Switch reasoning off for simple routing and extraction, and keep it on for harder tasks. In our test calls, DeepSeek V4.1 Flash answered in 0.5 seconds with reasoning off and in up to 1.5 seconds with it on. Without reasoning, its quality score falls from 29 to 19.
Which open-weight model can replace Claude Haiku, GPT-5 nano or GPT-5.4 mini for high-volume, low-latency tasks? It comes up when a small closed model handles millions of short requests and you want a model you can move between providers.
What teams use Claude Haiku, GPT-5 nano and GPT-5.4 mini for
Small closed models do the high-volume work in a product. They route support tickets, pull fields out of documents, label content and write short replies. Each request is short, the answer is a few dozen tokens, and the user is often waiting for it.
An open-weight model is one whose trained weights are published, so any provider can run the same model. If you move providers, the model and its behaviour stay the same. The gap to closed models has also become small. Since January 2026, the most capable open-weight models have trailed frontier closed models by an average of four months. So the best open model today is about where the best closed model was four months ago.
We are Lyceum, a sovereign European AI inference provider for open-source models: a European company that runs open models on GPUs in European data centres. Our serverless models run in Paris and Finland, and you call them through one OpenAI-compatible API.
For GPT-5 nano there is also a date. OpenAI removes gpt-5-nano from its API on 11 December 2026 and told developers on 11 June 2026.
Open-weight models that score near Claude Haiku 5.5
The Artificial Analysis Intelligence Index is a combined quality score built from 9 evaluations; we use its current version, v5.0.1. Effort is how long a model may think before it answers: low, high or max. The index scores Claude Haiku 5.5 at 30 with high effort. Two of the small open models we serve land within two points of that.
| Model | Index score | Input | Cached input | Output | Reasoning | Licence |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 29 (high), 33 (max), 19 (off) | $0.50 | $0.13 | $1.50 | Can be switched off | MIT |
| GLM-5.3 Flash | 28 | $0.20 | $0.05 | $0.50 | Keep on | MIT |
| DeepSeek V4 Flash | not scored | $0.25 | $0.06 | $0.30 | On by default | MIT |
| MiniMax M3 | 19 | $0.40 | $0.10 | $2.00 | Always on | MiniMax Community License |
Prices in US dollars per 1M tokens from our model pages, scores from Artificial Analysis, both read on 5 November 2026. For reference, Claude Haiku 5.5 scores 36 at max effort, 30 at high and 19 at low.
DeepSeek V4.1 Flash scores 29 at high effort and 33 at max. DeepSeek updated the model in October, and the index tests that update. Our DeepSeek V4.1 Flash model ID served the same October update in our test calls, so the score applies to it. GLM-5.3 Flash scores 28 and costs less on both input and output.
DeepSeek V4 Flash is the older, cheaper DeepSeek model. The current index has no score for it, so test it against the other two if price matters most.
MiniMax M3 scores 19, close to Haiku 5.5 at low effort. It reads long documents and images, with a 1M-token context, but its reasoning cannot be switched off.
A score is a starting point, not a verdict. It says which models are worth testing first, and your own tasks decide.
Replacing GPT-5 nano and GPT-5.4 mini without a current benchmark
Artificial Analysis marks GPT-5 nano and GPT-5.4 mini as deprecated and no longer gives them a current index score. So no independent number sets an open model next to either of them today. Anyone who names a "closest" model here is guessing.
Choose by the kind of task, then test.
- Short structured answers (routing, labels, field extraction): start with GLM-5.3 Flash, or with DeepSeek V4.1 Flash and its reasoning switched off.
- Harder tasks (multi-step instructions, short reasoning): start with DeepSeek V4.1 Flash with reasoning on.
Do you call Haiku through Claude Code? Our Anthropic-compatible endpoint accepts Claude Code's requests and answers them with our open models. Claude Code asks for a tier (Opus, Sonnet or Haiku), and our endpoint sends the Haiku tier to MiniMax M3 by default. That default does not follow the index above. Settings named ANTHROPIC_DEFAULT_* pick the model for each tier, so you can point Haiku at GLM-5.3 Flash. Our Claude Code setup guide shows how.
How the reasoning setting changes latency and cost at high volume
Reasoning is a thinking step the model runs before its answer. Its tokens count as output tokens, so they cost money and time. For a ticket label, the thinking can be longer than the answer.
