Baseten: Elephant 1

Baseten

•

Elephant 1

•

Released

October 2026

Baseten
Elephant 1

A 9B open-weight decision model that returns calibrated probabilities for routing, classification and scoring instead of generating text.

Modality:

Text

model ID

baseten/elephant-1

Start for free

View API Docs

Output Speed *

N/A

tok/s

Intelligence Index *

N/A

/ 100

Context Window *

16384

tokens

Input price

0.21

Anytoken

Output price

Anytoken

Calibrated Decisions in One Forward Pass, Without Generating Text elephant-1 is a 9B decision model from Base Labs, Baseten's research team, post-trained from Qwen3.5-9B. It is the smaller of two decision models in the lineup, next to the 27B mammoth-1. Rather than writing text, it reads a state and typed questions, then returns a calibrated probability for every answer option. That makes elephant-1 a fit for ticket routing, classification, moderation and confidence-gated automation, where an application needs a score to act on instead of prose to parse. Access elephant-1 via AnyAPI.ai

Performance

Why a Single Forward Pass Matters for Routing and Scoring

elephant-1 answers every question about a state from one shared-prefix prefill and returns a probability per option, so there is no generated text to parse. Baseten claims decision models can be up to 10x faster than reasoning models, which is a vendor figure. A third-party tracker citing Artificial Analysis reports a median of 165 ms per task through Baseten's API. In production, this suits high-volume triage where a confidence threshold decides between automation and human escalation. Knowledge-heavy questions are the weak spot.

Benchmarks

elephant-1 Decision Index Results: Strong Judgment, Weak Knowledge

On the Artificial Analysis Decision Index v1.0, elephant-1 scored 64.69, third of 17 decision models. It sat behind d1-ultra and SubQ Snap 1.1 and ahead of Jev 1.14 at 57.07, per a third-party tracker. Knowledge is the weak area: 11.6 on Omniscience accuracy, though non-hallucination is 93.2. On Jev Decision Index 0.4.1 its Full score is 60.26, 40th of 229, well below the 72.24 public score Base Labs submitted. Validate on your own data.

Output Speed

N/A

tok/s

Intelligence Index

N/A

/ 100

MMLU *

Broad world knowledge and problem-solving

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.

HLE *

Adherence to multi-step structured instructions.

LiveCodeBench *

Tool-calling reliability in long agentic loops.

Technical Specifications

What the model supports

elephant-1 takes a text or JSON state, plus images according to the model card, and typed yes/no, choice or score questions. It returns probabilities rather than text, so a conventional output-token limit does not apply. The 16,384-token input cap is the specification with the biggest production impact. The model card says longer requests are refused, not truncated, so long documents must be chunked or condensed upstream. BenchLM lists 262,144 positions from the released config, but the model card limit is the operating constraint.

Verified Specifications —

Elephant 1

Input modalities

Text

output modalities

Text

Context window

16384

tokens

Maximum output tokens

14745

Reasoning

Knowledge cutoff

October 2026

Pricing (standard)

0.21

AnyTokens in

AnyTokens out

Comparison

elephant-1 or mammoth-1 for Decision Workloads?

mammoth-1 is the 27B sibling, post-trained from Qwen3.5-27B with the same recipe, the same typed question formats and the same 16K input limit. On Base Labs' figures, mammoth-1 leads the Jev Decision Index 0.4 at 71.24 against 62.65 for elephant-1. The two are close on the Artificial Analysis replica (65.9 versus 65.2). The practical decision is whether broader general knowledge justifies the larger model, or whether quick, lightweight decisions matter more.

Dimension

Elephant 1

Mammoth-1

Context window *

16384

tokens

16384

tokens

Output speed *

N/A

tok/s

N/A

tok/s

Intelligence Index *

N/A

Input pricing

0.21

AnyToken

0.6

AnyToken

Output pricing

AnyToken

Knowledge cutoff *

October 2026

Choose elephant-1 when decisions are high-volume and need to be quick, such as routing, triage, moderation or guardrails, or when you want a 9B model that self-hosts on one CUDA GPU. Choose mammoth-1 when the task needs broader general knowledge, which Baseten recommends it for, or when the Jev margin shows up on your own labeled data. Neither generates text, so pair either with a generative model for drafting or explanation.

Limitations & Trade-offs

Where Elephant 1 falls short

No text generation. elephant-1 outputs probabilities over predefined options and writes no text. It cannot draft replies, explain a decision or extract free-form fields. Workloads that need generated answers, summaries or open-ended reasoning require a generative model, with elephant-1 acting only as a router or gate in front of it.

