Model Fallback

Overview

Model fallback enables specifying multiple models in a models array. The API tries each model in order until one succeeds, providing automatic failover when a model is unavailable.

How It Works

Provide a models array containing up to 5 models:

The API tries the first model in the array
If it fails or is unavailable, the next model is tried
This continues until one succeeds or all models are exhausted

The models array takes precedence over the single model field when both are provided.

Benefits:

Higher availability: Automatic failover when primary model is unavailable
Provider redundancy: Use models from different providers for maximum reliability
Seamless operation: No code refactoring needed, fallback is handled automatically by the API

Basic Example

from perplexity import Perplexity

client = Perplexity()

response = client.responses.create(
    models=["openai/gpt-5.6-sol", "anthropic/claude-sonnet-5-5", "google/gemini-3.8-flash"],
    input="Explain the original Transformer architecture from 'Attention Is All You Need' (Vaswani et al. 2017): encoder-decoder structure, multi-head self-attention, and positional encodings.",
    instructions="You have access to a web_search tool. Use it for questions about current events.",
    max_output_tokens=8192,
)

print(f"Model used: {response.model}")

import Perplexity from '@perplexity-ai/perplexity_ai';

const client = new Perplexity();

const response = await client.responses.create({
    models: ["openai/gpt-5.6-sol", "anthropic/claude-sonnet-5-5", "google/gemini-3.8-flash"],
    input: "Explain the original Transformer architecture from 'Attention Is All You Need' (Vaswani et al. 2017): encoder-decoder structure, multi-head self-attention, and positional encodings.",
    instructions: "You have access to a web_search tool. Use it for questions about current events.",
    max_output_tokens: 8192,
});

console.log(`Model used: ${response.model}`);

curl https://api.perplexity.ai/v1/agent \
  -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "models": ["openai/gpt-5.6-sol", "anthropic/claude-sonnet-5-5", "google/gemini-3.8-flash"],
    "input": "Explain the original Transformer architecture from \u0027Attention Is All You Need\u0027 (Vaswani et al. 2017): encoder-decoder structure, multi-head self-attention, and positional encodings.",
    "instructions": "You have access to a web_search tool. Use it for questions about current events.",
    "max_output_tokens": 8192
  }'

Include max_output_tokens whenever a fallback chain contains an anthropic/* model. Anthropic requests require it even though non-Anthropic providers accept requests without it.

