We’ll be at @HumanXCo Amsterdam! Find Requesty at booth B22.
Let’s talk AI in production: Building with AI is one thing, running it in production is another.
Come say hello!
1
139
The model announcement is one thing. What happens next is more interesting.
DeepSeek v4.1 Flash climbed from 3.20% to 26.73% of daily token share in our dataset in nine days—then eased back.
Watch the September adoption curve unfold.
117
Meet Jev, @typesafeai 's new decision model. Now on Requesty.
Ask yes/no, multiple-choice or scale questions. Get typed answers and probabilities, not prose.
Start building with typesafe/jev-latest
3
1,043
Same input/output list price. Different bills.
Observed cost per 1M tokens:
>Fable 5.1: $1.87
>Astra: $2.90
Caching saved 82% and 74%, respectively.
Reprice Astra's cache reads from $1 to $0.25/M: its bill would drop 23%.
Cache hits matter. Cache pricing does too.
2
139
Your model is fixed. Its fastest provider isn’t.
Our live latency leaderboard shows how provider leadership changes throughout the day.
See the shifts and let latency routing adapt.
requesty.ai/rankings/latency
129
We've partnered with Lyceum to bring more inference options to teams in Germany and across the EU.
> Kimi K3
> GLM-5.3 & GLM-5.3 Flash
> DeepSeek v4.1 Flash
More choice for teams building with Requesty.
12
1,004
We're at Berlin AI Day today!
At 16:00, our CEO Thibault Jaigu shares what production AI traffic teaches you that benchmarks don't.
16:00–16:30 · Aquarium · Festsaal Kreuzberg
Building with AI? Come say hi to the Requesty team.
buildersklub.com/
2
260
Model choice is a process, not a one-off decision.
Production data from Requesty per org, week 0 → 12:
• Models tried in total: 4.5 → 12.8
• Models used per week: 4.5 → 3.9
Exploration keeps growing. Weekly usage stays near 4.
Build for model choice to change.
159
Your AI policy shouldn’t live in a Confluence page. It should return a 403.
Who → Model → Provider → Region → Budget → Route/Block.
Requesty puts policy on the request path.
Policies become infrastructure and not a reading assignment.
4
259
DeepSeek V4.1 Flash from @deepseek_ai now on Requesty!
With reduced prices from DeepSeek v4 Flash:
Input: $0.44 → $0.30 (-32%)
Cache hit: $0.014 → $0.006 (-57%)
Output: $1.32 → $1.20 (-9%)
USD per 1M tokens.
Plus native image input and 1M-token context.
🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.
🔹 Introducing the smallest model in our new architecture family, with native visual understanding.
🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.
1/6
2
1
19
1,349