16,384-token input cap. The model card sets a maximum sequence length of 16,384 tokens and says longer requests are refused rather than truncated. Long contracts, transcripts or large retrieval payloads must be chunked or condensed before a decision call. Teams deciding over whole long documents should pre-process inputs or choose a long-context model.

Weak general knowledge. A third-party tracker reports Artificial Analysis sub-scores of 11.6 on Omniscience accuracy and 18.3 on HLE multiple choice. Decisions that depend on world knowledge outside the supplied state are risky. Baseten points tasks needing broader knowledge to the larger mammoth-1.

Benchmark gap and overlap caveats. Base Labs' submitted Jev public score of 72.24 sits 12 points above the Full score of 60.26 on the same index. The Jev index also records that training included train splits of Decision Index benchmarks, as the authors declared. Validate accuracy and calibration on your own labeled data before setting confidence thresholds.

Best-Fit Workloads

Where this model earns its place

Support ticket routing

‍
elephant-1 is built for this pattern: Baseten's own example sends a ticket with a team-selection question and gets a probability per team. One request can carry several questions, such as team, severity and refund intent, answered from a single prefill. It suits high-volume inboxes. Tickets over 16,384 tokens need trimming first.

Confidence-gated automation

‍
Baseten describes accepting an answer automatically at a set confidence, for example 90%, and escalating lower-confidence cases to a person. elephant-1 returns calibrated probabilities with a trained "I don't know" option, so thresholds can be tuned per action risk. Calibration should be checked on your own data before relying on a threshold.

Moderation and agent guardrails

‍
Baseten positions both decision models for safety checks on requests. A third-party tracker reports Artificial Analysis sub-scores of 93.3 on WildGuardTest F1 and 94.8 on When2Call. As an agent guardrail or judge, elephant-1 returns a score to branch on rather than text to parse. Edge cases that depend on outside knowledge may warrant mammoth-1 or human review.

Rubric and sentiment scoring

‍
Score questions return a probability per level on a scale of 2–10 levels, plus an expected level and a confidence. Base Labs' example asks how upset a customer is on a calm-to-furious scale. This fits content scoring, triage priority and quality review at volume. It will not justify its scores in prose.

Pricing in anytokens via AnyAPI

Input

0.21

₳

Output

₳

Cache write

—

₳

Cache read

—

₳

Get API access

Read Docs

Integration

Access Elephant 1 via AnyAPI.ai

Access Elephant 1 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Elephant 1 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

One API integration

Access Elephant 1 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

Easy model switching

Test Elephant 1 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

Flexible for production

Use Elephant 1 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

Multi-model applications

Use Elephant 1 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free

Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

What is elephant-1?

elephant-1 is a 9B decision model from Base Labs, Baseten's research team, released October 21, 2026. It is post-trained from Qwen3.5-9B on 6.21 million decision examples. Given a state and typed questions, it returns a calibrated probability for every answer option in one forward pass. Its weights are open under the MIT license on Hugging Face.

Does elephant-1 generate text like a chat model?

No. elephant-1 does not generate text. It returns a probability distribution over the options of each question, such as choice, yes/no or score, along with a confidence value. This avoids decoding and parsing free-form output. Use a generative model when you need written responses, summaries or explanations.

What is the elephant-1 context window?

elephant-1 has a 16,384-token maximum sequence length according to its model card. Requests longer than that are refused rather than truncated. Because the model returns probabilities rather than text, there is no conventional maximum output token setting. Split or summarize long documents before sending them for a decision.

elephant-1 vs mammoth-1: which should I use?

Use elephant-1 for quick, high-volume decisions such as routing, triage and moderation. Use mammoth-1, the 27B sibling, when the task needs broader general knowledge. Base Labs reports 71.24 for mammoth-1 versus 62.65 for elephant-1 on the Jev Decision Index 0.4, with a near-tie on the Artificial Analysis replica (65.9 versus 65.2).

How much does the elephant-1 API cost?

Baseten offered elephant-1 free for a launch week that ended October 27, 2026. A third-party tracker notes Baseten has not published pricing for after that week. Because elephant-1 produces no generated text, cost per decision is driven mainly by input size. See the AnyTokens pricing component on this page for current AnyAPI.ai pricing.

* Benchmark data source: Artificial Analysis artificialanalysis.ai