Response

{
  "id": "resp_157b10b1-4685-4f11-b372-382b8a04e4dd",
  "created_at": 1779391718,
  "model": "openai/gpt-5.6-sol",
  "object": "response",
  "output": [
    {
      "results": [
        {
          "id": 1,
          "snippet": "To the best of our knowledge, however, the Transformer is the first transduction model relying\n...\nThe Transformer follows this overall architecture using stacked self-attention and point-wise, fully...",
          "title": "[PDF] Attention is All you Need - NIPS",
          "url": "https://papers.neurips.cc/paper/7181-attention-is-all-you-need.pdf",
          "date": null,
          "last_updated": "2026-05-20",
          "source": "web"
        },
        {
          "id": 2,
          "snippet": "We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.",
          "title": "[1706.03762] Attention Is All You Need - arXiv",
          "url": "https://arxiv.org/abs/1706.03762",
          "date": "2017-06-12",
          "last_updated": "2026-05-19",
          "source": "web"
        },
        {
          "id": 3,
          "snippet": "{ts:907} encoder which is the positional encoding so what is positional encoding what we want is that each word should\n…\n{ts:963} have this partial information given by our eyes but the model cannot...",
          "title": "Attention is all you need (Transformer) - YouTube",
          "url": "https://www.youtube.com/watch?v=bCz4OMemCcA",
          "date": "2023-05-28",
          "last_updated": "2026-03-31",
          "source": "web"
        },
        {
          "id": 4,
          "snippet": "\"**Attention Is All You Need**\" is a 2017 research paper in machine learning authored by eight scientists and engineers working at Google.\nThe paper introduced a new deep learning architecture known...",
          "title": "Attention Is All You Need - Wikipedia",
          "url": "https://en.wikipedia.org/wiki/Attention_Is_All_You_Need",
          "date": "2023-12-04",
          "last_updated": "2026-05-17",
          "source": "web"
        },
        {
          "id": 5,
          "snippet": "●Presents a new neural architecture named the Transformer\n●Based solely on the attention mechanism widely used in SEQ2SEQ models\n...\nTransformer uses only self-attention \nwhich is attention onto the...",
          "title": "[PDF] Attention Is All You Need",
          "url": "https://ysu1989.github.io/courses/au20/cse5539/Transformer.pdf",
          "date": null,
          "last_updated": "2026-05-15",
          "source": "web"
        },
        {
          "id": 6,
          "snippet": "The Transformer is a neural network architecture, introduced by Vasami *et al.* (2017).\n...\nThe Transformer follows the classic**encoder–decoder ** design for sequence-to-sequence translation.\nThe...",
          "title": "Review of “Attention Is All You Need (Vaswani et al., 2017)”",
          "url": "https://harbisingh.wordpress.com/2025/08/12/review-of-attention-is-all-you-need-vaswani-et-al-2017/",
          "date": "2025-08-12",
          "last_updated": "2026-04-05",
          "source": "web"
        },
        {
          "id": 7,
          "snippet": "We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.\n...\nIn this work we propose the Transformer...",
          "title": "Attention Is All You Need - arXiv",
          "url": "https://arxiv.org/html/1706.03762v7",
          "date": null,
          "last_updated": "2026-05-18",
          "source": "web"
        },
        {
          "id": 8,
          "snippet": "We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.\n...\nIn this work we propose the Transformer...",
          "title": "Attention Is All You Need",
          "url": "https://arxiv.org/html/1706.03762?_immersive_translate_auto_translate=1",
          "date": null,
          "last_updated": "2026-03-23",
          "source": "web"
        },
        {
          "id": 9,
          "snippet": "To the best of our knowledge, however, the Transformer is the first transduction model relying\nentirely on self-attention to compute representations of its input and output without using sequence-...",
          "title": "[PDF] Attention is All you Need - NIPS",
          "url": "https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf",
          "date": null,
          "last_updated": "2026-05-20",
          "source": "web"
        },
        {
          "id": 10,
          "snippet": "In this work we propose the Transformer, a model architecture eschewing recurrence and instead\nrelying entirely on an attention mechanism to draw global dependencies between input and output.\n...\nTo...",
          "title": "Attention is All you Need",
          "url": "https://papers.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf",
          "date": null,
          "last_updated": "2025-02-04",
          "source": "web"
        }
      ],
      "type": "search_results",
      "queries": [
        "Attention is All You Need Vaswani 2017 transformer architecture encoder decoder multi-head self-attention positional encoding"
      ]
    },
    {
      "id": "msg_14924dfc-980f-4339-9b39-d0f6eb987f87",
      "content": [
        {
