Fastest LLM API: Output Speed per Model, Not per Vendor

Speed depends on the model, not just the vendor. Baseten’s Fast tier leads DeepSeek V4.1 Flash 1017 at 837.1 tokens/s with a 0.21 s first chunk and GLM-5.3 at 283.4 tokens/s; CoreWeave leads GLM-5.3-Flash at 292.4 tokens/s, and Cerebras still leads the GPT-OSS 120B baseline at 1,693.7 tokens/s. Novita AI ranks fourth or worse on every current model.

Updated · 18 sources · Methodology

NeedPickWhy
Fastest · GLM-5.3Baseten (Fast)283.4 tokens/s, 0.90 s first chunk1
Fastest · DeepSeek V4.1 Flash 1017Baseten (Fast)837.1 tokens/s, 0.21 s first chunk2
Fastest · GLM-5.3-FlashCoreWeave292.4 tokens/s, 1.38 s first chunk3
Fastest · GPT-OSS 120B baselineCerebras1,693.7 tokens/s, 0.38 s first chunk4
Shortest first chunk · DeepSeek V4.1 Flash 1017Baseten (Fast)0.21 s, against 0.27 s on Baseten’s standard tier2
Highest accuracy · GLM-5.3Z.ai98.74%, 3.12 points above Novita AI118

How to read these rankings

All values are Artificial Analysis (AA) provider medians on a 10,000-token input: output speed in tokens per second and time to first chunk, which excludes reasoning time. AA benchmarks each model at its own headline effort — Max for GLM-5.3 and DeepSeek V4.1 Flash 1017, High for GPT-OSS 120B, a single config for GLM-5.3-Flash — so providers are ranked within a model and never across models. Fireworks has no AA row for GLM-5.3 despite selling it, so it appears only where AA measures it.

GLM-5.3 output speed

Six named hosts, one clear leader. Baseten’s Fast tier is a separate, higher-priced SKU at $2.10/$6.60, and it is not comparable to the $1.40/$4.40 standard routes on cost.

Swipe to see all columns

ProviderOutput speed (tokens/s)First chunk (s)Endpoint Accuracy Index v1.0Price in / out per 1M
Baseten (Fast)15283.40.9094.18%$2.10 / $6.605
Together AI116164.30.6796.39%$1.40 / $4.4016
Baseten15150.42.6095.54%$1.40 / $4.405
Novita AI16105.52.5195.62%$1.40 / $4.406
Z.ai11898.22.3998.74%$1.40 / $4.4018
DeepInfra11797.10.9798.6%$0.90 / $4.0017
Output speed by provider · GLM-5.3 (tokens/s)
Baseten (Fast)283.4
Together AI164.3
Baseten150.4
Novita AI105.5
Z.ai98.2
DeepInfra97.1
Bars are scaled inside each model; the fastest host of one model is not comparable to the fastest host of another because AA’s effort settings differ.

Endpoint Accuracy Index is published only for GLM-5.3 and GPT-OSS 120B among these four models; DeepSeek V4.1 Flash 1017 and GLM-5.3-Flash have no AA accuracy value yet.

DeepSeek V4.1 Flash 1017 output speed

Seven hosts and the widest spread of the three current models: the fastest is 12.8× the slowest, all on the same model and the same AA config.

Swipe to see all columns

ProviderOutput speed (tokens/s)First chunk (s)Endpoint Accuracy Index v1.0Price in / out per 1M
Baseten (Fast)25837.10.21—$0.60 / $2.405
Together AI216365.00.32—$0.30 / $1.2016
Baseten25281.10.27—$0.30 / $1.205
Fireworks215210.41.15—$0.30 / $1.2015
DeepSeek211169.01.19—$0.30 / $1.2011
Novita AI27158.01.43—$0.30 / $1.207
DeepInfra21765.30.72—$0.20 / $0.6017
Output speed by provider · DeepSeek V4.1 Flash 1017 (tokens/s)
Baseten (Fast)837.1
Together AI365.0
Baseten281.1
Fireworks210.4
DeepSeek169.0
Novita AI158.0
DeepInfra65.3
Bars are scaled inside each model; the fastest host of one model is not comparable to the fastest host of another because AA’s effort settings differ.

Endpoint Accuracy Index is published only for GLM-5.3 and GPT-OSS 120B among these four models; DeepSeek V4.1 Flash 1017 and GLM-5.3-Flash have no AA accuracy value yet.

GLM-5.3-Flash output speed

Seven hosts, and the only current model where AA measures CoreWeave. CoreWeave’s 292.4 tokens/s is the fastest of the seven hosts in this table.