We tested this on 5 November 2026. A system prompt asked each model to route a support ticket and reply with JSON only. We sent three tickets per setting, one call each, at temperature 0. Every setting routed all three tickets to the same queues.
Then we priced the result for a hypothetical support tool: 1,000,000 tickets a month, each with 500 input tokens. Output per ticket is the average from our three calls.
monthly cost = input tokens x input rate + output tokens x output rate (rates per 1M tokens)
DeepSeek V4.1 Flash, reasoning off: 500M x $0.50 + 13M x $1.50 = $250.00 + $19.50 = $269.50| Model and setting | Output tokens per ticket | Full answer, our test | Cost per month |
|---|---|---|---|
| DeepSeek V4 Flash, default settings | 80 | 0.7 to 0.8 s | $149.00 |
| GLM-5.3 Flash, reasoning on | 129 | 1.3 to 1.8 s | $164.50 |
| DeepSeek V4.1 Flash, reasoning off | 13 | 0.5 s | $269.50 |
| MiniMax M3, reasoning always on | 64 | 1.6 to 7.3 s | $328.00 |
| DeepSeek V4.1 Flash, reasoning on | 54 | 0.8 to 1.5 s | $331.00 |
Hypothetical workload, calculated from our list prices on 5 November 2026 and the average output of three test calls per setting.
Input is 500M tokens in every row, and output is at most 129M. So the input rate sets most of the bill. Reasoning sets the answer time. Switching it off on DeepSeek V4.1 Flash cut output from 54 to 13 tokens and the answer to half a second.
Do not switch reasoning off on GLM-5.3 Flash. GLM-5.3 Flash Instant is the same model with reasoning switched off. In the same test, it began all three answers with prose, not the JSON the prompt asked for. One began "The user is reporting a billing issue". Our model page warns about this: with reasoning off, the model writes its reasoning into the answer.
The example does not show quality on your tickets. Three tickets is a smoke test, not a benchmark. It also leaves out caching. A system prompt that repeats on every ticket is billed at the lower cached input rate, which would cut every row.
Latency under load is on our public status page. Time to first token is the wait before the first word arrives. Over the last 7 days, its median was 366 ms for DeepSeek V4.1 Flash and 363 ms for DeepSeek V4 Flash. GLM-5.3 Flash was at 500 ms and MiniMax M3 at 1,103 ms. Our guide on what drives time to first token explains the gap.
Testing a replacement on your own tasks
Pick 50 real requests from your logs, including the hard ones. Send them to DeepSeek V4.1 Flash and GLM-5.3 Flash. Run DeepSeek V4.1 Flash twice, once with reasoning on and once with it off.
This call runs one request with reasoning off. Drop the reasoning_effort line for the run with reasoning on.
from openai import OpenAI
client = OpenAI(base_url="https://api.lyceum.technology/openai/v1", api_key="lk_...")
resp = client.chat.completions.create(
model="deepseek/deepseek-v4.1-flash",
messages=[
{"role": "system", "content": "Route the ticket. Reply with JSON only."},
{"role": "user", "content": "The export button returns a 500 error."},
],
reasoning_effort="none",
)
print(resp.choices[0].message.content, resp.usage.completion_tokens)Check three things for each run: whether the answer is right, whether it parses, and how long it took. The output token count in the usage field is what you pay for.
To compare answers on open-ended tasks, use a second model as the grader. Our guide on comparing model outputs before a provider switch shows the setup. Then price the winner with your own token counts, as in the table above, or against the full price comparison of open models.
Switch when the open model gets your hard cases right with the setting you chose. Create an API key in our dashboard and run your 50 requests today.
Frequently asked questions
Can I keep Claude Code and change only the model behind the Haiku tier?
Yes. Point Claude Code at our Anthropic-compatible endpoint. The Haiku tier maps to MiniMax M3 by default, and the ANTHROPIC_DEFAULT entries in your settings choose another of our models for each tier.
What licences do these open-weight models carry?
DeepSeek V4.1 Flash, DeepSeek V4 Flash and GLM-5.3 Flash are under the MIT licence. MiniMax M3 is under the MiniMax Community License, MiniMax's own licence. Read its terms before you build on it.
Do reasoning tokens show up on the bill?
Yes. The reasoning step's tokens count as output tokens. Set max_tokens high enough to cover the reasoning and the answer, or the answer can come back empty.