          "text": "The original Transformer is a stacked encoder–decoder sequence model that replaces recurrence with multi‑head self‑attention plus position‑wise feed‑forward layers, with sinusoidal positional encodings added to token embeddings to supply order information. [web:7][web:9] Multi‑head attention computes several parallel scaled dot‑product attentions using different learned projections of queries, keys, and values, then concatenates and linearly mixes them. [web:7][web:9]\n\n## Encoder–decoder structure\n\nThe Transformer is an encoder–decoder architecture for sequence transduction (e.g., machine translation), built entirely from attention and feed‑forward layers, without RNNs or convolutions. [web:7][web:9] The base model in the paper uses 6 encoder layers and 6 decoder layers, each operating on vectors of dimension \\(d_{\\text{model}} = 512\\). [web:7][web:9]\n\n- **Encoder** layers:\n  - Each encoder layer has:\n    - A multi‑head self‑attention sublayer.\n    - A position‑wise fully connected feed‑forward network. [web:7][web:9]\n  - Each sublayer is wrapped with residual connection and layer normalization (input + sublayer output, then LayerNorm). [web:7][web:9]\n\n- Decoder layers:\n  - Each decoder layer has three sublayers:\n    - Masked multi‑head self‑attention over the decoder’s own previous outputs.\n    - Multi‑head encoder–decoder attention, where queries come from the decoder and keys/values from the encoder outputs.\n    - A position‑wise feed‑forward network. [web:7][web:9]\n  - The self‑attention in the decoder is masked so position \\(i\\) cannot attend to positions \\(> i\\), preserving the autoregressive property. [web:7][web:9]\n\n## Scaled dot‑product attention\n\nThe core building block is scaled dot‑product attention, which maps queries \\(Q\\), keys \\(K\\), and values \\(V\\) to an output:  \n\\[\n\\text{Attention}(Q,K,V) = \\text{softmax}\\left(\\frac{QK^\\top}{\\sqrt{d_k}}\\right)V.\n\\]  \n[web:7][web:9]\n\n- **Queries, keys, values** are all learned linear projections of the input representations in that layer. [web:7][web:9]\n- The dot products \\(QK^\\top\\) produce attention scores, scaled by \\(1/\\sqrt{d_k}\\) to prevent large magnitudes from pushing softmax into very small gradients. [web:7][web:9]\n- The softmax over each query’s scores gives attention weights over positions; these weights form a weighted sum of the value vectors. [web:7][web:9]\n\nIn self‑attention, \\(Q\\), \\(K\\), and \\(V\\) all come from the same sequence representations; in encoder–decoder attention, \\(Q\\) comes from the decoder and \\(K,V\\) from the encoder. [web:7][web:9]\n\n## Multi‑head self‑attention\n\nMulti‑head attention repeats scaled dot‑product attention in parallel with different learned projections, letting the model focus on different relation patterns and positions simultaneously. [web:7][web:9]\n\n- For each of \\(h\\) heads (the paper uses \\(h = 8\\)):\n  - Project inputs linearly to per‑head \\(Q_i, K_i, V_i \\in \\mathbb{R}^{d_k}, \\mathbb{R}^{d_k}, \\mathbb{R}^{d_v}\\) with separate weight matrices. [web:7][web:9]\n  - Compute attention head \\(i\\) as \\(\\text{head}_i = \\text{Attention}(Q_i, K_i, V_i)\\). [web:7][web:9]\n- Concatenate all heads and apply a final linear projection:\n  \\[\n  \\text{MultiHead}(Q,K,V) = \\text{Concat}(\\text{head}_1,\\dots,\\text{head}_h)W^O.\n  \\]  \n  [web:7][web:9]\n\nKey uses:\n\n- Encoder self‑attention: Each position can attend to all positions in the encoder’s previous layer. [web:7][web:9]\n- Decoder masked self‑attention: Each position attends only to positions up to itself, using a mask that sets scores to \\(-\\infty\\) for future positions before softmax. [web:7][web:9]\n- Encoder–decoder attention: Decoder queries attend over all encoder outputs, with keys and values from the encoder and queries from the decoder. [web:7][web:9]\n\nThis design increases the **expressive** power since different heads can specialize to, for example, syntactic dependencies, coreference, or positional patterns. [web:4][web:5]\n\n## Positional encodings\n\nBecause the Transformer has no recurrence or convolution, it adds positional encodings to input (and output) token embeddings so the model can use order information. [web:7][web:4]\n\n- The paper uses **fixed sinusoidal** positional encodings:\n  - For position \\(\\text{pos}\\) and even dimension \\(2i\\):\n    \\[\n    PE_{\\text{pos}, 2i} = \\sin\\left(\\frac{\\text{pos}}{10000^{2i/d_{\\text{model}}}}\\right),\n    \\quad\n    PE_{\\text{pos}, 2i+1} = \\cos\\left(\\frac{\\text{pos}}{10000^{2i/d_{\\text{model}}}}\\right).\n    \\]\n    [web:4][web:7]\n  - Different dimensions use different frequencies, so relative positions correspond to phase shifts, which allows the model to learn relative as well as absolute distance patterns. [web:4]\n\n- Usage:\n  - For each token, the model sums its learned embedding with the positional encoding vector for its index before feeding into the first encoder or decoder layer. [web:7][web:5]\n  - The same scheme is applied to both source and target sequences. [web:7]\n\nThese fixed encodings avoid extra parameters and support sequences longer than seen in training, while providing smooth, distance‑aware position signals to self‑attention. [web:4][web:7]\n\n## Position‑wise feed‑forward networks\n\nEach layer’s attention sublayer is followed by a position‑wise feed‑forward network applied independently to each sequence position. [web:7][web:9]\n\n- The FFN has the form:\n  \\[\n  \\text{FFN}(x) = \\max(0, xW_1 + b_1)W_2 + b_2,\n  \\]\n  with an inner dimension \\(d_{\\text{ff}} = 2048\\) in the base model. [web:7][web:9]\n- The same FFN parameters are shared across positions but differ between layers. [web:7]\n\nTogether, multi‑head attention, sinusoidal positional encodings, and position‑wise feed‑forward layers form the original Transformer encoder–decoder architecture introduced by Vaswani et al. in 2017. [web:7][web:9]",
          "type": "output_text",
          "annotations": [],
          "logprobs": []
        }
      ],
      "role": "assistant",
      "status": "completed",
      "type": "message"
    }
  ],
  "status": "completed",
  "error": null,
  "usage": {
    "input_tokens": 8687,
    "output_tokens": 1599,
    "total_tokens": 10286,
    "cost": {
      "currency": "USD",
      "input_cost": 0.02041,
      "output_cost": 0.03198,
      "total_cost": 0.05632,
      "cache_creation_cost": null,
      "cache_read_cost": 0.00143,
      "tool_calls_cost": 0.0025
    },
    "input_tokens_details": {
      "cache_creation_input_tokens": 0,
      "cache_read_input_tokens": 3584,
      "cached_tokens": 3584
    },
    "tool_calls_details": {
      "search_web": {
        "invocation": 1
      }
    },
    "output_tokens_details": {
      "reasoning_tokens": 0
    }
  },
  "completed_at": 1779391718
}