Swipe to see all columns

ProviderOutput speed (tokens/s)First chunk (s)Endpoint Accuracy Index v1.0Price in / out per 1M
CoreWeave314292.41.38—$0.15 / $0.5014
Together AI316190.10.70—$0.15 / $0.5016
Fireworks315167.23.16—$0.15 / $0.5015
Baseten35147.40.41—$0.15 / $0.505
Novita AI3894.34.00—$0.15 / $0.508
Z.ai31873.52.84—$0.15 / $0.5018
DeepInfra31757.20.83—$0.15 / $0.5017
Output speed by provider · GLM-5.3-Flash (tokens/s)
CoreWeave292.4
Together AI190.1
Fireworks167.2
Baseten147.4
Novita AI94.3
Z.ai73.5
DeepInfra57.2
Bars are scaled inside each model; the fastest host of one model is not comparable to the fastest host of another because AA’s effort settings differ.

Endpoint Accuracy Index is published only for GLM-5.3 and GPT-OSS 120B among these four models; DeepSeek V4.1 Flash 1017 and GLM-5.3-Flash have no AA accuracy value yet.

GPT-OSS 120B baseline: where Groq, Cerebras and SambaNova still rank

Groq, Cerebras and SambaNova serve none of the three current models above, which is why the 2025 GPT-OSS 120B baseline stays: it is the only model where AA measures all three.

Swipe to see all columns

ProviderOutput speed (tokens/s)First chunk (s)Endpoint Accuracy Index v1.0Price in / out per 1M
Cerebras4121,693.70.3887%$0.35 / $0.7512
SambaNova413701.21.1898%$0.22 / $0.5913
Groq410473.40.6386%$0.15 / $0.6010
Novita AI49238.90.9291%$0.05 / $0.259
Together AI416169.40.5377%$0.15 / $0.6016
CoreWeave41498.31.4098%$0.03 / $0.1714
DeepInfra41756.60.6297%$0.037 / $0.1717
Output speed by provider · GPT-OSS 120B (tokens/s)
Cerebras1,693.7
SambaNova701.2
Groq473.4
Novita AI238.9
Together AI169.4
CoreWeave98.3
DeepInfra56.6
Bars are scaled inside each model; the fastest host of one model is not comparable to the fastest host of another because AA’s effort settings differ.

Provider by provider across the current models

Cards are ordered by each provider’s best output speed on any of the three current models. A dash means the provider is not measured on that model; Groq, Cerebras and SambaNova appear only in the GPT-OSS 120B baseline.

  1. Baseten

    Fastest on GLM-5.3 and DeepSeek V4.1 Flash 1017 through its higher-priced Fast tier.1235

    GLM-5.3 tokens/s
    150.4
    DeepSeek V4.1 Flash 1017 tokens/s
    281.1
    GLM-5.3-Flash tokens/s
    147.4

    Strengths

    Its Fast tier runs GLM-5.3 at 283.4 tokens/s with a 0.90 s first chunk and DeepSeek V4.1 Flash 1017 at 837.1 tokens/s with 0.21 s, the quickest start measured on any current model here.1235

    Limitations

    The Fast tier costs more than the standard one — $2.10/$6.60 against $1.40/$4.40 on GLM-5.3 — and it scores lower on Endpoint Accuracy (94.18 against 95.54 on the standard tier).1235

    1235

  2. Together AI

    The most consistent host here: second on output speed on all three current models.12316

    GLM-5.3 tokens/s
    164.3
    DeepSeek V4.1 Flash 1017 tokens/s
    365.0
    GLM-5.3-Flash tokens/s
    190.1

    Strengths

    It runs GLM-5.3 at 164.3 tokens/s with the shortest first chunk on that model at 0.67 s, and DeepSeek V4.1 Flash 1017 at 365.0 tokens/s.12316

    Limitations

    It never leads on output speed: Baseten’s Fast tier is faster on GLM-5.3 and DeepSeek V4.1 Flash 1017, and its Endpoint Accuracy on GLM-5.3, 96.39, sits below DeepInfra and Z.ai.12316

    12316

  3. CoreWeave

    The fastest of the seven GLM-5.3-Flash hosts here, and the only measurement CoreWeave gets on a current model.314

    GLM-5.3 tokens/s
    —
    DeepSeek V4.1 Flash 1017 tokens/s
    —
    GLM-5.3-Flash tokens/s
    292.4

    Strengths

    Its 292.4 tokens/s on GLM-5.3-Flash is 1.5× Together AI’s 190.1 and 3.1× Novita AI’s 94.3, on the model’s only AA-measured route at CoreWeave.314

    Limitations

    It has no AA row for GLM-5.3 or DeepSeek V4.1 Flash 1017, and its first chunk on GLM-5.3-Flash, 1.38 s, is 3.4× Baseten’s 0.41 s.314

    314

  4. Fireworks

    Mid-pack on both models AA measures it on, with a long first chunk on GLM-5.3-Flash.2315