Cross-Provider Fallback

For maximum reliability, use models from different providers:

from perplexity import Perplexity

client = Perplexity()

response = client.responses.create(
    models=[
        "openai/gpt-5.6-sol",
        "anthropic/claude-sonnet-5-5",
        "google/gemini-3.8-flash"
    ],
    input="What did the SEC v. Sam Bankman-Fried complaint (December 2022) allege about FTX, and what was the outcome of his 2023 criminal trial?",
    max_output_tokens=8192,
)

import Perplexity from '@perplexity-ai/perplexity_ai';

const client = new Perplexity();

const response = await client.responses.create({
    models: [
        "openai/gpt-5.6-sol",
        "anthropic/claude-sonnet-5-5",
        "google/gemini-3.8-flash"
    ],
    input: "What did the SEC v. Sam Bankman-Fried complaint (December 2022) allege about FTX, and what was the outcome of his 2023 criminal trial?",
    max_output_tokens: 8192,
});

curl https://api.perplexity.ai/v1/agent \
  -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "models": [
      "openai/gpt-5.6-sol",
      "anthropic/claude-sonnet-5-5",
      "google/gemini-3.8-flash"
    ],
    "input": "What did the SEC v. Sam Bankman-Fried complaint (December 2022) allege about FTX, and what was the outcome of his 2023 criminal trial?",
    "max_output_tokens": 8192
  }'

Response

{
  "id": "resp_e012062c-3b30-474e-82a9-3f8116442e6f",
  "created_at": 1779391825,
  "model": "openai/gpt-5.6-sol",
  "object": "response",
  "output": [
    {
      "results": [
        {
          "id": 1,
          "snippet": "The Securities and Exchange Commission today charged Samuel Bankman-Fried with orchestrating a scheme to defraud equity investors in FTX Trading Ltd.\n(FTX), the crypto trading platform of which he...",
          "title": "SEC Charges Samuel Bankman-Fried with Defrauding Investors in ...",
          "url": "https://www.sec.gov/newsroom/press-releases/2022-219",
          "date": "2022-12-13",
          "last_updated": "2025-10-23",
          "source": "web"
        },
        {
          "id": 2,
          "snippet": "The SEC’s Complaint alleges that SBF defrauded investors in FTX by concealing (1) that FTX directed billions of dollars of FTX customer assets to Alameda,[2] (2) that FTX provided special privileges...",
          "title": "An Overview of the SEC, SDNY and CFTC Cases Against Sam ...",
          "url": "https://www.pbwt.com/securities-enforcement-litigation-insider/an-overview-of-the-sec-sdny-and-cftc-cases-against-sam-bankman-fried",
          "date": "2022-12-21",
          "last_updated": "2026-05-16",
          "source": "web"
        },
        {
          "id": 3,
          "snippet": "On December 12, 2022, Bankman-Fried was arrested in the Bahamas and extradited to the United States, where he was indicted on seven criminal charges, including wire fraud, commodities fraud...",
          "title": "Sam Bankman-Fried - Wikipedia",
          "url": "https://en.wikipedia.org/wiki/Sam_Bankman-Fried",
          "date": "2021-04-17",
          "last_updated": "2026-05-17",
          "source": "web"
        },
        {
          "id": 4,
          "snippet": "The complaint charges all three defendants with fraud and material misrepresentations in connection with the sale of digital commodities in interstate commerce.\nFurther, the complaint asserts that...",
          "title": "CFTC Charges Sam Bankman-Fried, FTX Trading and Alameda with ...",
          "url": "https://www.cftc.gov/PressRoom/PressReleases/8638-22",
          "date": "2022-12-13",
          "last_updated": "2026-05-21",
          "source": "web"
        },
        {
          "id": 5,
          "snippet": "According to the allegations contained in the Indictment, the evidence offered at trial, and matters included in public filings:\nBANKMAN-FRIED was the founder and chief executive officer of FTX, an...",
          "title": "Samuel Bankman-Fried Sentenced To 25 Years In Prison",
          "url": "https://www.justice.gov/usao-sdny/pr/samuel-bankman-fried-sentenced-25-years-prison",
          "date": "2024-03-28",
          "last_updated": "2026-03-15",
          "source": "web"
        },
        {
          "id": 6,
          "snippet": "Prosecutors said Bankman-Fried had cost customers, investors and lenders over $10 billion by misappropriating billions of dollars to fuel his quest for influence and dominance in the new industry...",
          "title": "FTX founder Sam Bankman-Fried sentenced to 25 years in prison",
          "url": "https://www.marketplace.org/story/2024/03/28/sam-bankman-fried-sentenced-to-25-years-in-prison",
          "date": "2024-03-28",
          "last_updated": "2026-03-15",
          "source": "web"
        },
        {
          "id": 7,
          "snippet": "The SEC complaint that applied from May 2019 through November 2022 described activities at FTX\nand Alameda that were never disclosed to investors or customers.\nThe SEC complaint90 included the...",
          "title": "[PDF] Sam Bankman-Fried's FTX | MIT Sloan",
          "url": "https://mitsloan.mit.edu/sites/default/files/2024-06/Sam%20Bankman-Fried's%20FTX.pdf",
          "date": "2024-01-17",
          "last_updated": "2026-05-16",
          "source": "web"
        },
        {
          "id": 8,
          "snippet": "The defendant is “charged with a wide-ranging scheme to misappropriate billions of dollars of customer funds deposited with FTX and mislead investors and lenders to FTX and to Alameda Research,” a...",
          "title": "Sam Bankman-Fried found guilty on all seven counts - TechCrunch",
          "url": "https://techcrunch.com/2023/11/02/sam-bankman-fried-found-guilty-on-all-seven-counts/",
          "date": "2023-11-02",
          "last_updated": "2026-04-15",
          "source": "web"
        },
        {
          "id": 9,
          "snippet": "The charges broadly covered two categories: stealing the money of customers who put their money into FTX accounts and lying to investors and creditors.",
          "title": "Sam Bankman-Fried sentenced to 25 years in prison for his FTX ...",
          "url": "https://www.opb.org/article/2024/03/28/sam-bankman-fried-sentenced-to-25-years-in-prison-for-his-ftx-crimes/",
          "date": "2024-03-28",
          "last_updated": "2026-05-16",
          "source": "web"
        },
        {
          "id": 10,
          "snippet": "{ts:49} Bankman-Fried and his co-conspirators stole billions of dollars from FTX customers.\nThe indictment accuses Bankman-Fried of misappropriating FTX.com customers' deposits and using those to pay...",
          "title": "Sam Bankman-Fried Faces Fraud Charges, Lawsuits Tied to FTX ...",
          "url": "https://www.youtube.com/watch?v=oALv48BicV0",
          "date": "2022-12-13",
          "last_updated": "2026-01-25",
          "source": "web"
        },
        {
          "id": 11,
          "snippet": "Bankman-Fried, the 31-year old son of two Stanford legal scholars and graduate of Massachusetts Institute of Technology, had pleaded not guilty to charges including wire fraud, securities fraud and...",
          "title": "Sam Bankman-Fried found guilty on all seven criminal fraud counts",