    GLM-5.3 tokens/s
    —
    DeepSeek V4.1 Flash 1017 tokens/s
    210.4
    GLM-5.3-Flash tokens/s
    167.2

    Strengths

    It runs DeepSeek V4.1 Flash 1017 at 210.4 tokens/s, ahead of DeepSeek itself at 169.0 and Novita AI at 158.0.2315

    Limitations

    Its GLM-5.3-Flash first chunk of 3.16 s is second only to Novita AI’s 4.00 s among the seven measured hosts, and AA has no GLM-5.3 row for it at all.2315

    2315

  5. Novita AI

    Measured on all four models, but fourth to sixth on output speed in each ranking.123678

    GLM-5.3 tokens/s
    105.5
    DeepSeek V4.1 Flash 1017 tokens/s
    158.0
    GLM-5.3-Flash tokens/s
    94.3

    Strengths

    It is one of three hosts measured on all four models, with Together AI and DeepInfra, and its 238.9 tokens/s on GPT-OSS 120B puts it fourth of seven there.123678

    Limitations

    It runs GLM-5.3 at 105.5 tokens/s against Baseten’s Fast tier at 283.4, DeepSeek V4.1 Flash 1017 at 158.0 against 837.1, and GLM-5.3-Flash at 94.3 with the model’s longest first chunk at 4.00 s.123678

    123678

  6. Z.ai

    Near the back on GLM-5.3 speed but the most accurate: the highest Endpoint Accuracy Index on that model.1318

    GLM-5.3 tokens/s
    98.2
    DeepSeek V4.1 Flash 1017 tokens/s
    —
    GLM-5.3-Flash tokens/s
    73.5

    Strengths

    Its Endpoint Accuracy Index of 98.74% on GLM-5.3 is the highest of the six measured hosts, 3.12 points above Novita AI.1318

    Limitations

    It runs GLM-5.3 at 98.2 tokens/s, fifth of six and 60% of Together AI’s 164.3, and takes 2.39 s to the first chunk.1318

    1318

  7. DeepInfra

    The most accurate host on GLM-5.3 after Z.ai, and the slowest on DeepSeek V4.1 Flash 1017.12317

    GLM-5.3 tokens/s
    97.1
    DeepSeek V4.1 Flash 1017 tokens/s
    65.3
    GLM-5.3-Flash tokens/s
    57.2

    Strengths

    Its GLM-5.3 Endpoint Accuracy Index of 98.6% is second only to Z.ai, and its 0.97 s first chunk is the third-shortest on that model.12317

    Limitations

    Its 65.3 tokens/s on DeepSeek V4.1 Flash 1017 is the slowest of the seven hosts, 12.8× behind Baseten’s Fast tier, and its list price of $0.20/$0.60 on that model carries a current promotion at $0.098/$0.294.12317

    12317

Methodology and sources

Speed and first-chunk values are Artificial Analysis provider medians from one snapshot per model, all on a 10,000-token input; first chunk excludes reasoning time so it measures queueing and prefill, not thinking. Endpoint Accuracy Index v1.0 is expressed relative to a reference endpoint and exists only for GLM-5.3 and GPT-OSS 120B here. Prices are the providers’ standard serverless input/output rates, excluding cache discounts.

Medians are not guarantees: a single host can move several positions between snapshots, and geo-routing, batch tiers and peak load all shift the numbers. This page ranks measured hosts only and makes no vendor speed claim that AA does not support. Novita AI publishes it from public sources without a private workload benchmark.

Providers are ranked inside each model from the AA output-speed median, never across models, because AA’s headline effort differs by model: Max for GLM-5.3 and DeepSeek V4.1 Flash 1017, High for GPT-OSS 120B and one config for GLM-5.3-Flash. Ties in speed keep the AA order. Cards follow each provider’s best speed on any current model, so a host that leads one model can sit above a more consistent one.1234

  1. Artificial Analysis · GLM-5.3 providers
  2. Artificial Analysis · DeepSeek V4.1 Flash 1017 providers
  3. Artificial Analysis · GLM-5.3-Flash providers
  4. Artificial Analysis · GPT-OSS 120B providers
  5. Baseten · pricing
  6. Novita AI · GLM-5.3
  7. Novita AI · DeepSeek V4.1 Flash 1017
  8. Novita AI · GLM-5.3-Flash
  9. Novita AI · GPT-OSS 120B
  10. Groq · GPT-OSS 120B
  11. DeepSeek · API pricing
  12. Cerebras · pricing
  13. SambaNova · model token prices
  14. CoreWeave · Forge token prices
  15. Fireworks · serverless pricing
  16. Together AI · pricing
  17. DeepInfra · GLM-5.3
  18. Z.ai · pricing

Build with Novita Model APIs

Create an API key and connect your application.

Start BuildingExplore Models