          "url": "https://web.archive.org/web/20231102235813/https:/www.cnbc.com/2023/11/02/sam-bankman-fried-found-guilty-on-all-seven-criminal-fraud-counts.html",
          "date": "2023-11-02",
          "last_updated": "2024-09-14",
          "source": "web"
        },
        {
          "id": 12,
          "snippet": "He was accused of using customer deposits on the cryptocurrency trading platform FTX, the company he founded, to cover losses at his hedge fund, pay off loans and buy lavish real estate, among other...",
          "title": "Disgraced former FTX CEO Sam Bankman-Fried sentenced to 25 ...",
          "url": "https://abcnews.go.com/US/sam-bankman-fried-sentenced-thursday-financial-fraud/story?id=108554809",
          "date": "2024-03-28",
          "last_updated": "2026-02-10",
          "source": "web"
        }
      ],
      "type": "search_results",
      "queries": [
        "SEC v. Sam Bankman-Fried complaint December 2022 allegations FTX",
        "Sam Bankman-Fried criminal trial verdict November 2023 counts outcome",
        "Sam Bankman-Fried sentencing March 2024 length of sentence forfeiture"
      ]
    },
    {
      "id": "msg_fb282ebe-48cc-4585-b6f3-ce4607ee2df7",
      "content": [
        {
          "text": "The December 2022 SEC civil complaint alleged that Sam Bankman‑Fried ran a years‑long scheme to defraud FTX equity investors by secretly diverting billions of dollars of customer funds to Alameda Research, granting Alameda undisclosed special privileges on the FTX platform, and hiding the resulting risks and misuse of funds while publicly touting FTX as safe and well‑controlled.[web:1][web:2] In his 2023 federal criminal trial in New York, a jury found him guilty on all seven counts related to defrauding FTX customers, Alameda lenders, and FTX investors, and in March 2024 he was sentenced to 25 years in prison plus multi‑billion‑dollar forfeiture and restitution orders.[web:5][web:8][web:11]\n\n## SEC complaint allegations\n\nThe SEC’s December 13, 2022 complaint charged Bankman‑Fried with securities fraud under the antifraud provisions of the Securities Act of 1933 and the Securities Exchange Act of 1934.[web:1][web:2] It alleged that since at least May 2019, FTX raised over $1.8 billion from equity investors (about $1.1 billion from roughly 90 U.S. investors) based on false and misleading statements about how customer assets were handled and the risk profile of the business.[web:1]\n\nKey allegations about **FTX** and Alameda in the SEC complaint included:[web:1][web:2][web:7]  \n- Undisclosed diversion of FTX customer funds to Alameda Research, including having customers deposit billions into bank accounts controlled by Alameda or its subsidiaries.  \n- Special treatment for Alameda on the FTX platform, such as an effectively unlimited “line of credit,” exemptions from key risk controls, and code changes (like an “allow negative” balance flag) that let Alameda withdraw customer assets even when it lacked collateral.  \n- Concealment of FTX’s concentrated exposure to Alameda and to overvalued, illiquid assets such as the FTT token, while publicly claiming customer assets were segregated, safe, and not used for FTX’s own investments.  \n- Use of commingled customer funds for venture investments, luxury real estate in the Bahamas, large political donations, and “loans” to insiders, none of which were disclosed to investors.[web:1][web:2]\n\n## 2023 criminal trial verdict\n\nBankman‑Fried was indicted in the Southern District of New York on multiple fraud and conspiracy charges tied to the collapse of FTX and Alameda Research.[web:3][web:8] In November 2023, after a month‑long trial, a federal jury found him guilty on all seven counts, including wire fraud and conspiracy to commit wire fraud on FTX customers and Alameda lenders, as well as securities fraud and commodities fraud conspiracies and money‑laundering conspiracy.[web:8][web:11]\n\nProsecutors argued he orchestrated a “wide‑ranging scheme” to misappropriate billions in FTX customer deposits, using them to cover Alameda losses, repay lenders, make investments, and fund political spending, while lying to investors and lenders about FTX’s financial condition and Alameda’s special access to customer funds.[web:5][web:8][web:9] Evidence described altering FTX code to give Alameda effectively unlimited withdrawals, producing false financial statements for lenders and investors, and making public assurances that customer deposits were safe and segregated when they were not.[web:5][web:8]\n\n## Sentencing outcome\n\nOn March 28, 2024, the judge sentenced Bankman‑Fried to 25 years in federal prison.[web:5][web:6][web:12] The court also ordered forfeiture and financial penalties tied to what prosecutors said were losses exceeding $8–10 billion to customers, investors, and lenders, reflecting the scope of misappropriated funds and the scale of the fraud.[web:5][web:6][web:9]",
          "type": "output_text",
          "annotations": [],
          "logprobs": []
        }
      ],
      "role": "assistant",
      "status": "completed",
      "type": "message"
    }
  ],
  "status": "completed",
  "error": null,
  "usage": {
    "input_tokens": 7356,
    "output_tokens": 909,
    "total_tokens": 8265,
    "cost": {
      "currency": "USD",
      "input_cost": 0.01509,
      "output_cost": 0.01818,
      "total_cost": 0.0372,
      "cache_creation_cost": null,
      "cache_read_cost": 0.00143,
      "tool_calls_cost": 0.0025
    },
    "input_tokens_details": {
      "cache_creation_input_tokens": 0,
      "cache_read_input_tokens": 3584,
      "cached_tokens": 3584
    },
    "tool_calls_details": {
      "search_web": {
        "invocation": 1
      }
    },
    "output_tokens_details": {
      "reasoning_tokens": 0
    }
  },
  "completed_at": 1779391825
}

Pricing

Billing is based on the model that serves the request, not all models in the fallback chain.

The model field in the response indicates which model was used, and the usage field shows the token counts for that model.

Example

Request:

{
  "models": ["openai/gpt-5.6-sol", "openai/gpt-5.6-terra"],
  "input": "..."
}

Response (if first model failed):

{
  "model": "openai/gpt-5.6-sol",
  "usage": {
    "input_tokens": 150,
    "output_tokens": 320,
    "total_tokens": 470
  }
}

In this case, billing is based on gpt-5.1 pricing for 470 tokens.

Place preferred models first in the array. Consider pricing differences when ordering the fallback chain.

Next Steps

Agent API Models

Explore available models and their pricing.

Effort Levels

Explore available effort levels and their configurations.

Agent API Quickstart

Get started with your first Agent API call.

API Reference

View complete endpoint documentation.

Getting Started

Orchestrator API

Agent API

Search API

Embeddings API

Decisions API

Perplexity SDK

Admin & Management

Resources

Overview

How It Works

Basic Example

Cross-Provider Fallback

Pricing

Next Steps

Agent API Models

Effort Levels

Agent API Quickstart

API Reference

​Overview

​How It Works

​Basic Example

​Cross-Provider Fallback

​Pricing

​Next Steps

Agent API Models

Effort Levels

Agent API Quickstart

API Reference

Overview

How It Works

Basic Example

Cross-Provider Fallback

Pricing

Next Steps